Questory
A quest-and-story app for kids: habits, reading and a virtual pet.
Designed and built solo by AJ Pardilla, with AI agents as the engineering team.
- 1,158 commits in 8 weeks
- Privacy label: data not collected
AJ Pardilla
I spent twenty years designing and building products. Now I build the agents that build them with me.
VP of Product and AI, Unicity International
I lead product and AI at Unicity International: 13 direct reports, 8 product teams, and the checkout and payments stack for more than 50 markets. This is my design portfolio. The work on show is the agents I built, with what I taught each one and how I check it.
The register, current at 18 September 2026
76 agents across 12 systems.
30 run on a clock, 33 on command. The other 13 are built and waiting for command.
The register
12 systems, ordered by what actually runs. Open one to meet its agents
How I checked
55 of these were read out of the files on my own machine. 21 are Grok bots, seen running in my workspace on 18 September 2026 and not audited past that. Each entry says which.
The method comes with me. Agents built inside a company stay there, and get rebuilt inside the next one under its rules.
76 of 76 shown
No agent matches that. Clear the field, or choose All.
Mobile app team
16 agents, 16 running
A sixteen-role software team for a mobile app: research, plan, design, build, test, review, release.
The 3 receipts in this system come from one run on 21 August 2026, not 3 separate ones. Some stages passed only after I overrode a failing grade.
Show the 16 agentsHide the 16 agentsAnalytics Specialist, Backend Specialist, Chief of Staff and 13 more
001Analytics SpecialistDecides what a change should measure: which event, against which baseline, and how success gets judged.on command
- What I taught it
Never invent metrics or claim a launch proved success by itself.Do not fabricate data, declare KPI impact from release alone, query protected data without authority, write providers, approve work, or create agents.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read analytics contracts
- Read KPI strategy
- Run local instrumentation checks
- Emit analytics decision
002Backend SpecialistAdvises on API and service changes, request, response, error and compatibility behaviour, and names the contract tests that must pass.on command
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read backend code
- Read API contracts
- Run local contract tests
- Emit backend guidance
003Chief of StaffSorts an incoming task, ties it to the project's strategy pack, and assembles the smallest qualified team of other agents.on command
- What I taught it
The dispatcher can't change anything in a live system, approve work, or leak its thinking.Do not perform provider writes, approve a handoff, or expose private reasoning.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read project pack
- Classify risk
- Request role through control plane
- Emit work order draft
004Data SpecialistDesigns and reviews database schema and migration work on paper, order, rollback, verification queries, with no access to live data.on command
- What I taught it
It designs database changes but is never allowed to run them live.Do not apply a migration, change live data, use production credentials, approve the candidate, or create agents.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read schema
- Read migrations
- Run local database tests
- Run nonmutating dry run
- Emit data guidance
005DesignerTurns an approved plan into accessible mobile screen designs covering every state: loading, error, empty and translated text.on command
- What I taught it
Never touch or influence the independent QA checking your own design.Do not act as, select, prompt, dismiss, or edit the findings of Independent UX QA.
- How it is checked
- When design QA is required, a separate Independent UX QA agent judges the design before plan approval, in a session the Designer cannot share.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read project pack
- Read research packet
- Create design artifacts
- Emit prototype manifest
006EngineerWrites the code for one approved slice of a plan, in its own isolated workspace, and reports exactly what it covers.on command
- What I taught it
Never grade your own code or rewrite the rules to pass.Do not approve your own work, change acceptance criteria, or create other agents.
- How it is checked
- Always graded by the Evaluator in a separate session from the Engineer, and on the higher QA tiers also reviewed by a separate Independent Reviewer, against agents.
- What it made
- Wrote a small, fully tested code change inside its own isolated workspace, and a separate reviewer agent cleared it, start to finish in about two minutes, with no human writing any of the code.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read repository
- Workspace write
- Run local checks
- Emit candidate evidence
- Request role through control plane
007EvaluatorGrades finished work against a fixed rubric and hard pass/fail gates before it is allowed to ship.on command
- What I taught it
One serious problem can't be smoothed over by a good average score.Do not average away a veto, override a hard gate, or approve merge, data, build, submit, or release.
- How it is checked
- The Evaluator is the grader, so it is bounded from outside: a fixed craft rubric it is forbidden to improve plus hard gates enforced deterministically by the control plane, not by the Evaluator.
- What it made
- Independently graded a plan written by another AI agent against a fixed rubric before it was allowed to proceed, and recorded a 100-out-of-100 pass with zero hard blockers.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read handoff artifacts
- Run deterministic gate checks
- Grade rubric
- Emit coverage receipt
008i18n SpecialistChecks new customer-facing text is translatable, that every language stays in sync, and that translated text still fits on screen.on command
- What I taught it
Never make up a translation and pass it off as translator-approved.Do not invent translator-approved copy, overwrite shared translation sources, write providers, approve work, or create agents.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read locale files
- Read translation governance
- Run i18n checks
- Emit i18n decision
009Independent ReviewerReviews the engineer's code diff adversarially, without reading the engineer's own write-up first, and returns evidence-backed findings.on command
- What I taught it
A clean review must prove it actually looked, not just say fine.Limit ordinary findings to five, deduplicate them, and make zero-finding reviews falsifiable.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric, and required to run in a session separate from the Engineer it reviews.
- What it made
- Reviewed another AI agent's code diff blind, the system physically withholds that agent's self-assessment, and signed off only after checking the change against each acceptance criterion itself.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read candidate diff
- Read acceptance contract
- Run readonly checks
- Emit review findings
- Request role through control plane
010Independent UX QAChecks design quality before a plan is approved, then checks the finished build against the approved design.on command
- What I taught it
Every applicable screen state needs its own real screenshot, not a generic stand-in.Do not approve your own work or substitute generic screenshots for state coverage.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read intake
- Read research
- Read design artifacts
- Read candidate visual evidence
- Emit UX findings
011Mobile SpecialistAdvises on iOS and Android behaviour, navigation, lifecycle, storage, timing, and sets the device test matrix for a change.on command
- What I taught it
Never claim a device test ran when it didn't, and never relax the QA bar.Do not lower mobile QA, claim unrun device evidence, write providers, approve work, or create agents.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read mobile code
- Read platform contracts
- Run local mobile checks
- Emit mobile guidance
012PlannerTurns research and product direction into the written implementation plan the rest of the pipeline builds from.on command
- What I taught it
Unresolved product decisions have to surface, not get buried in engineering tasks.Do not hide unresolved product decisions inside engineering instructions.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- What it made
- Turned research into an implementation plan that a separate evaluator agent scored 100 out of 100 against a fixed rubric, after three rounds of being sent back, before any code was written.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read project pack
- Read research packet
- Read design artifacts
- Emit markdown plan
- Emit HTML explainer
- Request role through control plane
013QA EngineerTests a finished build against its acceptance criteria, including failure and offline paths, and captures screenshots and logs as evidence.on command
- What I taught it
A generic 'it loads' test doesn't count as real coverage.Do not substitute a generic smoke test for task-specific acceptance coverage.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read candidate
- Run test commands
- Request mobile runner job
- Read test artifacts
- Emit QA manifest
014Release OperatorRuns only the exact release step it was pre-authorised for, then confirms it happened using the provider's receipt.on command
- What I taught it
A good score, a passed QA run or a deadline never adds up to permission to build a release.Do not infer build authority from submit authority, QA, a rubric score, or urgency.
