superskills
Less is more. A minimal coding harness, shipped as a Claude Code plugin: 4 skills, 2 hooks, ~418 always-on tokens. Codex and Aone Copilot are covered by a single install script.
中文文档
Why
Heavyweight harnesses made sense when models needed guardrails at every step: hard process gates, multi-stage reviews, forced TDD loops. As models get stronger, most of that scaffolding turns into overhead. What still compounds in value:
- Memory — learnings from past sessions (corrections, pitfalls, decisions) that no model can infer from code
- Conventions — a minimal, evidenced spec of how this project actually works
- Clarification — resolving the genuinely undecided parts of a request before coding
- A final test pass — verified behavior, without ritualizing the path there
superskills keeps exactly these four things and deletes everything else.
Does it actually help?
Measured A/B on the same tasks, same model (Sonnet 4.6), real end-to-end runs, deterministic graders — full methodology and per-check tables in docs/benchmark.md:
| Scenario | Baseline (pure model) | With superskills | Δ |
|---|
| Cross-session memory (3 team decisions persisted as learnings) | 20% | 100% | +80pp |
| Requirement clarification (ambiguous feature request) | 0% asked | 67% asked | +67pp |
| Final test pass (2 planted bugs in "just developed" code) | 40% — tests locked the bugs in | 100% — both fixed at root cause | +60pp |
| Convention adherence (rules scattered in docs) | 100% | 100% | 0pp, ~equal time |
| Control: HumanEval/0–9 verbatim | 10/10 | 10/10 | no regression |
The pattern: when the knowledge is one obvious read away in a tiny fixture, a strong model already behaves (S1, control). The gains appear exactly where superskills operates — knowledge that exists nowhere in the repo (memory), questions nobody asked (clarification), and bugs that fresh tests happily cement in place (test pass). Baseline runs wrote passing test suites around both planted bugs in 3 of 3 trials; the test skill fixed both at root cause in 3 of 3.
What you get
| Component | Kind | What it does |
|---|
superskills:discover | skill | Scans an existing project and generates minimal spec files: .superskills/conventions.md (≤80 lines), AGENTS.md, CLAUDE.md. Refreshes them when stale, folding hardened learnings into conventions. |
superskills:learn | skill | Persists durable learnings (user corrections, pitfalls + fixes, invisible decisions) to .superskills/learnings/. |
superskills:clarify | skill | Surfaces only the questions whose answers change the implementation, with recommended answers, then starts coding. |
superskills:test | skill | One full unit-test pass after development is done. Result-driven, no fixed process. |
| SessionStart hook | hook | Injects the learnings index into each session; reminds you when conventions drift >30 commits behind HEAD; suggests discover for projects with no AI specs. |
| Stop hook | hook | Auto-learning: when a session did real work (≥5 user messages and file edits), asks the model once — with full session context — to persist anything durable before stopping. |
Everything shows up in the /plugin panel with per-component token costs. Total always-on cost: ~418 tokens.
Project artifacts (committed to your repo)
.superskills/
├── conventions.md # single source of truth, ≤80 lines
└── learnings/
├── INDEX.md # one line per learning, auto-injected at session start
└── 2026-06-12-use-pnpm.md
AGENTS.md # ≤20 lines, points at .superskills/
CLAUDE.md # @AGENTS.md + @.superskills/conventions.md
Install
Claude Code (plugin, recommended)
/plugin marketplace add Mrlyk/superskills
/plugin install superskills@superskills
Or from the CLI: claude plugin marketplace add Mrlyk/superskills && claude plugin install superskills@superskills. Hooks register automatically with the plugin; nothing touches your settings.json.
Codex / Aone Copilot
git clone https://github.com/Mrlyk/superskills.git && cd superskills
./install.sh # autodetects ~/.codex and ~/.aone_copilot
| Tool | Skills | Hooks (auto-learning + injection) |
|---|
| Claude Code | plugin: /superskills:discover etc. | yes |
| Aone Copilot | ~/.aone_copilot/skills/ss-* | yes |
| Codex | ~/.codex/prompts/ss-*.md (custom prompts) | no — relies on AGENTS.md pointers |
./install.sh --tools claude remains available as a legacy settings-based install for environments without marketplace access. --uninstall reverses everything and preserves your own settings.
Then, in each project, run the discover skill once and commit the generated files.
How knowledge flows back in
