Aby Thannikal Joseph
Menu

AI product team

How my AI team builds this site

One coordinator agent, five specialist agents, automated checks and me. Agents made implementation cheap; judgement, verification and accountability still need an owner.

Who’s who

hand-off way back checked

Choosing the direction

Step 1 of 4

Two independent audits, one decision that stays with me.

Two AI tools audit the old site independently. One recommends refactoring the Gatsby site, the other a rebuild.

The four workflows, step by step

Taken from this site’s repository: the rules every agent follows, the tracker, the agent definitions and the article runbook.

1Choosing the direction

Two independent audits, one decision that stays with me.

  1. Independent AI reviews → MeTwo AI tools audit the old site independently. One recommends refactoring the Gatsby site, the other a rebuild.
  2. MeThe disagreement makes the real question explicit: what would we actually be keeping? I decide to rebuild.
  3. Me → Coordinator agentI brief the coordinator agent, which plans and implements the work.
  4. Coordinator agent → RepositoryBefore any code: AGENTS.md, the tracker and an inventory of every live URL.

2A work package

The loop every foundation package ran, from F00 to F16.

  1. Coordinator agent → RepositoryThe coordinator writes a package into the tracker: dependencies, acceptance criteria and the evidence to collect.
  2. Coordinator agentIt implements the package, one at a time, switching between the architect and engineer hats.
  3. Coordinator agent → Automated checksnpm run check validates types and content. npm run verify builds the site and checks the generated pages.
  4. Automated checks → Coordinator agentA failure sends the package back. A broken link, missing alt text or a leaked draft fails the build.
  5. Coordinator agent → Automated checksFixed and re-run, the checks pass. They were proven first: seven injected faults gave nineteen reported failures.
  6. Automated checks → RepositoryEvidence goes against each criterion, and the commit names its package: “F13: add dist regression checks and preview build”.
  7. Repository → MeI review the rendered pages, not just the report.
  8. Me → Push · merge · deployPush, merge and deploy stay with me. Until I approve, the package says “blocked”, not “done”.

3When checks aren’t enough

Package F20: the look was never written down as something to check.

  1. Coordinator agent → Automated checksThe first foundation build passes every check.
  2. Automated checks → MeI review the rendered site. It had replaced the site’s look with a bare text layout: “heavily simplified”.
  3. Me → Coordinator agentMy feedback reopens the work as package F20. The coordinator records it as its own error.
  4. Coordinator agentIt restores the visual design: hero, contents, margin notes, infographics, colours and dark mode.
  5. Coordinator agent → RepositoryA new rule goes into AGENTS.md: no unrequested visual redesigns, and screenshots at 390 and 1440 px before reporting UI work.
  6. Repository → MeRules follow decisions. When I later ask for a full redesign, the rule is reworded.

4Writing an article

The main session conducts five agents and two skills. Files are the handoff.

  1. Me → Coordinator agentI bring notes, or ask what to write about.
  2. Coordinator agent → topic-scoutOptional: topic-scout checks recent releases and proposes five to eight ideas, without writing any files.
  3. topic-scout → MeI pick one and add what I know. My notes become the spine, kept word for word.
  4. Coordinator agent → article-researcherarticle-researcher turns the notes into a sourced brief: facts, counterpoints, structure and questions only I can answer.
  5. article-researcher → RepositoryThe brief is a file: docs/research/<slug>.md.
  6. Repository → MeGate 1: I answer its questions and approve the argument.
  7. Coordinator agent → article-writerarticle-writer drafts in my voice. Nothing invented: missing facts become visible placeholders.
  8. article-writer → RepositoryThe draft lands as index.mdx with draft: true and figure placeholders.
  9. Coordinator agent → SkillsThe main session runs the humanizer skill on the wording, because the writer has no skill access.
  10. Coordinator agent → diagram-builderdiagram-builder loads the diagram-design skill and draws each figure in the site’s style.
  11. diagram-builder → RepositoryFigures are exported as SVG into the article folder and placed in the draft.
  12. Coordinator agent → technical-reviewertechnical-reviewer checks claims against official docs, runs the code and redoes the arithmetic. It never edits the article.
  13. technical-reviewer → article-writerMust-fix items go back to the writer. A review loop runs at most twice, then the disagreement comes to me.
  14. Repository → MeGate 2: I check placeholders and tone, and sign off.
  15. Me → Automated checksPublish gate: lint:content, check and verify.
  16. Automated checks → Push · merge · deployNothing is published, committed or pushed without my say-so.