- How it is checked
- Executes only the exact operation authorized by a consumed, target-bound capability from the broker, and must emit a provider receipt plus fresh readback evidence for each named operation.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read release policy
- Run nonmutating release preflight
- Prepare brokered release intent
- Read provider receipt
015ResearcherProduces sourced competitor, product, UX and technical research to brief the planner before a feature is built.on command
- What I taught it
Research informs the decision; it never makes the decision up.Do not select the product decision, write the canonical plan, or fabricate market evidence.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- What it made
- Wrote a sourced research brief, citations, findings and explicit limitations, that kicked off a real planning cycle inside a self-built multi-agent engineering pipeline.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read project pack
- Read workspace
- Web research
- Emit research packet
016Security SpecialistReviews login, privacy, credential and permission changes for abuse cases and fail-closed behaviour, independently of the engineer.on command
- What I taught it
It flags security risks but never touches real secrets or live systems.Do not handle production secrets, weaken controls, run destructive tests, write providers, approve a release, or create agents.
- How it is checked
- Graded by the Evaluator agent against this role's fixed 15-point craft rubric.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Read security contracts
- Read candidate diff
- Run safe negative tests
- Emit security findings
Product Command Center
16 agents, 15 running
Gathers signals from Slack, ClickUp, mail and calendar, writes the morning brief, then reviews itself.
Show the 16 agentsHide the 16 agentsPCC Accuracy Reviewer, PCC Briefing Critic, PCC Briefing Writer and 13 more
017PCC Accuracy ReviewerChecks every briefing item written this run for a source link, real evidence and the right person, then debates findings with the other reviewers.on a clock
- How it is checked
- Its findings go through a live 2-3 round debate with the relevance and completeness reviewers before the lead applies any fix.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Synthesis and signals API (read)
- SendMessage (reviewer debate)
018PCC Briefing CriticChecks the draft briefing against the actual database items, nothing urgent missed, calendar conflicts noted, and sends feedback back for one revision.on a clock
- How it is checked
- It is itself the check on the briefing writer, and its own feedback triggers one mandatory writer revision before the lead posts the briefing.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Items and calendar (read)
- SendMessage (to the writer)
019PCC Briefing WriterDrafts AJ's daily briefing in plain spoken prose from the current items and calendar, most urgent first, then hands it to the critic.on a clock
- How it is checked
- A second agent, the briefing critic, reviews the draft against the real database items and sends it back for revision before it ships.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Items and calendar (read)
- SendMessage (to the critic)
020PCC Collector: ClickUpPulls active and blocked ClickUp tasks for AJ's tracked teams and posts them to the briefing system as signals.on a clock
- How it is checked
- Its output is cross-checked by the completeness-reviewer phase for gaps and duplicates; that review runs after the items are written, and the lead then applies fixes or drops the items the reviewers agree on.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- ClickUp (via MCP)
- Bash (posts to the ingest API)
021PCC Collector: Gmail + CalendarPulls Gmail and Google Calendar signals for the daily briefing, falling back to the Google connectors when the command-line collector is broken.on a clock
- How it is checked
- The pre-flight connector probe treats the Gmail/Calendar path as nice-to-have with a documented degraded mode rather than a hard stop, and the completeness-reviewer phase then checks its coverage before the lead finalizes items.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Bash (Google Workspace CLI collectors for Gmail and Calendar)
- Gmail and Google Calendar (via MCP, fallback only)
022PCC Collector: SlackSearches AJ's priority Slack channels and posts what it finds to the briefing system as signals.on a clock
- How it is checked
- Its signals are validated by the required-connector pre-flight probe (a broken Slack MCP call halts the whole refresh) and later by the completeness-reviewer.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Slack (via MCP: search and read threads)
- Bash (posts to the ingest API)
023PCC Collector: TelegramRuns the Telegram collector script, reads the signals it produces, and posts them to the briefing system.on a clock
- How it is checked
- Its liveness is checked by an automated heartbeat watchdog that pages if the listener goes silent for more than 10 minutes.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Bash (Telegram collector script)
- Bash (posts to the ingest API)
024PCC Completeness ReviewerChecks that every important signal this run, incidents, blockers, direct asks to AJ, became an item, and flags duplicates.on a clock
- How it is checked
- Its own findings are debated live with the accuracy and relevance reviewers for 2-3 rounds before the lead applies any fix.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Signals API (read)
- SendMessage (reviewer debate)
025PCC Pattern DetectiveGroups this run's poorly rated items by cause and proposes config fixes: new stop words, tighter thresholds, better filters.on a clock
- How it is checked
- Its proposed fixes go through a debate to consensus with the signal-auditor and synthesis-auditor before the lead applies only the safe ones.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Ratings API (read)
- SendMessage (debug-team debate)
026PCC Relevance ReviewerCompares this run's items against what AJ has rated badly before, and pushes the team to demote or drop repeat offenders.on a clock
- How it is checked
- Its challenges are debated live with the accuracy and completeness reviewers before the lead drops or keeps an item.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Ratings feedback API (read)
- SendMessage (reviewer debate)
027PCC Signal AuditorReviews the last run's raw signals for self-referencing content, stale ClickUp tasks slipping through, and wrong attribution.on a clock
- How it is checked
- Spawned only when 3 or more items were rated 2 or lower in the last 24 hours (the threshold is a config setting), and its findings must reach debate consensus with two other auditors before any fix is applied.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Signals API (read)
- SendMessage (debug-team debate)
028PCC Snapshot Team WriterWrites one tracked team's person and team snapshots, proposes items needing attention, and cross-checks with the other team writers before anything is saved.on a clock
- What I taught it
When people disagree, report both positions and who is deciding, never a winner.Never silently pick one side.
