Projects

Project 5 · Codify the Body

⚠️ Throwaway repo. Set a limit first. Practice the skill on a disposable clone before you ever point it at something real, and cap the number of loop…

⚠️ Throwaway repo. Set a limit first. Practice the skill on a disposable clone before you ever point it at something real, and cap the number of loop invocations you run while iterating.

Difficulty	Time	Concepts	Loop type
Medium (T2/T3)	60 min	skills, the cold-start problem, rules vs. procedure	conditional loop with a codified skill body

The scenario

You've been hand-writing the same three-paragraph prompt every time you ask an agent to add a changelog entry for a merged PR: which file to touch, the heading format, the tone, which PRs to skip (drafts, reverts). It works, but every session starts cold — the agent re-derives "how we do changelogs here" from scratch, and it drifts a little each time. This project turns that repeated prose into a skill: a small, named, reusable procedure the loop's body invokes instead of re-explaining itself every beat.

What you'll build

Take a real repository with a CHANGELOG.md and a real PR history (your own project, or any OSS repo with a maintained changelog you can fork). Extract the changelog convention into a SKILL.md (both tools' shape), then wire a schedule loop's body to invoke it instead of an inline prompt:

The skill: changelog-entry/SKILL.md — heading format, which PR fields to pull, what to skip, one worked example.
Heartbeat: schedule, once per merge or once daily.
Body: for each new merged PR since last beat, invoke the skill to draft (not commit) a changelog entry into a scratch file — never CHANGELOG.md directly at L1.
Spine: changelog-state.md — last PR number processed.

# Claude Code
# .claude/skills/changelog-entry/SKILL.md exists with the convention + one example
> /loop 1d For each PR merged since changelog-state.md's last-seen number, use the
  changelog-entry skill to draft an entry into changelog-draft.md (never edit
  CHANGELOG.md directly). Skip drafts and reverts, per the skill. Update
  changelog-state.md. Log one line.

# OpenCode — AGENTS.md points at the same convention as a referenced doc
opencode run "For each PR merged since changelog-state.md's last-seen number, follow
  docs/conventions/changelog-entry.md to draft an entry into changelog-draft.md.
  Skip drafts/reverts. Update changelog-state.md. Log one line." --schedule daily

Done when (provable stop)

A SKILL.md (or equivalent AGENTS.md-referenced doc) exists, versioned, with the convention spelled out and one worked example.
Two separate loop beats, run in fresh sessions, produce changelog entries in the same format without you re-explaining the convention in the prompt either time.
changelog-draft.md never wrote to the real CHANGELOG.md — this project stays L1.
changelog-state.md correctly tracks the last PR processed across a restart.

Going deeper

This is the exact cold-start problem Step 9 names, and the exact fix starters/changelog-drafter/ ships as a finished kit. Compare your SKILL.md against that kit's once you're done — the difference between a fat, repeated prompt and a thin prompt plus a named skill is the whole lesson of this project.

When it goes wrong

Symptom	Cause	Fix
Every entry looks slightly different	The skill was read as inspiration, not procedure	Make the skill's heading/format non-negotiable; keep the "one worked example" verbatim-close
Skill grows into a second prompt just as long	Trying to encode every edge case up front	Start thin; add an edge case to the skill only after it actually occurs, not speculatively
Draft entries pick up PRs already processed	`changelog-state.md` not read at beat start	Read the spine before touching the PR list — never re-derive "what's new" from scratch

Sources: the cold-start problem and SKILL.md shape from Step 9, Panaversity's Loop Engineering: A Crash Course (S1) and Agentic Coding Crash Course (S2). Full attribution: resources/sources.md.