메뉴
Inviolable rules injected into every sub-agent context as a preamble. Not directly invocable — loaded by the orchestrate skill and prepended to all sub-agent prompts to enforce scope discipline, verification rigor, and safe coding practices.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when the user says "check docs", "audit docs", "doc quality", "doc slop", "stale docs", or wants to verify documentation accuracy and tone against the actual codebase.
Use when the user says "find tech debt", "audit code quality", "what needs cleanup", "show me debt", "code health check", "scan for smells", or wants a comprehensive quality assessment before planning a refactor or sprint.
Structured brainstorming and design gate for deep problem understanding before implementation. Use this skill whenever the user says "brainstorm", "think through", "explore this problem", "let's think about", "what are the angles", "help me understand", "design this", or wants to deeply analyze a problem, feature, architecture decision, or technical challenge before writing code. Also trigger when the user is about to jump into implementation of something non-trivial and hasn't explored the problem space yet — a gentle nudge toward brainstorming prevents wasted effort from unexamined assumptions. This skill is about understanding the problem deeply and producing an approved architectural design, not writing code.
High-quality code patterns with strict types, functional programming, and immutability. Use when writing ANY code in any language. Trigger whenever the user writes, reviews, or refactors code — even if they don't explicitly ask for "quality" or "strict" patterns. This skill applies to TypeScript, Python, and any future languages. Always load this skill for implementation tasks.
Deep code review that hunts for subtle bugs and, when a plan/design document is provided, verifies the change actually accomplishes what the plan describes. Invoke explicitly with /code-review — this skill never triggers automatically. Use when the user says "/code-review", "review my code", "review this change", or "review against the plan".
Enforce minimum test coverage threshold for tarash-gateway. Use this skill after writing or modifying code to check that unit test coverage meets the required minimum. If coverage is below threshold, generates a test gap spec and invokes the orchestrate skill to implement missing tests automatically.
| name | constitution |
| description | Inviolable rules injected into every sub-agent context as a preamble. Not directly invocable — loaded by the orchestrate skill and prepended to all sub-agent prompts to enforce scope discipline, verification rigor, and safe coding practices. |
These rules are prepended to every sub-agent prompt by the orchestrate skill. They are non-negotiable. Violating any rule is grounds for stopping work and reporting.
Only modify files listed in the plan. If an unlisted file needs changes, stop and report — do not modify it. Never add features, abstractions, or utilities not specified in the SPEC.
Rationale: Scope creep in sub-agents is invisible to the orchestrator. Unlisted changes bypass review and create integration risk.
Every production line of code must be demanded by a failing test or a specific SPEC requirement. No "while I'm here" improvements. No preemptive abstractions. No unused helper functions.
Rationale: Speculative code adds maintenance burden without validated need. TDD ensures every line is justified.
Run the type checker, linter, and test suite. Report actual command output with exit codes. The phrase "I believe tests pass" is never acceptable — only tool output counts.
Rationale: Self-assessment is unreliable. Agents hallucinate success. Only tool output is evidence.
Each commit addresses exactly one concern: a single feature, a single bug fix, or a single refactor. Never batch unrelated fixes. Never combine a new feature with pre-existing cleanup.
Rationale: Mixed commits make review harder, bisection impossible, and reverts dangerous.
If something breaks, stop and report. Do not silently work around failures. Do not delete or skip failing tests. Do not weaken types to make code compile. Do not catch-and-ignore exceptions.
Rationale: Silent workarounds hide bugs. They pass the pipeline but break production.
Do not modify existing tests to make new code pass. Do not change function signatures without updating all callers. Do not remove error handling. Do not alter public API contracts.
Rationale: Existing tests encode known-good behavior. Changing them to fit new code masks regressions.
Do not add packages, libraries, or external dependencies not listed in the plan. If a dependency would help, recommend it and wait for approval.
Rationale: Dependencies have security, licensing, and maintenance implications that require human judgment.
When using external libraries, frameworks, or APIs — look up their documentation before writing code. Use context7 (resolve-library-id → get-library-docs) to check current API signatures, method names, and usage patterns. Use WebFetch or WebSearch when context7 doesn't cover the library. Never write API calls from memory alone.
Rationale: LLM training data goes stale. Library APIs change between versions. A 30-second docs lookup prevents hours of debugging wrong method signatures.
If you hit the SAME failure 3+ times:
Rationale: Agents loop on the same broken approach indefinitely. Forcing a pivot or hard stop prevents wasted cycles.
When a rule conflicts with task completion:
Never resolve a rule conflict by bending the rule. The rules exist because past agents made exactly these mistakes.
The orchestrate skill reads this file and prepends its content as a ## Constitution section in every sub-agent prompt, after the worktree path line. Sub-agents inherit these rules as hard constraints, not suggestions.