- How it is checked
- Every writer cross-checks its proposed items against the other 8 team writers via a live message round before the lead writes anything to the database.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Tools it may use
- Synthesis API (person and team snapshots, item proposals)
- SendMessage (teammate cross-check)
029PCC Synthesis AuditorReviews briefing items that scored under 40 for confidence, looking for weak grounding or missing evidence.on a clock
- How it is checked
- Runs only when 3 or more items were rated 2 or lower in the last 24 hours (the threshold is a config setting), and its findings must reach debate consensus with the signal-auditor and pattern-detective.
- How I know
- Audited against the files on AJ's machine. Running status confirmed by AJ on 19 September 2026.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Items API (read)
- SendMessage (debug-team debate)
030PCC v3 Daemon (pccd)AJ's live briefing system: collectors gather signals, a gated Claude step writes the items and morning briefing, and urgent alerts reach him immediately.on a clock
- What I taught it
Don't make a solved problem sound like it's still open.Resolved means resolved. Never present a resolved item as pending.
- How it is checked
- Its own heartbeat is monitored by a watchdog process that pages if the daemon goes silent for more than 10 minutes, and separately if no briefing exists by 9am.
- What it made
- One always-on agent runs a full collect-synthesize-brief cycle roughly every half hour and leaves a dated output folder as a receipt for each run.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 18 September 2026
- How it runs
- Running under pm2 as process 'pccd' (status: online), alongside five sibling v3 processes also online: pcc-v3-incident-poller, pcc-v3-feed-sweep, pcc-v3-watchdog, pcc-v3-vip, pcc-v3-self-improve. It spawns a gated 'claude -p' synthesis session (SYNTHESIS_MODEL, BRIEFING_PROMPT_PATH) on each cycle.
- Runs on
- Anthropic Claude, Sonnet 5
- Tools it may use
- Read
- Write
- Grep
- Glob
031PCC Dashboard AssistantAnswers AJ's questions about team status, incidents and meetings inside his dashboard, reading a curated knowledge wiki before calling any live tool.on command
- How I know
- Audited against the files on AJ's machine.
- How it runs
- Chat assistant embedded in the PCC v2 web dashboard; it fires only when AJ sends it a message and has no schedule. Its standing instructions are the system preamble built in the dashboard's own code ('You are embedded in the PCC dashboard...'). The pm2 process that serves the dashboard ('pcc-web') is online today, so the surface is reachable.
- Tools it may use
- Slack MCP
- Gmail MCP
- Calendar MCP
- ClickUp MCP
- Supabase MCP
- Mixpanel MCP
032PCC v2 Morning RefreshThe older 7:15am weekday briefing run: collector scripts gather signals, then one Claude session does triage, items, snapshots and the briefing.built
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 15 July 2026 (date the trigger was disabled; no run evidenced since)
- How it runs
- Its only trigger, the launchd job com.pcc.scheduled-refresh (weekdays 7:15am plus a 7:45am retry), is disabled: the plist file was renamed com.pcc.scheduled-refresh.plist.disabled-2026-07-15. No crontab exists on this machine at all, and no pm2 process replaces it. Superseded by the PCC v3 Daemon.
- Tools it may use
- Bash (curl)
- Read
- Grep
- Glob
16 agents in Product Command CenterPCC Accuracy ReviewerPCC Briefing CriticPCC Briefing WriterPCC Collector: ClickUpPCC Collector: Gmail + CalendarPCC Collector: SlackPCC Collector: TelegramPCC Completeness ReviewerPCC Pattern DetectivePCC Relevance ReviewerPCC Signal AuditorPCC Snapshot Team WriterPCC Synthesis AuditorPCC v3 Daemon (pccd)PCC Dashboard AssistantPCC v2 Morning Refresh
The engineering pipeline
9 agents, 9 running
Takes one feature or bug from research to a graded, tested, committed change.
The 9 receipts in this system come from one run on 6 September 2026, not 9 separate ones. Some stages passed only after I overrode a failing grade.
Show the 9 agentsHide the 9 agentspipe-doc-writer, pipe-grader, pipe-implementer and 6 more
033pipe-doc-writerUpdates the project's wiki, README and changelog so the documentation matches what was actually built.on command
- How it is checked
- pipe-grader scores it against its written rubric.
- What it made
- Updated six separate documents so a project's written record matched what was actually built, as part of a real, shipped production change.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 14 September 2026
- How it runs
- Invoked when a person types the pipeline command in a target repo. The standing role file lives in agent-pipeline; each invocation writes its stage-06 receipt into the target repo's own run folder. A stage receipt exists in three of the four runs found.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Edit
- Write
- Grep
- Glob
034pipe-graderScores one pipeline stage's output against that stage's rubric, blind, and writes the pass/fail the coordinator checks before continuing.on command
- What I taught it
Score the actual output, not how hard it tried.Grade the artifact, not the effort.
- How it is checked
- No downstream agent grades the grader; a deterministic lint hook checks the grade file's frontmatter is complete and refuses a 'pass' verdict written over any criterion scored below the rubric's threshold.
- What it made
- Independently scored every stage of a real production feature build against that stage's fixed rubric, and the next stage could not start until the score was in.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 15 September 2026
- How it runs
- The same pipeline mechanism. Standing role file; its output is the grades folder every run produces, one graded file per stage and retry.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Grep
- Glob
- Write
035pipe-implementerBuilds one work package from the brief at a time, test first, inside the approved file list, and finishes only when the checks pass.on command
- What I taught it
Prove the test fails first, then write the fix.Never write implementation code before its failing test exists.
- How it is checked
- pipe-grader scores it against its written rubric, and a separate stage (pipe-test-verifier) writes outside-in acceptance tests for every criterion and runs the full suite itself, rather than trusting the implementer's account.
- What it made
- Implemented the code for a real feature that was later merged, deployed and health-checked in production, working test-first inside its own isolated workspace.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 14 September 2026
- How it runs
- The same pipeline mechanism. Standing role file; stage-04 output is written per work package inside each run folder.
