メニュー
Blueprint for a Claude 5 Fable coding agent that reads a ticket, plans, edits, runs tests, opens a PR, and self-reviews — with guardrails against destructive ops.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/kalilurrahman/kr-claudiator-skills --skill autonomous-coding-agentコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
SOC 職業分類に基づく
| name | autonomous-coding-agent |
| description | Blueprint for a Claude 5 Fable coding agent that reads a ticket, plans, edits, runs tests, opens a PR, and self-reviews — with guardrails against destructive ops. |
| argument-hint | ["repo","ticket source","CI system","merge policy"] |
| allowed-tools | Read, Write, Bash |
| model | claude-fable-5 |
| capability | Agentic Workflows |
Claude 5 Mythos + Fable skill — Agentic Workflows. Optimised for Claude 5 Mythos (reasoning) and Claude 5 Fable (action / tool use), and backward-compatible with Claude Sonnet 4.5 and Opus 4.
Ship a coding agent that behaves like a careful junior: reads the ticket, writes a plan comment, edits in a branch, runs tests, iterates on failures, opens a PR with a self-review checklist, and stops on destructive commands.
| Field | Required | Notes |
|---|---|---|
context_source | ✅ | Repo path, doc set, ticket, screenshot, or MCP resource id |
goal | ✅ | The single sentence definition of done |
audience | ✅ | Reviewer, end-user, downstream agent, or auditor |
constraints | ✅ | Token budget, step budget, permitted tools, forbidden ops |
output_shape | ✅ | Free text / markdown / JSON schema / patch / PR |
citations | ⛔ optional | Required for research, legal, medical, financial outputs |
escalation | ⛔ optional | When to stop and ask a human |
<file>, <doc>, <ticket>) to make regions addressable.rm, DROP, git push --force) through an approval gate.The skill produces:
Example template head:
<system>
You are Claude 5 Agentic operating under the "autonomous-coding-agent" skill.
Follow the output contract exactly. If a required field is unknown, return
null with a `reason` explaining what is missing. Never fabricate citations.
</system>
<context>
<!-- structured, ordered, addressable -->
</context>
<task>
<goal>{goal}</goal>
<constraints>{constraints}</constraints>
<output_shape>{output_shape}</output_shape>
</task>
Goal. Ship a coding agent that behaves like a careful junior: reads the ticket, writes a plan comment, edits in a branch, runs tests, iterates on failures, opens a PR with a self-review checklist, and stops on destructive commands.
Inputs.
context_source: (task-specific — see argument-hint)goal: single-sentence definition of doneoutput_shape: JSON matching the schema in /schemas/autonomous-coding-agent.jsonRun.
/schemas/autonomous-coding-agent.json.Expected artefact. A machine-readable output plus a human-readable summary that a reviewer can approve in under 5 minutes.
| Cadence | Action |
|---|---|
| Per run | Emit trace: prompt hash, model id, tokens in/out, tool calls, validator result |
| Per week | Review failure traces; tighten the schema, examples, or guardrails |
| Per month | Rotate examples; retire ones that no longer surface real failure modes |
| Per quarter | Re-benchmark against the current Claude 5 point release |
prompt-engineering-playbook — reusable prompt patterns across every model.mcp-agent-architecture — deeper wiring for MCP-based tool use.rag-knowledge-systems — when the corpus outgrows even a 1M window.executive-briefs — for turning Claude 5 output into partner-ready one-pagers.Claude 5 Mythos and Fable are powerful because they behave like a careful colleague, not a magic oracle. This skill is the frame that keeps the colleague honest, cheap, and reviewable.