codex-skills

Execution-first Codex skills for repository cleanup, long-running agent work, and low-noise verified progress.

中文文档 · Contributing · ai-repo-cleanup · long-run-execution

codex-skills is a small open-source repository of practical Codex skills built for one narrow goal: make execution-oriented agent work more reliable, more reusable, and less noisy.

The current center of gravity is:

repo cleanup -> verified execution -> low-noise continuity

If a skill does not help Codex produce clearer execution, safer cleanup, or stronger verification, it probably does not belong here.

Overview

• execution-first Codex skills
• verification before claims
• repository hygiene by default
• low-noise outputs over analysis theater
• one repository, multiple focused skills

Why This Exists

Many prompt collections look impressive but are hard to reuse, hard to maintain, and too soft when real work begins. codex-skills is intentionally narrower.

It focuses on skills that improve real execution behavior, such as:

• slimming AI-generated repositories without damaging active contracts
• keeping long tasks moving through verified slices instead of drifting into handoff theater
• preventing tools and intermediate outputs from polluting user repositories
• turning repeatable execution patterns into reusable Codex skills instead of rewriting prompts every time

At A Glance

Skill	Responsibility
ai-repo-cleanup	surface delete-ready, high-probability, and backlog cleanup candidates in AI / agent repositories with explicit missing proof
long-run-execution	keep Codex executing through verified slices until the next real milestone instead of drifting into summaries, handoffs, or fake blockers

Included Skills

ai-repo-cleanup

A deletion-oriented cleanup audit for AI / agent codebases that challenges what should be removed, merged back, narrowed, retired, or archived before edits.

Best for:

• dead code suspicion
• false-alive tests
• weak helper/support splits
• duplicate implementation or duplicate intent
• wrappers, adapters, compat layers, and other support seams
• stale docs/history noise and non-core surface growth that should not dominate the active repo

Key characteristics:

• maximizes candidate discovery before verdict compression
• produces an execution package, not a vague audit report
• names the missing proof and fastest next check for non-ready candidates
• supports zero-action compression when no cleanup should happen
• requires tool artifacts to stay outside the repo by default
• explicitly forbids GitNexus-style instruction-file pollution such as AGENTS.md / CLAUDE.md

long-run-execution

A skill for keeping Codex moving through long tasks without collapsing into midpoint summaries, fake handoffs, blocker inflation, or user-question loops for locally discoverable context.

Best for:

• uninterrupted execution
• verification after each slice
• progress to the next real milestone
• reducing “next agent can continue” style drift
• audit-then-repair tasks that should not stop at issue lists

Key characteristics:

• locks the current slice around target, boundary, and proof
• adds an explicit continue / stop decision ladder
• makes verification part of execution cadence
• classifies side issues and returns to the active slice
• treats handoff-as-substitute-for-work as failure
• keeps continuity short and reality-based

Typical Use Cases

Repository cleanup

Use ai-repo-cleanup to audit this repo and generate a cleanup list.

Expected outcome:

• delete-ready and high-probability items are front-loaded
• tool outputs stay out of the repo
• a coding agent can act on the result directly

Long execution session

Keep going until this feature is implemented and verified.

Expected outcome:

• Codex does not stop at a midpoint summary
• each slice is verified before moving on
• continuity reflects actual progress instead of speculative notes

Design Principles

codex-skills prefers skills that are:

• execution-first
• verification-driven
• small in scope, strong in rules
• useful across repositories
• careful about repository hygiene

Non-goals:

• building a giant skill framework
• packaging every personal prompt into a public repo
• adding heavy setup just to make simple skills look productized

Repository Layout

codex-skills/
  README.md
  README_CN.md
  CONTRIBUTING.md
  LICENSE
  scripts/
    install.sh
  skills/
    ai-repo-cleanup/
      README.md
      SKILL.md
      references/
      scripts/
    long-run-execution/
      README.md
      SKILL.md
      long-run-review-template.md

Quick Install

Install all bundled skills:

cd codex-skills
./scripts/install.sh

Install only one skill:

cd codex-skills
./scripts/install.sh ai-repo-cleanup

Install to a custom Codex home:

CODEX_HOME=/path/to/custom/codex ./scripts/install.sh

Default target:

$HOME/.codex/skills

Development

Edit skills in place under:

skills/<skill-name>/

Reinstall locally after changes:

./scripts/install.sh

Test without touching your real Codex home:

CODEX_HOME=$PWD/.tmp/codex-home ./scripts/install.sh ai-repo-cleanup

Roadmap

Possible future additions:

• more execution-oriented skills from the same workflow family
• richer examples and before/after outputs
• lightweight release notes for behavior changes

The repository will stay intentionally small.

🏠 Commuity （Acknowledge）

First shared on LINUX DO — where possible begins.

License

MIT