- Runs on
- Anthropic Claude, Opus
- Tools it may use
- Read
- Edit
- Write
- Bash
- Grep
- Glob
036pipe-plan-auditorRuns a Claude-versus-Codex adversarial audit of the technical brief before AJ approves it: ship, ship with caveats, rescope or incomplete.on command
- What I taught it
Do not fake agreement between the two reviewing models; dissent is preserved word for word.Never manufacture consensus.
- How it is checked
- pipe-grader scores it against its written rubric; its own output is a cross-model debate (Claude seat vs Codex seat) rather than a single self-check.
- What it made
- Audited a technical brief before it was approved and correctly flagged it as not ready, forcing the human owner to explicitly override the block before work continued.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 14 September 2026
- How it runs
- The same pipeline mechanism. Standing role file; stage-03b output is written inside each run folder.
- Runs on
- Anthropic Claude
- Tools it may use
- Read
- Grep
- Glob
- Bash
- Write
037pipe-plannerWrites the technical brief every later stage follows: data model, flow, API, frontend, tests, risks, docs and the ordered work packages.on command
- How it is checked
- pipe-grader scores the brief against its written rubric, and pipe-plan-auditor separately pressure-tests it in a cross-model Claude-vs-Codex audit before AJ approves it.
- What it made
- Wrote the complete technical brief, data model, process flow, an ordered work-package list, the tests required and thirty named risks, that every later stage of a production feature build followed.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 14 September 2026
- How it runs
- The same pipeline mechanism. Standing role file; stage-03 output is written inside each run folder.
- Runs on
- Anthropic Claude
- Tools it may use
- Read
- Grep
- Glob
- Write
038pipe-researcherMaps the part of a codebase a feature or bug touches, finds the test and lint commands, and ranks likely causes.on command
- How it is checked
- pipe-grader scores its receipt against its written rubric and writes a pass/fail grade file before the next stage can run.
- What it made
- Wrote the research stage of a multi-agent engineering pipeline that mapped a real production codebase and its test commands before any code was written.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 14 September 2026
- How it runs
- The same pipeline mechanism. Standing role file; stage-01 output is written inside each run folder.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Grep
- Glob
- Bash
039pipe-spec-writerTurns the research into one user story with acceptance criteria a test can check, or an expected-versus-actual write-up for a bug.on command
- What I taught it
Do not add filler requirements just to look complete.Never pad the criteria list to look thorough; each one must earn its place.
- How it is checked
- pipe-grader scores it against its written rubric; a fail routes the spec back before the next stage runs.
- What it made
- Turned research into a full acceptance-criteria spec for a real production feature, the kind of spec a test can actually check against, not a wish list.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 14 September 2026
- How it runs
- The same pipeline mechanism. Standing role file; stage-02 output is written inside each run folder.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Grep
- Glob
- Write
040pipe-test-verifierProves the build meets every acceptance criterion by writing its own outside-in tests, without trusting the implementer's claims.on command
- What I taught it
Do not claim a test proves something it does not check.Never mark a criterion covered when the test does not actually assert it.
- How it is checked
- pipe-grader scores it against its written rubric; the acceptance test files are frozen by SHA-256 digest, which the gate recomputes before the run can close, so a later stage cannot quietly weaken them.
- What it made
- Independently proved a finished build satisfied every one of its acceptance criteria by writing its own outside-in tests, rather than trusting the builder's word for it.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 14 September 2026
- How it runs
- The same pipeline mechanism. Standing role file; stage-05 output is written inside each run folder.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Edit
- Write
- Bash
- Grep
- Glob
041pipe-validatorReviews the finished change blind, from the spec, the brief and the raw diff only, grading each finding critical, important or minor.on command
- What I taught it
Do not invent problems for show, and do not shrink a real one to let the work through.Never manufacture findings to look rigorous, and never downgrade a real one to unblock the gate.
- How it is checked
- pipe-grader scores it against its written rubric; it reviews blind (spec, brief and raw git diff only, never the builders' own account) and the run cannot close while it reports any critical or important finding.
- What it made
- Ran a blind review of the final code diff, without seeing what the other agents said about their own work, and cleared it with zero critical or important findings.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 15 September 2026
- How it runs
- The same pipeline mechanism. Standing role file; stage-07 output is written inside each run folder.
- Runs on
- Anthropic Claude
- Tools it may use
- Read
- Grep
- Glob
- Bash
9 agents in The engineering pipelinepipe-doc-writerpipe-graderpipe-implementerpipe-plan-auditorpipe-plannerpipe-researcherpipe-spec-writerpipe-test-verifierpipe-validator
The trading desk
8 agents, 8 running
Eight bots argue a trade. Risk can veto it, and no order is placed without AJ.
Show the 8 agentsHide the 8 agentsBear, Bull, Fundamentals and 5 more
042BearBuilds the case against a trade on AJ's trading desk.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- Research stack
- Desk handoff
043BullBuilds the case for a trade on AJ's trading desk.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- Research stack
- Desk handoff
044FundamentalsTrading-desk analyst covering company filings and fundamentals.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- Filings
- Fundamentals data
045NewsTrading-desk analyst covering market news.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- Market news
046RiskReviews position size and risk before any live trade on AJ's desk.on a clock
- What it may never do
- Holds a veto: the chair is blocked until Risk clears the trade.
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- Position and risk review
047SentimentTrading-desk analyst covering market sentiment, read from X and market data.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- X
- Market data
048TechnicalTrading-desk analyst covering quotes, charts and technicals.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- Quotes and charts
049TraderTurns the trading desk's analysis into concrete trade ideas for the chair to decide on.on a clock
- What it may never do
- Cannot place or cancel an order itself.
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- Market stack
The research bench
5 agents, 5 running
Finds evidence, reports what people actually said, and grades its own packets.
Show the 5 agentsHide the 5 agentsCompeti, Overheard, 30 and 2 more
050CompetiWatches competitors' pricing, product and hiring pages for changes, and sends a weekly brief.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Weekly brief
- Runs on
- xAI Grok
- Tools it may use
- Watch-list pages
- Page diff
- Weekly brief
051OverheardTracks mentions of AJ's name, brand and URLs across Reddit, Hacker News, news and X.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Subscriptions and digest
- Runs on
- xAI Grok
- Tools it may use
- Hacker News
- News
- X
- Digest
05230Reports what people actually said in the last 30 days on Reddit, X, Hacker News and YouTube.on command
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- Runs on
- xAI Grok
- Tools it may use
- X
- Hacker News
- YouTube
- Web
053EveGrades the research packets the other bots produce against a fixed rubric.on command
- What it may never do
- Grades only. Never does the research she is grading.
