antfu-simplify — Claude Code skill: what it does, how to install
antfu-simplify — a skill for Claude Code and Codex that cuts working tree diffs down to minimal working code through systematic ablation.
MIT 312 installs ★ 6,092 updated 2026-10-23
npx skills add antfu/skills --skill antfu-simplify -a claude-codeWhat antfu-simplify does
The skill evaluates pending changes from git diff HEAD or the PR merge base to strip speculative code. It tests functions, options, extra types, and new dependencies one by one, inlining or deleting pieces that lack clear justification. Surrounding project files remain untouched, leaving existing behavior unchanged.
Every ablation step gets verified with the project test runner, typechecker, and linter. Failing tests trigger an immediate rollback, confirming that piece is essential. Work finishes with a markdown summary table detailing every cut, alongside git diff --stat metrics without creating unrequested git commits.
What it helps with
- cut bloated PR diffs down to minimal changes
- inline single-use helper functions and shallow wrappers
- verify added branch conditions with project tests
- remove unneeded defensive guards and speculative generic types
Example prompts for the agent
- Simplify the current PR changes with an ablation study
- Slim down the working tree diff and keep only proven code
- Clean up the diff before submitting for review
How to install antfu-simplify
1Skills CLI (Vercel)
npx skills add antfu/skills --skill antfu-simplify -a claude-code2Manually: put the files in place
~/.claude/skills/antfu-simplify/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/antfu-simplify/
1Skills CLI (Vercel)
npx skills add antfu/skills --skill antfu-simplify -a codex2Manually: put the files in place
.agents/skills/antfu-simplify/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add antfu/skills --skill antfu-simplify -a cursor2Manually: put the files in place
.cursor/skills/antfu-simplify/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install antfu/skills skills/antfu-simplify2Manually: put the files in place
.github/skills/antfu-simplify/SKILL.md1Skills CLI (Vercel)
npx skills add antfu/skills --skill antfu-simplify -a gemini-cli2Platform CLI
gemini skills install https://github.com/antfu/skills --path skills/antfu-simplifyManually: put the files in place
.devin/skills/antfu-simplify/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add antfu/skills --skill antfu-simplify -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about antfu-simplify
How to install antfu-simplify in Claude Code?
Run npx skills add antfu/skills --skill antfu-simplify -a claude-code. The file is placed at ~/.claude/skills/antfu-simplify/SKILL.md.
What does antfu-simplify do?
It examines incoming diffs, attempts targeted deletions of new code components, and verifies test outcomes to discard unnecessary additions.
What is required for antfu-simplify to run?
Working test, typecheck, or lint commands in the target repository, as each deletion step relies on those checks for verification.
Source text
SKILL.md · MIT
--- name: antfu-simplify description: Simplify the changes of the current PR or working tree with an ablation study - remove each piece of the diff and keep only what is proven needed - so the result has just-right abstraction, minimal clean code, and stays robust and universal. Use when asked to simplify, slim down, clean up, or reduce the current changes or PR. metadata: author: Anthony Fu version: "2026.10.31" --- # Simplify Changes Review the **current changes** (PR branch or working tree) and cut them down to the smallest correct diff, the way a skilled developer would. The method is an **ablation study**: remove each piece, check whether anything that matters breaks, and keep only what earns its place. Goal: clean, minimal code with the **just-right abstraction**, that is also **robust** (handles real edge cases) and **universal** (solves the general case, not only the one that triggered the change). ## Hard Rules - **Behavior stays the same.** Simplify only. Do not add features, change public APIs, or fix unrelated issues. Report unrelated findings instead. - **Scope is the diff.** Only touch code the change introduced or modified, plus code that must change to remove it. Do not refactor the surrounding codebase. - **Every removal is verified.** Run the repo's tests, typecheck, and lint after each batch. If a check has no coverage for the removed piece, reason about it explicitly or keep it. - Match the surrounding code's style, naming, and comment density. ## Workflow 1. **Get the diff.** Pick the base: PR target branch merge base (`git merge-base HEAD origin/<default>`), else `git diff HEAD` for uncommitted work. Read the full diff and the files around it. Note the goal of the change in one sentence; it is the yardstick for everything below. 