Growing Organisational Context with AGENTS.md, Skills, Hooks and cron

TIMEWELL Inc. · Published 2026-09-12 · Ryuta Hamamoto

Source: https://timewell.jp/en/columns/ai-native-context-operations-agents-md-skills-hooks-cron

Subtitle: AI-Native New Business Development, Part 4 — Recurring Work Is Decided by the Context Machinery

Summary (verbatim-leaning)

Weekly reports, monitoring, organising interview records. When you hand recurring work to an AI agent, what decides performance is not the cleverness of the prompt but the machinery that gives the agent its context. This article lays out four parts — AGENTS.md (facts read every time), Skills (procedures read only when needed), Hooks (enforcing what a document cannot), and cron (scheduling) — plus how to grow internal context as a knowledge-graph-shaped dataset. Running this as an organisation comes down to three disciplines: keep updating the context, exclude what lies outside it, and raise the frequency of skill updates.

Key claims (verbatim-leaning English extract)

• Recurring agent work fails by third run usually from stale / cluttered / missing organisational context — not from the prompt.
• Context engineering (Lütke / Karpathy): fill the window with just the right information; Anthropic “context rot” — too many tokens degrade recall.
• Organisational reread: tacit assumptions (customer address forms, forbidden product names, approval routes, last year’s failure) must be “just the right amount, every time”.
• Four parts:
  • AGENTS.md / CLAUDE.md — standing facts/conventions/prohibitions; always read; aim under 200 lines; concrete verifiable instructions; import @AGENTS.md into CLAUDE.md.
  • Skills (SKILL.md) — procedures + optional scripts; load only when used; move procedures out of always-read docs.
  • Hooks — PreToolUse etc. mechanically stop actions; docs are context not enforced configuration.
  • cron / schedules — /loop, Desktop scheduled tasks, cloud Routines; choose cloud vs local by file/access needs; decide stopping line first for unattended work.
• When to add to AGENTS.md: same mistake twice; review catch the agent should have known; same correction typed twice; new teammate would need it — otherwise do not write.
• Internal knowledge as LLM-maintained Markdown wiki (Karpathy LLM Wiki): raw sources immutable; wiki owned by LLM; schema in AGENTS.md/CLAUDE.md; periodic lint for contradictions / stale / orphans.
• Three organisational disciplines:
  1. Keep updating — agent drafts, human reviews via PR; triggers predefined.
  2. Exclude — never write what code/docs can derive; path-scoped rules; agent proposes deletion candidates, human decides.
  3. Raise skill-update frequency — weekly lint/grade; unused skills fix description or delete.
• Worked Friday interview-extraction example: 20-line AGENTS facts + skill with Part-1 prompt + helper PII script + two hooks (masking gate; outbound approval) + Friday 17:00 cron; Monday 10-minute human diff review.

Remainder

Full original English column (footnotes, WARP programme links, series cross-links): see source_url.