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- Runs on
- xAI Grok
- Tools it may use
- Evaluation rubric
054OogleDeep research on marketing, competitors and history, using web search and GitHub as sources.on command
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- Runs on
- xAI Grok
- Tools it may use
- Web search and fetch
- Deep-research skill
- GitHub as a source
Personal staff
4 agents, 4 running
Chief of staff, critic, coach, and the chair of the trading desk.
Show the 4 agentsHide the 4 agentsMonie, Skeppie, Warp and 1 more
055MonieChairs AJ's trading desk, and is the only bot allowed to place a trade, through a Robinhood agentic account.on a clock
- What it may never do
- Every live entry needs AJ's tap to accept, and Risk must clear it first.
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Midday desk cycle
- Runs on
- xAI Grok
- Tools it may use
- Robinhood (Agentic account)
- Accept widgets
- Trading-desk handoffs
056SkeppieIndependent critic: reads AJ's day and pushes back on it, in a weekday digest.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Weekday digest routine
- Runs on
- xAI Grok
- Tools it may use
- Chat
- Weekday digest routine
057WarpAJ's personal chief of staff: trips, documents and his personal inbox.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Scheduled routines
- Runs on
- xAI Grok
- Tools it may use
- Personal Gmail
- Personal GitHub
- Vercel
- Browser
- Scheduled routines
058GymmieGym coach: programmes sessions, reviews lifting form from photos and video, and sets recovery rules.on command
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- Runs on
- xAI Grok
- Tools it may use
- Session programming
- Photo and video form review
Work staff
4 agents, 4 running
The work lane: weekday brief, HR drafts, brand, and build.
Show the 4 agentsHide the 4 agentsUni, Atelier, Codie and 1 more
059UniAJ's work chief of staff at Unicity: weekday brief, Friday wrap, incidents, brewing risks and meeting prep.on a clock
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- How it runs
- Weekday 9am brief, Friday wrap, incident routines
- Runs on
- xAI Grok
- Tools it may use
- Work Gmail
- Slack
- ClickUp
- Google Calendar
- Scheduled routines
060AtelierBrand, pitch and visuals: the story goes to Figma before anything gets built.on command
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- Runs on
- xAI Grok
- Tools it may use
- Figma
- Higgsfield
- ElevenLabs
061CodieImplementation engineer across AJ's work and personal codebases, through GitHub and Cursor cloud agents.on command
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- Runs on
- xAI Grok
- Tools it may use
- GitHub (work and personal repos)
- Cursor cloud agents
062PollieUnicity HR lane: hiring, compensation and personal-data questions. It drafts only.on command
- What it may never do
- Draft-only. Never sends anything itself.
- How I know
- Seen running in AJ's Grok workspace on 18 September 2026.
- Last ran
- 18 September 2026
- Runs on
- xAI Grok
- Tools it may use
- Unicity HR workspace
Department reporting
1 agent, 1 running
Checks each morning that the AI department's own reporting actually ran.
Show the 1 agentHide the 1 agentKPI Integrity Reviewer
063KPI Integrity ReviewerChecks each morning whether the AI department's KPI reporting actually ran: data freshness, deploy status, measured or carried-over numbers, then reports to AJ.on a clock
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 31 July 2026
- How it runs
- Claude cloud scheduled routine, daily 10:30 UTC, read-only, DMs AJ. It executes on claude.ai infrastructure, so it cannot be observed with launchctl/pm2/crontab on this machine; the mechanism is proven only from the project brain's own dated write-up, not from a running process.
1 agent in Department reportingKPI Integrity Reviewer
HR answer bot
1 agent, 1 running
Answers employees' HR policy questions in Slack with a source on every answer.
Show the 1 agentHide the 1 agentHR Assistant (hr-answer-bot / @hr_bot)
064HR Assistant (hr-answer-bot / @hr_bot)Answers US employees' HR policy questions in Slack from the employee handbook, cites a source every time, and routes anything else to a human.on a clock
- What it made
- An internal Slack-based question bot answered about 6 in 10 of the decisioned questions put to it during a pilot running from late July to mid-September, and every answer it gives carries a source citation by design.
- How I know
- Audited against the files on AJ's machine.
- Last ran
- 17 September 2026
- How it runs
- Deployed Slack Socket Mode bot on Google Cloud Run. The project's append-only wiki log records dated merges through 2026-09-17, each of which redeploys a new Cloud Run revision via the repo's build trigger.
- Runs on
- Anthropic Claude, Sonnet 5
- Tools it may use
- Slack Socket Mode (Bolt)
- Anthropic SDK (citations)
- Google Drive corpus sync
- Document AI (PDF/DOCX parsing)
1 agent in HR answer botHR Assistant (hr-answer-bot / @hr_bot)
Storefront reviewers
5 agents, 0 running
Reads changes to a global storefront before they are allowed near a pull request.
Show the 5 agentsHide the 5 agentscode-reviewer, goal-verifier, pre-pr and 2 more
065code-reviewerReviews code changes in Unicity's Shop2 storefront for quality, conventions and correctness before a person reviews them.built
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Grep
- Glob
- Bash
- Write
- Edit
066goal-verifierRe-runs every acceptance check itself in a fresh context instead of trusting the implementer's evidence, then writes a pass or fail verdict.built
- What I taught it
Wrongly approving bad work is the worst outcome; failing a borderline change is cheaper.A false PASS is the worst possible outcome
- How it is checked
- This agent is the check: it is the only agent allowed to write the PASS/FAIL verdict into the goal-plan ledger, and it re-runs the deterministic checks itself (format/lint, build, git diff) instead of trusting the implementer's pasted evidence.
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Grep
- Glob
- Bash
- Edit
- WebFetch
067pre-prChecks lint, types, build and documentation freshness before a pull request on the Shop2 storefront is opened.built
- How it is checked
- Runs lint/format, build, a convention scan and a doc-freshness scan itself and prints a READY / NOT READY verdict; no further agent checks it.
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Grep
- Glob
- Bash
068quick-fixHandles small isolated fixes under about ten lines, typos, one-line bugs, bad imports, without a full planning pass.built
- What I taught it
When unsure, stop and ask instead of guessing.If you're unsure about the fix, STOP and ask rather than guessing
- How it is checked
- Runs `npm run formatAndLint` itself after the edit and reports passed/failed; no separate reviewer agent is defined for it.
