claude-harness
August 25, 2026 · View on GitHub
Language: English | 日本語
claude-harness
Public snapshot of the Claude Code harness (skills / agents / rules / hooks) that shimo4228 uses day-to-day.
Skills, agents, and rules are a mechanical aggregation of assets tagged origin: shimo4228 from ~/.claude/; ECC-derived material (origin: ECC / ECC-customized) and auto-extracted artifacts (origin: auto-extracted) are excluded. ADRs are synced wholesale, and hooks come from a curated allowlist — publication there is a judgement about reuse outside this machine, not about who wrote the file.
Positioning
- Audience: Claude Code (CLI + IDE extensions) users, and developers researching agent skill / rule ecosystems
- Source of truth:
~/.claude/is canonical; this repo is a one-way export produced byscripts/sync-from-local.sh(origin filter + hook allowlist → secret scan → subtree replacement) - License: MIT. Free to copy, modify, and redistribute. Forking and customizing for personal use is encouraged
Contents
Skills
| Skill | Purpose |
|---|---|
| search-first | Research-before-coding workflow. Invokes the scout agent to discover existing tools |
| learn-eval | Extracts reusable patterns from sessions, evaluates quality, and decides where to save |
| skill-stocktake | Skill quality audit — inline Glob inventory + single-context holistic evaluation, Keep/Improve/Update/Retire/Merge verdicts |
| skill-health | Structural skill-library debt scan — flags "missing artifacts" (SKILL.md references to scripts / agents / sibling skills that don't resolve on disk). Deterministic; delegates quality / risk / validation to skill-stocktake / security-scan / skill-comply |
| rules-distill | Extracts cross-cutting principles from skills and promotes them to rules |
| rules-stocktake | Rules quality audit — residency-cost model (every line is a per-session token tax), staleness / substrate-absorption checks, Keep/Improve/Update/Merge/Demote/Dissolve/Retire verdicts. The inverse of rules-distill |
| skill-comply | Measures actual compliance of skills / rules / agents. Classifies behavioral sequences across 3 prompt strictness levels |
| context-sync | Audits and fixes project documentation. Detects role overlap, checks freshness, creates missing docs |
| codex-review | Cross-model second opinion from the OpenAI Codex CLI (a different model family), read-only on both the argv and config face — (1) code review of the current diff folded into the review chain, (2) plan-stage premise challenge of a design packet (refute / missing / alternative, never a design) |
| llms-txt-writer | Writes AI-facing docs (llms.txt / llms-full.txt). Answer.AI standard + GEO/AEO static analysis |
| jsonld-knowledge-graph | Designs and ships a companion JSON-LD knowledge graph (graph.jsonld) next to llms.txt. Encodes domain entities and relationships as schema.org triples for LLM citation |
| writing-ecosystem | Orchestrator for human-facing writing & review. Coordinates editor / essay-reviewer / fact-checker |
| collect-context | Gathers in-session and external context into source material for article writing |
| authorship-strategy | 4-layer framework (Authenticity / Attribution diffusion / Idea-vs-scaffold / Tactics) for DOI-registered idea-rescue research repos |
| release-doi | Cuts a versioned release of a DOI-registered research repo (Zenodo concept DOI semantics, CHANGELOG / tag / asset packaging) |
| adr-writer | Records design decisions as numbered ADRs — directory detection, sequence numbering, index update; prose delegated to the adr-writer agent |
| paper-ecosystem | Orchestrator for academic paper writing & review — role boundaries for paper-writing plus five reviewer agents; holds Source Fidelity / Vocabulary / Voice / Clarity / Citation rules |
| paper-writing | Drafting procedure for academic papers — title, outline, section drafting, abstract, references with claim-cite 1:1 mapping |
| paper-deposit | Deposits a finished, reviewed paper to Zenodo as a standalone DOI record, optionally cross-posts to SSRN, cross-links the DOI back into the research repo |
| ai-native-preprint-submission | Submits a deposited paper to AI-native preprint platforms (aiXiv / AiraXiv) — Web UI browser automation with human gates, or author-delegated API/MCP submission |
| readme-writer | Writes human-facing READMEs — deterministic structural lint plus holistic LLM review (no scores) |
| hf-sync | Mirrors graph.jsonld-bearing research repos to Hugging Face Datasets |
| citation-sync | Audits the three citation layers of a research repo (docs / .zenodo.json / graph.jsonld) and syncs them bottom-up |
| spawn-session | Launches a new detached Claude Code Remote Control session in a Herdr pane, visible in the mobile app session list |
| harness-sync | One-way export of origin-filtered components from the live harness into this repo — collection, secret scan, subtree replacement |
| cited-source-mirror-verification | Verify an access-blocked or digest-sourced numeric claim against an open mirror before citing it in a durable artifact |
