One Repo Per AI Skill: A skill.md + logs + memory.md Layout for Agent State
Weary_Razzmatazz5077 · reddit · 2026-09-29
A Reddit user shares how he organizes AI skills: one repo per skill with three parts.
- skill.md: static instructions, rarely changes — the reusable core.
- logs/: every agent+skill session gets logged. His GitHub runner skill's log, for example, records how local runners were set up and for which repos, so months later he can ask the agent to recall what was done.
- memory.md: per-instance preferences and constraints. His research skill stores rules like "for repo ABC, only touch these websites"; the runner skill keeps local credentials there so setup can run automatically. He deliberately keeps this out of skill.md so reusable instructions don't get contaminated by one project's constraints.
The loop: logs update automatically after each run; memory and skill files only change when he gives feedback; everything is scoped to the skill — nothing touches it when the agent isn't invoked.
His open question: most setups stash this state in the project repo's own config/skill files, but memory is about the skill's subject matter, not the repo — the wrong layer of abstraction.
More from coding & agent
- Open-Source P(doom) Music Video Fully Built With Claude Code — StefanoGogioso · 2026-09-30
- Twin launches unlimited agent plans from $29/mo, drops credit system entirely — socialwithaayan · 2026-09-30
- Celesto Launches GitHub Actions Runners, Claims 12x Cheaper Than GitHub — aniketmaurya · 2026-09-30
- Wasmer's Pi runs AI agents unmodified in the browser and on iPhone via WebAssembly — JosephJacks_ · 2026-09-30
- RecursiveMAS: Agent swarms share latent thoughts like looped Transformers, cutting tokens by 75% — james_y_zou · 2026-09-30
- When code review needs memory: adding persistent context with Hindsight — Ambitious_Menu1385 · 2026-09-30