← All projects

Flagship · AI governance

266 skills, one governance layer: agents that log, gate and reflect

Every edit passes a gate: agent, pre-tool gate, edit, stop hook, lesson store, next session.
Every edit passes a gate: agent, pre-tool gate, edit, stop hook, lesson store, next session.
Lesson lifecycle: logged the same turn, ranked while open, resolved only with pasted evidence.
Lesson lifecycle: logged the same turn, ranked while open, resolved only with pasted evidence.
The planning standard: discovery, one clarifying touchpoint, frozen scope, tests first, build waves, QA, live check.
The planning standard: discovery, one clarifying touchpoint, frozen scope, tests first, build waves, QA, live check.
  • AI governance
  • Internal system

Context

Several agent runtimes shared one working tree. Policy lived in markdown and drifted silently, and two core files were wiped by an unguarded writer before anyone noticed. Prompt instructions were not holding.

The decision

Keep writing softer standard operating procedures, or make governance mechanical: pre-tool and stop hooks that refuse the edit, tests that prove a guard fires, and a database as the only source of truth?

What was built

266 skill implementations, each with a companion guide kept in step by a drift check. A SQLite lesson store with severity ranking and recurrence tracking. Pre-tool and session-end hooks, concurrent-session and destructive-git guards, a sixteen-section planning standard with reflection before and after every wave, and an interactive dashboard generated for every architecture.

What it demonstrates

Governance you can test rather than read. Parity across runtimes. The claim of AI without tool sprawl backed by a mechanism, not a policy document.

Reconstructed with synthetic data.

  • Role

    Architect and operator

  • Stack

    Claude Code, Codex, Gemini, Python, SQLite, git hooks

  • Duration

    February to September 2026

  • Outcome

    266 skills, 110 recurring failure classes tracked to closure

See the build, then book a call to talk through yours. Every engagement ends with a handover your team can run.

See the services Discuss this workflow