| wiki-harvest | Read-only harvest from an Obsidian LLM wiki (wiki/concept/) into a research repo — extracts only next-action-changing candidates into a ranked, source-cited ledger under the repo's .notes/ |
| wiki-query | Read-only query over an Obsidian LLM wiki (wiki/concept/) with [[ ]] source-cited synthesis |
| repo-asset-stocktake | Audits a project repo's non-code assets (tool configs, CI workflows, runbooks) for diminished value — flags assets whose consumer has vanished, with Keep/Update/Retire/Merge verdicts |
| task-stocktake | Audits and consolidates a repo's pending-task tracking into its single task ledger — bootstraps the ledger, sweeps stray task lines, verifies entries against git log and actual code |
| llm-as-judge | Design pattern for LLM-as-judge evaluators — binary checks as evidence, one named holistic verdict, no score aggregation |
| implementation-chain | Decides the task type (feat / fix / refactor / chore / prototype / writing) and front-loads its agent chain into the plan — Chain Matrix, reviewer routing, early-stop conditions |
| public-comment | Replies in public technical threads (GitHub discussions / issues / PRs, HF discussions) — AI-slop tell removal, thread grounding, and a human gate with a Japanese translation before posting |
| agent-stocktake | Audit subagent definitions with a hybrid cost model (description = per-session residency, body = invocation) — flags suppression instructions and substrate absorption; third sibling of skill-/rules-stocktake |
| generation-audit | On a model-generation change, capture the live runtime layer (system prompt + tool descriptions), classify mismatches as conflict / redundancy / drift, and hand the evidence to the stocktake skills for verdicts |
| git-workflow | Permission-friction discipline for git in this environment — one Bash call per git command; chaining with && or pipes breaks the Bash(git:*) auto-allow and stalls on manual prompts |
| headline-craft | Craft skill for the one line that makes readers open — title / tagline / subtitle / SNS-post candidates, generated with concrete techniques and scored per traffic channel (search vs feed) |
| herdr-delegate | Hand a whole implementation task to a different CLI agent running in a Herdr pane (Codex, etc.). Gated on an explicit user request — parallelism alone is not a reason |
| prompt-perturb | Diversity injection. A deliberately context-starved forager agent fetches prompts from external creativity-technique catalogs, so the angles come from outside the session's own habits |
| session-judgment-mining | Mine past session transcripts for judgements the user made repeatedly, and promote the recurring ones into skills or rules |
| verify-bootstrap | Stand up a repo's machine gates (format / lint / type check / security / dependency / test), or take stock of gates that have gone stale. Tool choice is researched at bootstrap time rather than baked into the skill |
| x-draft | Turn a research report into one long-form social post. Pull-only — no quota, no notification, invoked only when the author already wants to post. Rechecks the primary source, gates on staleness, strips the AI tells, and stops at the draft |
| task-triage | One cycle of the task-triage loop: judge every open ledger task (premise, start condition, worth), dispatch the ready ones to fresh build sessions, verify their output independently — the human keeps the merge word |
| harness-boundary | Design-time lens for any proposed mechanism (rule / skill / hook / agent / workflow): which of 6 layers it belongs to, whether the model could own it instead, and whether it survives a runtime swap — keep only what outlives the harness |
| skill-creator | Write or rewrite a skill / agent definition — intent packet, library-wide boundary check, Fable-era writing rules, a fresh-context draft gate (Publishable / Fix / Drop, no scoring), author read-through. Replaces the upstream anthropics skill-creator in place (ADR-0046) |
| measurement-discipline | 測定に基づく主張・閾値・ガード・実験結果を設計または評価するときの規律。Use when the user says 「この実験結果で判断していい？」「閾値を決めたい」「ガード/検査を足したい」「1 回通ったから大丈夫」, when a design places a numer |
| prose-translation | 日本語⇄英語の voice 保持翻訳スキル（両方向）。エッセイ・記事・README・ADR 等の人間向け prose を、出力先の publication channel contract が宣言する register と原文の確度を保って自然に訳す。逐語訳でも MT で |
| quality-gate | 人間向け公開物の受け入れゲート。完成稿と project の publication channel contract を読み、必須 reviewer verdict・機械検査・最新 title-reviewer findings が揃ったかを集約して PASS / FAIL / |
| repair-discipline | バグ修正・残課題・schema/storage 変更に着手するときの規律。Use when the user says 「このバグ直して」「残課題をやって」「この schema を変えたい」, when picking up a stale task file, or when |
| session-theme-mining | 過去の Claude Code / Codex セッションを横断し、記事になりうる未解決の問いを 0〜3 件の同格な候補カードとして発見する。Use when — 「過去セッションから記事テーマを探して」「まだ書いていない問いを発掘して」「セッション履歴から collect-co |
| rfc-writer | 公開 rfcs/ 台帳へ 1 エントリを起票する手順と規約の唯一の正本（足切り → 採番 → 様式 → 公開規約 → spawn 接続 → index 行）。Use when the user says 「これ起票して」「RFC にしておいて」「提案を台帳に載せて」, when |
The first six (search-first, learn-eval, skill-stocktake, rules-distill, skill-comply, context-sync) are components of the Agent Knowledge Cycle (AKC). Each is also published as its own standalone repo, but they are bundled here so the harness can be read end-to-end.