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Edit
- Grep
- Glob
- Bash
069researcherExplores the Shop2 storefront codebase to find patterns, locate files and explain the architecture before a feature is planned.built
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Grep
- Glob
Idea appraisal
4 agents, 0 running
Turns a one-line idea into sourced evidence, a critic's attack, and a go or no-go.
Show the 4 agentsHide the 4 agentsdisc-critic, disc-designer, disc-presenter and 1 more
070disc-criticAttacks the research's weakest assumptions, checks the cited sources are real, hunts counter-evidence, and returns a go / no-go verdict.built
- What I taught it
Stopping a bad idea early counts as the process working.A killed idea early is a success for this process, not a failure.
- How it is checked
- pipe-grader scores it against its written rubric; it works blind, receiving file paths only and never the researcher's reasoning or confidence.
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude
- Tools it may use
- Read
- Grep
- Glob
- Write
- WebSearch
- WebFetch
- ToolSearch
071disc-designerBuilds a clickable HTML prototype in the right design system when an idea needs a user interface, with borrowed patterns sourced.built
- How it is checked
- pipe-grader scores it against its written rubric; the prototype must open standalone with no console errors, and every borrowed pattern must be listed with its source.
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude, Opus
- Tools it may use
- Read
- Write
- Edit
- Grep
- Glob
- WebSearch
- WebFetch
- ToolSearch
072disc-presenterWrites the stakeholder summary page for an idea: sourced numbers split into proven, built but unproven, and dependent on someone else.built
- What I taught it
Only label something 'verified' if it was actually checked.Never move something into "working now (verified)" that no one has verified.
- How it is checked
- pipe-grader scores it against its written rubric, and it scores its own prose with the stop-slop checklist: below 35/50 it must revise the page and score again, and may not return under 35.
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Write
- Edit
- Grep
- Glob
073disc-researcherTurns a one-line business idea into cited competitor evidence, a recommended approach, success metrics, an effort estimate and kill criteria.built
- What I taught it
Every fact and figure must be real, never made up.Never invent a number, a baseline, a competitor, or a product feature.
- How it is checked
- pipe-grader scores it against its written rubric, and disc-critic then attacks the same receipt with the stated job of trying to kill the idea.
- How I know
- Audited against the files on AJ's machine.
- Runs on
- Anthropic Claude, Sonnet
- Tools it may use
- Read
- Grep
- Glob
- Bash
- WebSearch
- WebFetch
- ToolSearch
Growth pipeline
3 agents, 0 running
Pulls app analytics and feedback, ranks what to build next, and builds one.
Show the 3 agentsHide the 3 agentsBuilder (growth pipeline), Gatherer (growth pipeline), Synthesizer (growth pipeline)
074Builder (growth pipeline)Builds one approved feature for The Method app on its own branch, runs checks, updates docs, and reports the commit to ClickUp and Slack.built
- How it is checked
- Runs lint, a TypeScript typecheck (`npx tsc, noEmit`) and lint:expo itself after implementing and must fix any failure before proceeding; no separate grader agent exists in this pipeline.
- How I know
- Audited against the files on AJ's machine.
- Tools it may use
- ClickUp (read task)
- ClickUp (update task)
075Gatherer (growth pipeline)Pulls The Method app's Mixpanel analytics, Supabase feedback rows and Slack wishlist ideas into one structured file per cycle.built
- How I know
- Audited against the files on AJ's machine.
- Tools it may use
- Mixpanel (via MCP)
- Supabase (SQL via MCP)
- Slack (read channel)
076Synthesizer (growth pipeline)Cross-references the gathered analytics, feedback and ideas, and ranks the features most likely to move the chosen growth metric.built
- How it is checked
- Runs its own Reality Check step before finalizing: each proposal must be actionable, scoped, measurable, and aligned to the success metric.
- How I know
- Audited against the files on AJ's machine.
3 agents in Growth pipelineBuilder (growth pipeline)Gatherer (growth pipeline)Synthesizer (growth pipeline)
House rules
4 of 11 are enforced by code
Databases are read-only by default
Never run a database write without the human explicitly approving that exact change first.
Enforced by a hook that stops the call and makes the human approve itPersonal and work accounts never mix
Keep personal and work identities separate on every service; a guard blocks the wrong one automatically.
Enforced by a hook that blocks the actionName the model on every call
Always set the model explicitly per agent call; the strongest model is reserved for the hardest judgment seats.
Enforced by a hook that blocks the actionNothing goes out unscored
Score outbound writing on five clarity dimensions and revise anything below the pass threshold before sending.
Enforced by a hook that fires before every outbound send and forces the checkBe a critic, not a yes-man
Stress-test ideas and lead with what's wrong before offering any praise.
A rule I followPlain language, three honesty buckets
Explain work in plain terms and sort progress into working, unproven, and dependent-on-others, honestly.
A rule I followPipelines run only when invited
Never start the multi-agent pipeline on an ordinary request; recommend it and wait for explicit approval.
A rule I followA second model audits the plan
Run an adversarial Claude-versus-Codex debate on a plan and resolve blockers before building starts.
A rule I followTests first, then the fix
Write the failing test before writing the fix, and never edit a test just to make it pass.
A rule I followExternal code stays external
Never install internal tooling, tracking, or process into a third-party repo without explicit case-by-case approval.
A rule I followOne writer at a time
Never let two agents edit the same repository at once unless they touch completely separate files.
A rule I followHard limits on 5 agents
- Risk, on the trading desk. Holds a veto: the chair is blocked until Risk clears the trade.
- Trader, on the trading desk. Cannot place or cancel an order itself.
- Eve, on the research bench. Grades only. Never does the research she is grading.
- Monie, on personal staff. Every live entry needs AJ's tap to accept, and Risk must clear it first.
- Pollie, on work staff. Draft-only. Never sends anything itself.
Two apps in the stores
Live in the stores
The Method by Unicity
The daily-practice companion for graduates of The Method Seminar: goals, vision, breathwork and streaks, in nine languages.
Designed and built solo by AJ Pardilla, with AI agents as the engineering team. Published by Unicity.
- Nine languages, narrated audio in each
- On the App Store and Google Play
In the week of 31 August 2026: 808 people used The Method, 466 a day, 517 of them practised, 4,726 sessions.
