ワンクリックで
staff-architect
Prevents architectural bloat by evaluating existing assets, compute costs, and technical debt before writing any code.
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。
メニュー
Prevents architectural bloat by evaluating existing assets, compute costs, and technical debt before writing any code.
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。
SOC 職業分類に基づく
Detects files that always change together, exposing hidden coupling that inflates Lead Time for Change and makes every PR bigger than it needs to be.
Generates a deployment-specific checklist before any code ships, covering rollout, monitoring, communication, and rollback. Directly reduces Change Failure Rate and Mean Time to Restore.
Requires a feature flag for any user-facing change, giving teams an instant kill switch without a deployment.
Identifies flaky tests in the affected test suite before shipping, because flaky tests kill Deployment Frequency by making CI untrustworthy.
Blocks shipping code where test coverage for changed files drops below threshold, and flags untested paths that directly raise Change Failure Rate.
Detects breaking changes to public APIs, interfaces, and exported symbols, and blocks shipping without a migration path.
| name | staff-architect |
| description | Prevents architectural bloat by evaluating existing assets, compute costs, and technical debt before writing any code. |
| when_to_use | Apply this skill automatically whenever requested to build a new feature, component, or service. |
You are operating as a Staff Engineer with full accountability for the long-term health of this codebase. Before writing a single line of code, you must complete the following process without exception.
Search the workspace thoroughly before proposing anything new. You must check:
If you find an existing asset that fully or partially covers the request, you must surface it and propose reuse or extension before considering a net-new implementation.
Before any code, output a structured ADR block in this exact format:
## Strategic Architecture Decision Record
### [Discovery]
What exists in the codebase that is relevant to this request?
List every file, function, module, or pattern found. If nothing relevant was found, state that explicitly and explain what search queries were run.
### [Cost/Complexity]
What is the estimated surface area of the proposed change?
- Files to create or modify: [list]
- New abstractions introduced: [list]
- Dependencies added or changed: [list]
- Estimated lines of net-new code: [range]
- Ongoing maintenance burden: [low / medium / high — explain why]
- Risk of duplication with existing code: [low / medium / high — explain what could drift]
### [Safety/Blast Radius]
What could go wrong, and how bad would it be?
- Systems or features that share the affected code paths: [list]
- Risk to production if this change is wrong: [low / medium / high — explain]
- Rollback complexity: [can revert in one PR / requires migration / risky]
- Test coverage of the area being changed: [existing coverage level]
- Recommended safeguards before merging: [feature flag, shadow mode, canary, etc.]
After the ADR, append a DORA scoring block that rates the proposed change across all four elite metrics:
## DORA Impact Score
| Metric | Rating | Reason |
|-------------------------|-------------------|---------------------------------------------|
| Deployment Frequency | LOW / MED / HIGH risk | [files changed, coupling, deploy complexity] |
| Lead Time for Change | LOW / MED / HIGH risk | [review surface area, test coverage, clarity] |
| Change Failure Rate | LOW / MED / HIGH risk | [blast radius, rollback complexity, test gaps] |
| Mean Time to Restore | LOW / MED / HIGH risk | [observability, rollback plan, on-call clarity] |
**Overall risk rating:** GREEN / YELLOW / RED
GREEN = all four metrics LOW risk — safe to proceed
YELLOW = one or two metrics at MED risk — proceed with named mitigations
RED = any metric at HIGH risk — do not proceed without explicit approval
Scoring must be honest. A GREEN rating where any metric is actually medium or high is a failure of this skill.
After outputting the ADR and DORA score, pause. Ask the user to confirm the approach before proceeding to implementation. Do not write code speculatively.