Agents
| Agent | Purpose |
|---|---|
| scout | Pre-implementation solution discovery. Searches npm / PyPI / MCP registries / GitHub for existing solutions |
| prompt-writer | Generates concise prompts using a lightweight model. Creates and rewrites LLM prompt templates |
| editor | Strict technical article editor. Rigorously reviews code accuracy, AI slop, narrative flow, and terminology consistency |
| essay-reviewer | Strict essay editor. Targets idea pieces mixing social theory / organizational analysis / design philosophy / personal narrative |
| fact-checker | Fact verification specialist. Extracts verifiable claims from articles and verifies them via web sources |
| adr-writer | Generates the 7-section ADR body (incl. Review-when expiry conditions) from supplied input only — never invents context, expiry conditions, or alternatives |
| codemap-writer | Generates / refreshes docs/CODEMAPS/ — token-lean architecture documentation, ~1000 tokens per map |
| paper-reviewer | Academic paper structure review — argument flow, section transitions, claim sharpness, evidence-claim alignment |
| source-fidelity-checker | Reads each cited primary source directly and flags drift between paper claims and source content |
| vocabulary-consistency-checker | Verifies term definitions stay consistent and sub-classifications are explicit at introduction |
| clarity-reviewer | First-contact reader clarity review — coined-term budget, title-axis alignment, meta-commentary, insider-context dependency |
| citation-formatter | Verifies in-text citations against the reference list — format consistency, DOI / arXiv ID validity |
| readme-reviewer | Strict README / repo top-page review — LLM-read floor, lead clarity, human hook, scannability, length discipline, visual effectiveness. Companion to readme-writer |
| readme-clarity-reviewer | First-contact reader clarity review for READMEs — coined-term budget, insider-context dependency, Japanese register (ですます). Parallel partner of readme-reviewer |
| adr-reviewer | Checks an ADR's record, not its decision — whether Context carries verifiable evidence, Review-when names an observable expiry trigger, Alternatives are real rather than straw men (a live 「未決」 rival is allowed), Consequences show both sides, and override relations with prior ADRs are stated (dated 注記 on partial weakening) |
| prompt-forager | The context-starved half of prompt-perturb. Receives one line of purpose and deliberately nothing else, so what it finds is not shaped by the session that asked |
| swift-reviewer | Swift / SwiftUI review — Swift 6 strict concurrency, value semantics, SwiftUI state ownership, retain cycles, HIG compliance |
| readme-judge | Fresh-context README judge: reads evidence JSON + the README once, answers a fixed checklist with quoted evidence, returns a named verdict (Publishable / Fix / Rewrite) |
| prose-clarity-reviewer | First-contact reader clarity reviewer for human-primary articles, essays, blog posts, and newsletters |
| theme-reviewer | 人間向け記事・エッセイの執筆前テーマレビュアー。選択済みの問い一文と素材を fresh context で読み、非自明性・一次アクセス・読者接続・外部言説との差分を点検して findings と深化の問いだけを返す。Use before editorial brief |
| title-reviewer | 凍結した人間向け原稿のタイトルレビュアー。headline-craft の候補と現行タイトルを fresh context で読み、中心命題との軸一致・誠実さ・具体性・好奇心の回収・channel 制約を点検して findings だけを返す。Use after 本文の構造凍結、 |
Rules
Behavioral principles auto-loaded every session (under rules/common/):
| Rule | Purpose |
|---|---|
| agents | Agent orchestration conventions. When to use which agent, parallel execution patterns |
| akc-cycle | Six-phase behavioral conventions of the Agent Knowledge Cycle (Research / Extract / Curate / Promote / Measure / Maintain) |
| debugging | Root-cause-first debugging flow (hypothesis → evidence → confirm → fix), AI recency-bias guards, retry-with-context |
| planning | Required items for planning (What / Why / Alternatives). Mandates Phase 0 external research |
| skills | Skill origin tracking spec and knowledge placement principles |
| contemplative-axioms | Contemplative Constitutional AI clauses from Laukkonen et al. (2025), verbatim |
| task-tracking | Single task ledger per repo — one canonical pending-task file, Done-section history, pointer-only discipline for MEMORY.md and detail documents |
| knowledge-staleness | Treats external LLM-domain knowledge as going stale on a one-week scale — never assert tooling, specs, or going rates from memory; check at search time, date the evidence, and attach an expiry condition to any recommendation |
| practitioner-identity | Author's self-definition, verbatim — searching for what counts as a good idea and a good means in the AI era; DOI is one means, not a researcher career; code fades, ideas persist |
Hooks
hooks/ carries five PreToolUse hooks that run at the git commit boundary — a secret scan, a runner for the repo's own machine gate, a bandit scan, a ruff format --check, and a review reminder — plus the two parts they need, and two session-surface hooks published with the rfcs/ ledger: a ledger-etiquette reminder (with scripts/claims.py, the ledger CLI) and a judge-tier review-routing guard. Several ADRs argue about their internals, so the code lives here rather than leaving those decisions pointing at nothing. Unlike skills and rules, hooks need manual wiring into settings.json. All carry bats tests, each checked with a negative control — the hook mutated to remove the property, the test confirmed to fail against the mutant. Install steps, the approval model behind the verify gate, and what is deliberately left out: docs/hooks.md.