58% of weekly users open the app on any given day. Average daily users grew from 75 in April to 475 in September, with a peak of 493 on 9 August.
New members still practising, by day
Not there yet
- Average daily users are at 47% of the 1,000 launch target
- New-member first-practice conversion was 17.6% in the latest mature weekly cohort (17 new profiles)
- Day 66+ share fell from 42.7% to 26.8% over five months as the member base grew
The Method Weekly Pulse, production analytics in Mixpanel, project 4001739, generated 11 September 2026. Small internal and testing traffic may be included.
The pipeline behind it
Eight steps, and where each one really runs
Cloud agents that take a feature from a Slack message to a reviewed, device-tested build AJ approves from his phone. About ten dollars a month.
Between 5 and 13 September 2026, three Slack requests were planned, coded and independently reviewed by agents in the cloud. One came back as a build AJ approved from his phone, and the cloud recorded that approval and advanced the ticket by itself.
30 agents in the register run on a clock and need nobody. This pipeline cannot yet: two of its eight steps wait on me.
- AJ posts a request in Slack.ran in the cloud
- A Chief of Staff agent scopes it.ran in the cloud
- Researcher, Planner and Evaluator agents write and score a plan.ran in the cloud
- AJ approves in Slack. First human gate.ran in the cloud
- An Engineer agent writes code. An independent Reviewer checks it.ran in the cloud
- QA agents write device tests. A second Mac runs them with the screen locked.ran with nobody watching, on the test Mac
- A preview build publishes and a QR code posts to Slack. Second human gate.has needed AJ's machine every time
- A pull request opens, a human merges, and store submission stays manual.has not happened yet
Sixteen agent roles are defined and eight have run. Running with the computer shut is not proven, and nothing from this pipeline has merged into the app yet.
Receipts
Five, each with the thing you would check
- Shipped
- A pipeline of agents took a real feature from request to a merged, deployed, health-checked change. I reviewed and decided. I did not write the code.
- PR #29, commit 8060517. Deployed 6 September 2026.
- Measured
- An internal registry of agents and projects, with a test suite above 95% statement coverage. Its dashboard prints “not measured” where no data exists.
- Commit 8378a9f, 9 July 2026. 428 automated tests.
- Cited
- An HR question bot in Slack answered about 6 in 10 decided questions. Every answer carries a source. When it cannot cite one, it refuses.
- Seven-week pilot. 54 cited answers, about 100 questions.
- Stopped
- I ran a ten-minute bug fix through the full pipeline, stopped it when the cost outgrew the fix, and changed the rule for when the pipeline may be used.
- 61 agent runs, 9 roles, about 20 hours, for one fix.
- Running
- One always-on agent runs a collect, synthesise and brief cycle about every half hour, and leaves a dated receipt for each one.
- The pccd daemon. 35 dated runs on 18 September 2026.
Work
56 pieces. 10 can be opened from here, and those come first
56 of 56 shown
The Method19 pieces
The daily-practice companion for graduates of The Method Seminar: goals, vision, breathwork and streaks, in nine languages.
app
live
The early one-page marketing site that also carries the privacy policy and terms the App Store submission needed.
website
live
A public site where anyone types a goal and an AI interview turns it into an editable, first-person Vision Statement, with no sign-in.
website
live
An independent Spanish edition of the Vision Workshop interview, forked from the English source and published at its own address.
website
live
Breathwork engine
An animated breathing engine inside the daily practice, with two methods that ship: Tummo and Box Breathing.
app feature
live
Imagineering guided audio
Guided visualisation sessions with professionally narrated audio in nine languages, plus a self-guided mode that runs to your own music.
app feature
live
Leaderboard V2
A rebuilt community leaderboard with a global weekly ranking authority, automatic enrolment and immediate midweek participation.
app feature
live
OCR manuscript capture
Photograph a handwritten page of the paper seminar workbook and the typed text lands in the app's editor.
app feature
live
Thai language rollout
Thai added as a first-class language across the interface, dates, notifications, dictation and goal validation.
app feature
Agentic Pipeline Core (V3)
Cloud agents that take a feature from a Slack message to a reviewed, device-tested build AJ approves from his phone. About ten dollars a month.
pipeline
The Method Agentic Pipeline V1.5 (retired)
The laptop-bound predecessor that did drive real merges into the shipped app, gated on an approval phrase AJ typed in Slack.
pipeline
The Method iPhone Duo Site
A browser prototype that puts four of the app's screens on a photorealistic dual-iPhone 3D model, with the full five-step practice playable, nothing to install.
prototype
live
Method project pack
A local-only pack of Method product strategy, release rules and QA policy that the agentic pipeline reads, with no power to act.
system
The Method MCP server
A server shipped inside the app's repo that lets any AI assistant query the app's documentation, API surface and analytics vocabulary, read-only by design.
system
The Method release and evidence system
The release machinery that makes every build, submission, migration and deploy a dated record naming its exact commit, its build ID and what was deliberately not done.
system
The Method wiki as project brain
A 97-page in-repo wiki that both AI coding agents must read before a task and update after it, replacing about 49 scattered documents.
system
The Method cinematic marketing page
A scroll-driven marketing page for The Method, its video generated with AI, built for Unicity's corporate site with verified store QR codes.
website
The Method growth pipeline
Three agents that pull analytics, user feedback and wishlist ideas, rank the features most likely to move a chosen metric, and build an approved one.
workflow
Weekly stakeholder analytics pulse
A weekly KPI report written by agents from production analytics. It prints "working thesis, not proof of causality" and "threshold not met" rather than over-explain.
workflow
live
Everything else37 pieces
A finished Next.js itinerary site for AJ's own family trip to Tokyo, El Nido, Siargao and Manila, live on his personal domain.
app
open
A weekly pipeline, GitHub Actions, Apify scraping, Supabase, a one-tap Slack approval, that ranks Unicity distributors' TikTok content on a live dashboard.
pipeline
open
A study of Unicity distributors' TikTok affiliate accounts, finding prohibited health-claim wording in usernames and bios alongside near-zero orders for most accounts.
study
open
A packaging and market research brief for a Unimate seven-day trial pack: four box concepts and 35 renders, written for Unicity's President as one HTML page.
study
open
Unicity's internal AI Piloting Academy training: a built Session 1 slide deck on a public hub site, plus an outline for Sessions 2-10 on Claude Code.
