minimalist-skill

◇ Minimalist

Every line your agent didn't write is a line nobody has to fix at 3 a.m.

Ask your agent for a date picker. It installs a library, writes a wrapper component, a stylesheet, and a paragraph about timezones you didn't ask for.

Minimalist makes it stop and check first: does this need to exist → is it already in the codebase → does the stdlib have it → does the platform have it → is it one line → only then write new code. Every step it skips, it says so — and now it writes that down too, not just says it once and forgets.

<!-- descent step 4: the platform already has one -->
<input type="date">

The part nobody else ships: receipts

Every "less code" tool tells you to trust the vibes. Minimalist keeps a ledger. Every time the descent rejects a dependency, an abstraction, or a speculative config option, it's logged — not just mentioned once in chat and gone:

$ node scripts/log-rejection.js --report
3 rejected-scope entries in ~/.minimalist/ledger.jsonl

- [platform] date picker library -> <input type="date">
- [yagni] config flag for a value that never changes
- [stdlib] custom debounce helper -> AbortController + setTimeout

By step:
  platform: 1
  yagni: 1
  stdlib: 1

/minimalist-gain reads that file, not the model's memory of the last five minutes. Cross-session, inspectable, grep-able. If your tech lead wants to know what the agent talked itself out of building this sprint, that's a ten-second cat, not a re-prompt and a shrug.

Not lazy — accountable

Minimal never means unsafe. These are never cut, at any level: input validation, authN/authZ, rate limits, error handling on fallible ops, resource cleanup, tests. If a shorter version drops one of these, the shorter version is wrong — and it doesn't ship.

Install

Node.js 18+ on PATH (for the two lifecycle hooks; skills work without it).

Claude Code

/plugin marketplace add DivyeshJayswal/minimalist
/plugin install minimalist@minimalist

Desktop app: Customize → personal plugins → Add from repository → repo URL.

Codex

codex plugin marketplace add DivyeshJayswal/minimalist

/plugins → install Minimalist → /hooks → trust its two hooks → new thread.

GitHub Copilot CLI

copilot plugin marketplace add DivyeshJayswal/minimalist
copilot plugin install minimalist@minimalist

Commands are namespaced: /minimalist:minimalist ultra.

Gemini CLI

gemini extensions install https://github.com/DivyeshJayswal/minimalist

Antigravity CLI

agy plugin install https://github.com/DivyeshJayswal/minimalist

Reuses the Gemini extension; also ships .agents/rules/ for always-on rules.

Pi agent harness

pi install git:github.com/DivyeshJayswal/minimalist

OpenCode

{ "plugin": ["minimalist-skill"] }

Or from a checkout: { "plugin": ["./.opencode/plugins/minimalist.mjs"] }. OpenCode also auto-loads this repo's AGENTS.md.

Hermes

hermes plugins install DivyeshJayswal/minimalist --enable

OpenClaw — skills ship in .openclaw/skills/; point your skills path at the repo.

Cursor · Windsurf · Cline · Kiro · Devin · CodeWhale · Swival — rule files are pre-generated, copy the one you need: .cursor/rules/minimalist.mdc, .windsurf/rules/minimalist.md, .clinerules/minimalist.md, .kiro/steering/minimalist.md, .github/copilot-instructions.md, .devin-plugin/, or generic AGENTS.md.

Uninstall

node scripts/uninstall.js

Commands

command	does
/minimalist [lite|full|ultra|off]	set intensity (default full)
/minimalist-review	review a diff for bloat — and for missing guards
/minimalist-audit	ranked deletion candidates across the codebase
/minimalist-gain	the ledger, read back — measured, never invented
/minimalist-help	the whole tool in 15 lines

lite advises, full enforces, ultra ships the diff with ≤3 lines of prose. The guardrail holds at all three. stop minimalist / normal mode also works.

Numbers, reproduced by you

node benchmarks/run.js --repo <target-repo> --model haiku --n 4   # real run
node benchmarks/run.js --mock                                      # free pipeline check

Scores the git diff a real agent leaves behind, not a single-shot prompt: LOC, tokens, cost, wall time, plus a 6-task adversarial safety tier (path traversal, SQLi, token forgery, malformed input, rate limits, authz) that fails a run if a guard line is deleted and not restored. Mock runs are watermarked "do not publish" — only real runs are. Every benchmarks/arms/*.txt becomes an arm, so you can drop in another project's ruleset and compare directly, yourself, on your own repo. We don't publish a number we didn't run.

How it works

One skill file (skills/minimalist/SKILL.md) is the source of truth; every per-agent copy is generated from it (npm run mirrors), and CI fails on drift — across ubuntu, macOS, and Windows. For Claude Code/Codex/Copilot it rides two lifecycle hooks (UserPromptSubmit injects the ruleset and parses /minimalist switches; SubagentStart makes subagents inherit the level). For Gemini/OpenCode/Pi it loads as extension context. For rule-file agents it's a generated copy. Level state lives in ~/.minimalist/config.json, written atomically, merged never overwritten. Rejected scope lives in ~/.minimalist/ledger.jsonl, appended never truncated.

If another instruction-injecting skill is detected (output-style or code-minimal), minimalist yields prose-style decisions to it and keeps code volume for itself, or applies the stricter rule on overlap — never contradictory instructions.

Development

npm test              # 49 tests, node:test, zero dependencies
npm run mirrors       # regenerate per-agent rule copies from SKILL.md
npm run mirrors:check # CI drift gate
npm run bench:mock    # validate the benchmark pipeline for free

CI runs ubuntu / macos / windows × Node 18/20/22.

Write less. Prove it. Break nothing.