compound

星标0

分支0

更新时间2026年7月24日 11:16

Run the compound engineering step — a short retrospective after a piece of work that folds what was learned back into CLAUDE.md, the skills, or the hoard. Use at the end of every significant task.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

moritalous

moritalous/agentic-engineering-patterns

打开 GitHub 仓库查看创作者相关仓库

下载

Compound step

Agents follow instructions, and we can evolve those instructions over time. Every completed piece of work is an opportunity to make the next run better. Small improvements compound.

Procedure

Review the session. What was just accomplished? Look at the diff and the conversation.
Identify what worked and what didn't:
- A prompt, command, or approach that worked well and will recur.
- A mistake or dead end that a future session should be warned about.
- A trick figured out for the first time.
- Friction: something the project setup made harder than it should be.
Fold each learning into the right place:
- Recurring workflow guidance → CLAUDE.md.
- A repeatable multi-step procedure → a new or updated skill in .claude/skills/.
- A working code trick → a hoard/ entry (use the /hoard skill).
- Setup friction → fix it now if cheap (it's a refactor: ideal agent work).
Keep additions short. CLAUDE.md is loaded into every session; each line must earn its context cost. Prefer one sharp sentence over a paragraph.
Commit the documentation changes with the work they came from, or as an immediate follow-up commit.

If the session genuinely produced no reusable learning, say so and stop — do not invent filler.

同仓库更多 Skills

同仓库

git-rescue

moritalous/agentic-engineering-patterns

Sort out git messes and perform history surgery — undo/rewrite commits, resolve conflicted merges and rebases, recover lost code via reflog, find regressions with git bisect.

2026-07-240

hoard

moritalous/agentic-engineering-patterns

Save a newly-figured-out trick as a minimal working example in hoard/, so future sessions can reuse or recombine it. Use after solving something non-obvious for the first time.

2026-07-240

interactive-explanation

moritalous/agentic-engineering-patterns

Build a self-contained interactive or animated HTML page that explains how an algorithm or piece of code works. Use to pay down cognitive debt when a written explanation isn't enough.

2026-07-240

manual-test

moritalous/agentic-engineering-patterns

Manually exercise code beyond the automated tests — python -c for libraries, curl for APIs, browser automation with screenshots for web UIs — and document the results. Use after tests pass, before declaring work done.

2026-07-240

pr-ready

moritalous/agentic-engineering-patterns

Pre-pull-request checklist — verify the code works, the diff is reviewable, and the PR carries context and evidence of testing. Run before opening any pull request.

2026-07-240

proofread

moritalous/agentic-engineering-patterns

Proofread a document before publication — spelling, grammar, repetition, logical and factual errors, weak arguments, placeholder links. Reports issues; never rewrites the author's voice.

2026-07-240

name	compound
description	Run the compound engineering step — a short retrospective after a piece of work that folds what was learned back into CLAUDE.md, the skills, or the hoard. Use at the end of every significant task.

Compound step

Agents follow instructions, and we can evolve those instructions over time. Every completed piece of work is an opportunity to make the next run better. Small improvements compound.

Procedure

Review the session. What was just accomplished? Look at the diff and the conversation.
Identify what worked and what didn't:
- A prompt, command, or approach that worked well and will recur.
- A mistake or dead end that a future session should be warned about.
- A trick figured out for the first time.
- Friction: something the project setup made harder than it should be.
Fold each learning into the right place:
- Recurring workflow guidance → CLAUDE.md.
- A repeatable multi-step procedure → a new or updated skill in .claude/skills/.
- A working code trick → a hoard/ entry (use the /hoard skill).
- Setup friction → fix it now if cheap (it's a refactor: ideal agent work).
Keep additions short. CLAUDE.md is loaded into every session; each line must earn its context cost. Prefer one sharp sentence over a paragraph.
Commit the documentation changes with the work they came from, or as an immediate follow-up commit.

If the session genuinely produced no reusable learning, say so and stop — do not invent filler.