spec-fortify（规范强化）

Spec Fortify

Use this skill for the fortification phase of the spec-driven workflow.

Instructions

You are reviewing a project spec after its decisions have been resolved.

The user will reference a spec file. Read it in full.

If the work is not obviously greenfield, do a shallow codebase scan of the parts most likely to constrain the spec before proposing changes. Look for existing interfaces, migrations, deployment assumptions, auth or trust boundaries, and operator-visible behavior. This is a reality check, not a redesign pass.

Fortification Posture

Fortification is an integration pass, not an adversarial hunt for more work. It is valid to conclude that the spec needs no changes.

Prefer the smallest change that resolves a real issue:

1. clarify a constraint, explicit assumption, or accepted-risk boundary
2. tighten verification requirements
3. add a new pending decision only when implementation would otherwise proceed in materially different ways
4. reopen a resolved decision only as a last resort

Do not turn a non-blocking unknown into a pending decision just because it is still interesting. If different outcomes would not materially change what gets built now, clarify the boundary and move on.

Reopen a decision only when at least one of these is true:

• the current text permits two materially incompatible implementations
• two resolved decisions, or a resolved decision and a stated constraint, truly conflict
• the spec claims a safety, trust, ownership, or delivery guarantee that the current text cannot actually support
• leaving the decision resolved would likely cause incorrect implementation, unsafe defaults, or destructive operator behavior

Do not reopen a decision merely because:

• there is residual risk the spec already accepts explicitly
• the system has a deliberate version-one limitation or non-goal
• a concern can be handled by making an implicit assumption explicit
• you can imagine a cleaner future extension

This skill can also help with cross-tool review. If the user asks for a prompt for another model, treat that as a request for a second-opinion review, not as delegated fortification. Preserve the user's actual request and stated review lens. Do not smuggle this skill's decision-management rules or fortification workflow into the external prompt. Add only the minimum extra context, constraints, and likely hotspots needed to make the question answerable well. Unless the user explicitly asks for a rubric or strict structure, do not overprescribe the response format. Do not ask the other model to fortify the spec, manage decision state, or decide whether decisions should be added, reopened, or resolved. Prefer telling the user to attach or paste the spec for that model, include any additional context explicitly, and do not assume the other model can see the current conversation, files, or tool state. If the user brings back feedback from another model, evaluate it against the spec under this same fortification posture and integrate only what actually holds up.

Check For

1. Contradictions. Do any resolved decisions conflict with each other or violate a stated constraint? If a short clarification removes the contradiction, prefer that over reopening.
2. Gaps. Is there something the resolutions assume but never explicitly decided? If the missing piece is only an implicit assumption, accepted risk, or dependency, make it explicit instead of creating a new decision.
3. Accidental complexity. Read all resolutions as a whole. Is the total complexity proportional to the summary? Flag anything that crept in across sessions that nobody would have chosen deliberately. Do not reopen just because a simpler future exists; the complexity must be disproportionate to the stated goals as written.
4. Security risks. Do the current decisions introduce obvious security concerns, unsafe defaults, or unclear trust boundaries? Prefer adding an explicit constraint or assumption first. Reopen only when the current text would otherwise misstate the safety or trust posture.
5. Scaling risks. Scaling here means the system's ability to maintain acceptable performance and reliability as request volume, concurrency, or data size increases. Only flag concerns that are concrete and plausible from the spec as written. Explicitly accepted capacity limits are not gaps; only flag limits that contradict the claimed contract or are missing from it.
6. Verification steps. Do the current decisions make clear how the implementation will be verified? Prefer adding concrete verification requirements over reopening the underlying decision when that is enough.

Update the Spec

Don't add review notes or commentary to the spec file.

If the review surfaces a clarification-only change that preserves the meaning of the current decisions, apply it directly to the spec and report it briefly.

Require the user's approval before making any change that alters decision state or meaning.

It is valid to make no changes if the spec is already coherent enough to implement safely.

Clarification-only changes that may be applied directly include:

• adding or clarifying an explicit assumption
• adding or clarifying an accepted-risk boundary
• tightening verification requirements
• tightening wording to remove ambiguity without changing the resolved contract
• moving explicitly deferred nice-to-haves into Future Ideas

Changes that require the user's approval include:

• adding a new pending decision
• reopening a resolved decision
• changing the meaning of a settled contract or constraint