2. **Inventory the pieces.** List each unit the diff adds: function, parameter, option, branch, type, file, dependency, abstraction layer, comment, test. See [Ablation Study](#ablation-study). 3. **Ablate.** For each piece, try removing, inlining, or merging it. Decide keep or cut using [Keep vs Cut](#keep-vs-cut). 4. **Apply** the cuts in small steps. After each batch run focused checks (tests, typecheck, lint). Revert a cut that breaks behavior; that piece is now proven needed. 5. **Check the result.** Re-read the final diff as a reviewer. Is each remaining line explained by the one-sentence goal? Is it still robust and general? 6. **Report.** Show a short table of what was cut or kept and why, plus the diff size before and after (`git diff --stat`). Do not commit unless asked. ## Ablation Study For every piece in the diff, ask: *if I delete this, what concretely breaks?* | Piece | Ablation | Keep if | |-------|----------|---------| | New function / helper | Inline it at the call site | Called from 2+ places, or the name explains a non-obvious step | | Wrapper / abstraction layer | Call the underlying thing directly | It hides real complexity or a likely-to-change boundary | | Parameter / option / config | Hardcode the only used value | A real caller needs a different value now | | Type / interface / generic | Use the inferred or existing type | It is exported, or inference gives a worse or wrong type | | Branch / guard / try-catch | Delete it and run tests | It guards a reachable case (real input, real failure), not an impossible one | | New file / module | Merge it into its only consumer | It has multiple consumers or a distinct responsibility | | Dependency | Replace with a few lines or an existing dep | It handles hard edge cases you would otherwise get wrong | | Duplicate logic | Reuse an existing util in the repo | No equivalent exists (search first) | | Comment | Delete it | It explains *why* (constraint, workaround, surprising choice), not *what* | | Test | Remove it and see what is unprotected | It covers behavior or a regression nothing else covers | | Changed lines unrelated to the goal | Revert them | Never keep drive-by edits | ## Keep vs Cut Cut when the piece is: - Speculative: built for a future need nobody has (extra options, hooks, plugin points, "just in case" generics). - Redundant: repeats what the language, framework, types, or an existing util already provides. - Defensive against the impossible: null checks on values the types or callers guarantee, fallbacks that never run, catch blocks that swallow errors. - Indirection without payoff: one-use helpers, pass-through wrappers, classes where a function works, factories that build one thing. - Noise: commented-out code, narrating comments, unused imports and exports, formatting-only churn. Keep (or even add) when the piece is: - **Robustness:** handles real edge cases (empty input, concurrency, cleanup, cross-platform paths, error propagation with context). - **Universality:** solves the general problem instead of special-casing the reported one. If the fix hardcodes a name, path, or value from the bug report, generalize it. - **Right-sized abstraction:** the third repetition of something gets extracted; the first and second usually do not. A good abstraction removes more code than it adds. If an abstraction is too small, merge; if a function does several unrelated things, split. The target is neither fewest lines nor most layers, it is the fewest concepts a reader must hold. ## Smells to Look For - A fix that patches the symptom at the call site instead of the cause. - Special cases (`if (name === 'foo')`) where a general rule would cover them. - Two code paths that do the same thing with small differences; unify them. - State, flags, or caches that can be derived from existing data. - Long parameter lists or option bags where most values are never varied. - Types duplicated by hand when they can be derived (`typeof`, `ReturnType`, `Parameters`, utility types). - Hand-rolled utilities that the repo, runtime, or an existing dependency already has. - Tests that assert implementation details or mirror the code line by line. ## Report Format ```md | # | Piece | Action | Why | |---|-------|--------|-----| | 1 | `parseOptions()` in src/cli.ts | inlined | one caller, 3 lines | | 2 | `retry` option | removed | no caller passes it; default is enough | | 3 | `try/catch` in fetchData | kept | network errors reach users; test covers it | Diff: +142 -8 -> +61 -8. Tests, typecheck, lint pass. ``` List anything noticed but out of scope (unrelated bugs, wider refactors) separately, and leave it alone.
© Anthony Fu · MIT · antfu/skills