AGENT SKILL

retro

Use when the user asks for a retrospective at the end of a work session on a Godot game.

License 许可
Apache-2.0
所属仓库
nick-vas/hyperdot
最近更新
2026-11-13
Stars
★ 1
Forks
0

01 前言 Front Matter

name
retro
description
Use when the user asks for a retrospective at the end of a work session on a Godot game.
disable-model-invocation
true

03 SKILL.md 正文

Retro: lessons from this session

Curated lessons make agents better; lessons an agent writes for itself, unchecked, make them worse (SkillsBench measured -1 to -11 points for self-written skills against +16 for curated ones). So every lesson here needs evidence, the user's approval, and, for kit changes, a measured result. Never write a new skill from scratch.

1. Gather evidence (read, do not summarise from memory)

  • The run history: .godot/hyperdot-history.jsonl in the project. Error codes and failed runners that recur.

  • The handoff log: .godot/hyperdot-handoffs.jsonl (or handoffs in status -Json), what the subagents were asked and answered. A FAIL verdict followed by "done" with no passing handoff or test run after it, or the same check asked again and again, is evidence.

  • This conversation: where the user corrected you, where you retried a command, where a fix was reverted.

  • Auto memory: MEMORY.md in this project's Claude Code memory folder, if it exists.

  • git log --oneline -30 and reverted or amended work.

  • docs/lessons.md in the game, so nothing is recorded twice.

  • docs/kit-notes.md in the game: one line each time the kit got in the way. Each line without a pull request link is a kit candidate (step 3), with the line itself as its evidence.

  • For a whole-project post-mortem: the people (or agents) who built it, over a playthrough recording or a playthrough log, not memory. Take each account separately before merging, and record disagreements instead of averaging them: "Sometimes those contradictions are because one person is wrong. But sometimes, everyone is correct." (Yusaku Yamamoto, Untold History).

  • The team's shape and its recurring jobs. Who did what this time (the user, agents, subagents, scheduled jobs), and the jobs that come back every week or release: the build, playtests, the store page, the board, backups, answering players. Ask the user: "What recurring job has no owner?" A job nobody owns is a lesson with its evidence (the last time it was done, from git or the history). Note the answer in docs/lessons.md, the job and who now owns it.

2. Write each lesson as one line with its evidence

<what to do differently> - evidence: <log line, error code with count, or the user's words>. Drop anything without evidence, anything Claude would do anyway, and project overviews (they do not help agents and cost tokens).

3. Sort them

  • About this game (its scenes, its rules, its quirks): append to the game's docs/lessons.md.
  • About Godot or the kit (true for any game): a candidate kit change. Name the smallest place it fits: one row in godot-pitfalls, one line in an existing skill, a doctor check, a test, a script fix. Not a new skill.

4. Ask the user

One multiple-choice question per kit candidate: the exact proposed text or diff, the evidence, and the options "Open a kit PR", "Keep it game-only", "Drop it". Nothing is changed without a choice.

5. Prove a kit change before proposing it

Work in the kit clone (the folder above ci/godot/hyperdot.ps1; AGENTS.md names it). Never on main:

git -C <kit> switch -c feature/<six random lowercase letters or digits> origin/main
# make the change, then:
pwsh <kit>/ci/godot/tools/Test-Skills.ps1
pwsh <kit>/tests/Test-SkillFixtures.ps1

If the change touches a skill, agent or hook, run the kit's evals before and after (from the kit folder): claude plugin eval . --runs 3 --scaffold --trust-plugin --allow-tools Bash Edit Write --no-publish (the flags of the kit's .github/workflows/evals.yml). Put both score tables (with, without, delta, cost) in the PR description. Do not open the PR if the validators fail or the score went down; tell the user instead.

Commit as the user, with no AI attribution, push the branch and open a pull request (gh pr create when gh is installed; otherwise print the branch and the compare URL). Never merge it.

6. Clean up

Remove promoted lessons from auto memory, so they are not loaded twice. Add each kit PR's link at the end of the docs/kit-notes.md lines it answers; delete lines the user chose to drop. Report: lessons recorded, PRs opened with links, candidates dropped and why, kit files committed or left (step 7).

7. Leave the game clean

The kit writes files into the game (AGENTS.md, CLAUDE.md, docs/kit-help.md, .claude/settings.json, the workflow and the like), and a fresh clone lacks them until they are committed. Read what git sees as new or changed:

pwsh <kit>/ci/godot/hyperdot.ps1 doctor -Json -Quiet -Project <game>

uncommittedKitFiles[] in the result is empty: say so and stop. Otherwise ask one multiple-choice question that lists the files, marks each with machinePath: true as naming a path of this machine, and offers "Commit them on a bugfix/ branch", "Keep them local" and "Leave them". Commit only the files without a machine path, plus any the user names; commit as the user with no AI attribution and do not push. Keep a file local with a line in .git/info/exclude. Never commit .claude/settings.local.json. To tell the kit's maintainers, doctor -ReportIssue shows the issue and files it only after a yes: run it when the user asks, not by yourself.

04 同仓库更多 Skills