harness
/harness
Wyznacza etapy i zasady budowy oraz ulepszania harnessu Claude Code: skille, agenci, hooki i CLAUDE.md, z ewaluacją i kosztami.
- Autor
- mhmdsalameh ↗
- Repozytorium
- ai-native-harness ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 11.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 11.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Wyznacza etapy i zasady budowy oraz ulepszania harnessu Claude Code: skille, agenci, hooki i CLAUDE.md, z ewaluacją i kosztami.
Oryginalny opis (EN)
Stage map and operating rules for building or improving a Claude Code harness (skills, agents, commands, hooks, CLAUDE.md, scripts) with research, evals against a no-harness baseline, and cost optimization. Use at the start of any harness-building session, when the user asks to create, audit, evaluate, optimize or improve a harness, plugin, skill set or CLAUDE.md, or to decide which harness-builder skill to load next.
Jak działa
- Session start (every session)
- Stages
- hb commands
- Project folder
- Rules that hold in every stage
- Platform limits to plan around
Podgląd SKILL.md
---
name: harness
description: Stage map and operating rules for building or improving a Claude Code harness (skills, agents, commands, hooks, CLAUDE.md, scripts) with research, evals against a no-harness baseline, and cost optimization. Use at the start of any harness-building session, when the user asks to create, audit, evaluate, optimize or improve a harness, plugin, skill set or CLAUDE.md, or to decide which harness-builder skill to load next.
---
# Harness builder — stage map
A harness is everything around the model: CLAUDE.md and rules, skills, agents, commands, hooks,
scripts, MCP servers. The goal is a harness that **measurably** does the work better than Claude
with no harness, at the lowest token cost. Every claim of "better" comes from an eval run.
## Session start (every session)
1. Run `node "${CLAUDE_PLUGIN_ROOT}/scripts/hb.cjs" status` in the project folder. It prints the stage and the next step.
2. Read `.harness/STATE.md` if it exists. Do not re-derive facts it records.
3. Stage `none`: ask the user (AskUserQuestion) whether this is a **new** harness or an **improve**
job, and for the target path. Then run `hb init --mode new|improve --target <path>`.
## Stages
| Mode | Stage | Skill to load | Output | Gate (stop and ask) |
|---|---|---|---|---|
| new | research | harness-research | `.harness/research/findings.md` | approve the question list and sources first |
| new | design | harness-design | `.harness/design.md` | approve the design before building |
| new | build | harness-design | the harness files, with a `<harness>-orchestrator` agent for `claude --agent` | — |
| improve | audit | harness-audit | `.harness/audit.md` | approve which findings to act on |
| both | eval-suite | harness-eval | eval cases | approve the case list |
| both | wrap (non-plugin targets) | harness-eval | `.harness/wrapped/` | — |
| both | baseline | harness-eval | first ledger entry | approve the cost estimate |
| both | optimize | harness-optimize | ledger entries, changed files | approve each eval run and each edit to an existing harness |
Load the stage's skill with the Skill tool before doing the stage. Work one stage at a time; when a
stage's output exists, `hb status` moves on by itself.
## hb commands
`node "${CLAUDE_PLUGIN_ROOT}/scripts/hb.cjs" <command>` (Node built-ins only):
| Command | Use |
|---|---|
| `init --mode new\|improve --target <path>` | create `.harness/` (never overwrites) |
| `status` | stage + next step |
| `inventory <path> [--out file] [--json]` | static audit: counts, always-loaded tokens, budget findings |
| `estimate <eval dir> [--runs n] [--ablation none]` | agent runs, judge calls, cost range, suggested ceiling |
| `record <results dir> --label "<one change>"` | add an eval run to `.harness/experiments.jsonl` |
| `compare [a] [b]` | per-case change between two runs: improved / regressed / noise |
| `wrap <.claude dir or project> .harness/wrapped --evals .harness/evals` | eval-ready plugin copy of a non-plugin harness |
| `variant <plugin> --from evals --to evals-ctx --append <context.md>` | suite that carries CLAUDE.md text as `append_system_prompt` |
## Project folder
```
.harness/
harness.json mode, target, models
STATE.md verified facts, decisions, open failures, last session
research/ questions.md, notes/<source>.md, findings.md
design.md one row per control
audit.md hb inventory output + your reading of it
evals/ eval suite for non-plugin targets (plugin targets keep evals/ in the plugin)
wrapped/ hb wrap output (generated; safe to replace)
experiments.jsonl one line per eval run
```
## Rules that hold in every stage
- **Measure before claiming.** "Better" means a positive Δ against the no-harness arm in a recorded
eval run. Say "not measured" when there is none.