training
open
Everyday Protein: AI-generated marketing creative
An AI-generated marketing concept for Unicity's Everyday Protein: a mocked product page plus several video clips made with Higgsfield.
ai video
Unicon 2026: AI-generated event creative
AI-generated hero images and cinematic clips, made with Seedance and upscaled with Topaz, as visual creative for Unicity's Unicon 2026 event.
ai video
FeelGreat Fasting App (mobile)
A React Native mobile app for Unicity's FeelGreat fasting product.
app
Shop2: global e-commerce platform rebuild
A from-scratch rebuild of Unicity's global storefront in Next.js, multi-market, multi-locale, built with its own in-repo AI development workflow.
app
AI Department KPI site + HR-bot telemetry pipeline
A daily-refreshing internal site reporting the AI department's KPIs, fed by identity-free telemetry from the HR bot's Slack usage.
dashboard
Conduct: the AI Department's agent & project atlas
AJ's visual map of Unicity's AI department: people, knowledge bases, agents, data feeds and access grants, each marked verified or merely planned.
dashboard
CS Hub: Knowledge & Learning Centre
A prototype knowledge and learning hub for Unicity's customer support team, paired with a written security audit of its own design.
dashboard
Product Command Center (PCC)
An ops dashboard for a product leader: about 29 agents pull from Slack, Gmail, Calendar, ClickUp and Telegram into one page of actions and team status.
dashboard
Vibe-Trading paper-trading agent setup
AJ set up, rather than built, an open-source trading agent on a simulated Alpaca paper account to experiment with LLM-driven trading.
other
/discover: evidence-backed go/no-go idea pipeline
An opt-in pipeline of four AI workers that turns a two-sentence idea into cited competitor research, a blind critique, a clickable prototype and a go/no-go summary.
pipeline
/pipeline: graded multi-agent engineering pipeline
An opt-in assembly line of nine AI workers taking a feature or bug from research through planning, an adversarial audit, test-first build and blind review.
pipeline
Agentic Pipeline (agentic-pipeline-core)
A multi-agent software factory AJ built for himself: role-bound agents research, plan, design, build, review, QA and release features for The Method app.
pipeline
HR Answer-Quality Eval Lane
An automated quality check that tests the HR Assistant against a held-out bank of real HR questions on every deploy and document update.
pipeline
HR Bot CI/CD + Ops Automation
The push-to-deploy pipeline for the HR Assistant: every change is tested, built and shipped automatically, with alerts if the service goes down.
pipeline
HR Corpus Ingest Loop
A background loop that checks the approved HR document folder every five minutes, so policy changes reach the bot without anyone redeploying it.
pipeline
KPI Metrics Refresh Pipeline
The daily automation behind the AI department's KPI numbers: it reads the HR Assistant's own logs, recomputes every metric, and commits the result to git.
pipeline
99 Nights: Roblox game build
An in-progress Roblox game built in Roblox Studio; no notes survive saying how much of it, if any, was AI-assisted.
prototype
Everyday Protein: scroll-driven product page prototype
A scroll-driven product page prototype for Unicity's Everyday Protein that scrubs a 243-frame image sequence on canvas, built with GSAP and Lenis.
prototype
Marketing Chatbot Prototype
A local Streamlit app that reads a company's website, drafts on-brand marketing emails, and runs a rule-based regional compliance check.
prototype
Kissimmee / Orlando Airbnb research site
An AI-assisted search and comparison of Airbnb options near Kissimmee, Florida, packaged as a browsable HTML page with a written research plan.
study
Shopify migration report
A written assessment of moving Unicity's storefront infrastructure onto, or off, Shopify.
study
Account-routing guard layer (personal vs. work)
Shell hooks and shims that work out whether a project is personal or work, then route Git, GitHub, Vercel and Google to the right account.
system
AI-agent pipeline safety guards (model + write gates)
Hooks that halt a pipeline run when a worker's model isn't declared, block database writes with no human watching, and gate “done” until tests pass.
system
Brain Vault
A company-knowledge system of curated, git-tracked notes pointing back to raw sources, so any AI can answer with citations instead of being retrained.
system
Database write-safety guard
A safety hook that spots any database command that would write, and forces a human confirmation before it is allowed to run.
system
Outbound-writing quality gate (stop-slop)
A hook that fires before any email or message is sent, scoring the text against a writing-quality rubric so nothing obviously AI-written goes out.
system
Project Brain Standard (project-brain-kit)
A reusable in-repo documentation standard, a wiki plus automated lint and personal-data checks, installed in every new AI department project before its first feature.
system
unicity-starter skills
A starter kit of three Claude skills, context lookup, status reporting, decision support, that any Unicity department can install for itself.
system
Department Engagement Method
The playbook the AI department uses when it embeds with a non-engineering team: fixed questions, a discover-record-present process, and always exactly two proposals.
workflow
Onboarding Automation control room
A coordinator for new-hire onboarding at Unicity: when a hire lands in ADP it prepares IT, badge, photo, calendar and credit steps for HR to approve.
workflow
PCC v4: greenfield redesign proposal
A written plan, not yet built, to rebuild the Product Command Center as a cloud control room with a lighter always-on daemon.
workflow
How I got here
In order
01Design
02Marketing
03Ecommerce and payments
04Product
05AI
Nutrigold. Studies Weekly. Ridgway. Unicity International.
Unicity International, 2022 to now
| Checkout approval, EU and Japan | 64% | to | 94% |
|---|---|---|---|
| Product launches per year, counted by market | 3 | to | 40 |
| Fraud rate on those checkouts | about 30% | to | under 1% |
These three are from my own record of the work.
Off the clock
Elsewhere
Father to Astra. Husband to Mariko.
About ten countries a year. Space and technology are the two subjects I never get tired of reading about.
How this site was made
By the agents it describes
I directed agents to build this page. One set swept my machine and my Grok workspace and listed every agent it could find. A second set audited the first, opened the files behind each claim, and cut my own headline from "hundreds" to the 76 on this page.
Four designs were then built from the same audited file, and three judges scored them. This page is the merge they asked for. Every number on it is computed from the audited file when the page is built, so it cannot drift from the data.
The thread ends here.
If any of this is useful to you, or you think one of these numbers is wrong, I would like to hear it.
Every number on this page was counted from the audited file when the page was built. 76 agents in the register, 63 of them running.
Start a conversationRegister current at 18 September 2026
On LinkedIn