Design decisions (ADRs) and proposals (RFCs)
docs/adr/ records why this harness is shaped the way it is: adoptions, retirements, and reversals, each as a dated Architecture Decision Record synced from the live harness alongside the components. The skills, agents, and rules above are the what; the ADRs are the why — the audit trail behind the harness, failures included. Start from the ADR index. ADRs are written in Japanese.
rfcs/ is the harness's public task-and-proposal ledger (ADR-0049): one entry per proposal or work item, body in Rust-RFC-template form, state in the frontmatter, terminal entries left in place so rejected proposals stay readable with their reasons. ADRs record decisions; rfcs/ holds what is still open — including the entries that will never be built, which is the point.
Usage
Full install
git clone https://github.com/shimo4228/claude-harness.git ~/.claude-harness
# Copy skills / agents / rules into ~/.claude/
cp -r ~/.claude-harness/skills/* ~/.claude/skills/
cp -r ~/.claude-harness/agents/* ~/.claude/agents/
cp -r ~/.claude-harness/rules/common/* ~/.claude/rules/common/
Hooks are separate: they must live under ~/.claude and be wired into settings.json by hand. See docs/hooks.md.
Cherry-pick
Copy only what you want:
cp -r ~/.claude-harness/skills/search-first ~/.claude/skills/
Setup for skills with Python implementations
llms-txt-writer, skill-comply, rules-distill, skill-stocktake, and skill-health ship with Python code. In each skill directory:
cd ~/.claude/skills/<skill-name>
uv sync # or: pip install -e .
Origin tags
Each file's frontmatter (YAML or HTML comment) carries an origin field:
| origin | Meaning |
|---|---|
shimo4228 | Authored by shimo4228. The scope of this repo |
ECC | From Everything Claude Code. Content not included — named below |
ECC-customized | ECC derivative + shimo4228 modifications. Content not included — named below |
auto-extracted | Learned skill auto-extracted by learn-eval. Not included |
This repo is the result of a mechanical collection limited to origin: shimo4228.
Upstream components (names only)
The live harness also runs components from external upstreams. Their content — including any local modifications to it — is not redistributed here; the names alone are listed so the full composition stays visible. ECC = Everything Claude Code.
| Upstream | Skills | Agents | Rules |
|---|---|---|---|
| ECC + local modifications | agent-harness-construction, ai-regression-testing, config-gc, e2e, loop-design-check, python-patterns, refactor-clean, tdd, update-codemaps | architect, e2e-runner, refactor-cleaner, security-reviewer | common/coding-style, common/security, common/testing |
| herdrdev/herdr | herdr | — | — |
| mattpocock/skills + local modifications | grill-me, wait-what | — | — |
| modem-dev/hunk | hunk-review | — | — |
Related repos
- shimo4228 — Hub repo aggregating the five practice lines (AKC / Contemplative Agent / AAP / Authorship Strategy / Attention Not Self) and the supporting ecosystem. This repo's clone/view traffic is published on its public dashboard
- agent-knowledge-cycle — AKC concept and DOI release (Zenodo: 10.5281/zenodo.19200726)
- contemplative-agent-rules — Rule implementation of Contemplative Constitutional AI
claude-skill-*standalone repos — Individual versions of each AKC skill (search-first / learn-eval / skill-stocktake / rules-distill / skill-comply / context-sync) plus the adjacent skills (llms-txt-writer / daily-research / jsonld-knowledge-graph / writing-ecosystem / rules-stocktake)
Contributing
This repo is shimo4228's personal harness artifact, so external PRs are not accepted. Instead:
- Fork it and customize freely
- Issues for questions or suggestions are welcome
Bug fixes flow upstream into ~/.claude/ when shimo4228 incorporates them.
License
MIT License. See LICENSE.