- **Spend only with approval.** Before any `claude plugin eval`, run `hb estimate` and show the range.
Always pass `--max-cost-usd`. The plugin's hook stops each eval run for the user's approval and
denies a run without a ceiling.
- **Never edit an existing harness without approval.** Work on a git branch when the target is a
repo. Show the diff of each change before keeping it.
- **One agent with skills first.** Do the stages yourself. Spawn a subagent only for 3+ independent
research sources or a fresh-context review; brief it with objective, output format, sources,
scope and a tool budget; it returns ≤400 words plus a file path.
- **Never guess platform facts.** Frontmatter keys, flags and limits come from the docs or a live
check. Record verified facts in STATE.md with how they were verified.
- **End of session:** update STATE.md "Last session" (tried / passed / failed / next).
## Platform limits to plan around
- `claude plugin eval` loads only plugins. Non-plugin harnesses go through `hb wrap`.
- Eval runs on native Windows cannot grant Bash or PowerShell (no OS sandbox backend; WSL2 works).
Cases there use Read, Write, Edit, Glob, Grep and Skill, granted with `--allow-tools Write Edit`.
- Eval runs load nothing from the user's own config: no CLAUDE.md, user skills or other plugins.
- Traces are deleted after a run unless `--keep-temp` is passed. Pass it when you need to read transcripts.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · mhmdsalameh/ai-native-harness | 100 | MIT |
| ericrisco/rsc-harness | 100 | MIT |
| 90p-labs/harness | 100 | MIT |
| AbdullahYA1/claude-codex-harness | 100 | MIT |
| kcenon/claude-config | 100 | BSD-3-Clause |
| even-realities/everything-evenhub | 100 | MIT |
| SensorsIot/Universal-ESP32-Workbench | 100 | MIT |
| HikaruEgashira/agent-skills | 100 | Apache-2.0 |
| stellarlinkco/myclaude | 100 | AGPL-3.0 |
Migawka bazy z 13.11.2026 · wersja paczki 2681130db71b (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100harness-design
Decyduje, gdzie umieścić każdy element harnessu Claude Code: CLAUDE.md, reguły, skille, hooki, agentów lub skrypty, z budżetem tokenów. - 100harness-audit
Audytuje istniejący harness Claude Code: koszt stałych tokenów, źle umieszczone reguły, sprzeczne instrukcje i słabe opisy skilli. - 100setup
Konfiguruje harness frontendu w React lub Next.js: hooki Claude Code, reguły, skille, sprawdzenia i hooki gita. - 95harness-maintenance
Sprawdza i odświeża stan możliwości harnessów AI, takich jak Codex, Claude Code czy Copilot CLI: skille, hooki i agenci. - 100harness-engineering
Pomaga poprawiać i budować harness AI repozytorium: AGENTS.md, reguły, skille, komendy, hooki, zabezpieczenia i czujniki CI, a także wdrażać wnioski z audytu harness-score i podnosić poziom dojrzałości. - 100harness-engineering
Pomaga poprawiać i budować harness AI repozytorium: AGENTS.md, reguły, skille, komendy, hooki, zabezpieczenia i czujniki CI, a także wdrażać wnioski z audytu harness-score i podnosić poziom dojrzałości.