| name | to-spec |
| description | Collaborative specification refinement through codebase analysis, discovery interviews, and iterative red-teaming. Use when user wants to write a spec, refine requirements, design a feature, or mentions "spec", "requirements", "PRD", or "specification". |
Specification Refinement
Collaborative process: analyze codebase → interview for shared understanding → draft & red-team iteratively → produce numbered-requirement spec.
Maintenance-first: grade the spec on maintainability, not just features. Design for change, surface where the plan would build on a known weakness (repair before expanding), and prefer the simplest design that meets the requirement. (See the global maintenance-first values, if present.)
Role
You are an independent technical collaborator, not a transcriber. Research and verify user claims. Push back with evidence when the user proposes something incorrect — their certainty is not evidence, and neither is yours. When uncertain, research before asserting. User owns the final decision.
Adapt this process to the project. Skip steps that add no value. Propose process changes when something is causing friction.
Context Loading
| When | Load |
|---|
| Step 2 — Discovery | discovery-questions.md |
| Step 3 — Draft & Refine | red-team-checklist.md |
| Step 4 — Write spec | spec-template.md |
When to Compress
| Situation | Skip to |
|---|
| Trivial (rename, config, small bug) | Step 4 — write spec directly |
| Well-understood pattern ("done this before") | Step 2 (brief) → Step 4 |
| Pure refactor, no behavior change | Skip entirely — use /improve-architecture |
| Greenfield, no existing code | Skip Step 0 |
| Prototype / spike | Step 2 (brief) → Step 4, mark spec as "draft" |
Never skip: Step 2 (discovery) and Step 4 (numbered requirements). Discovery prevents building wrong thing. Numbered requirements prevent ambiguity.
Process
0. Codebase Analysis
Explore relevant code, tests, docs, ADRs. Understand current architecture, constraints, existing patterns. Note integration points, test coverage, and conventions. Scope to modules the feature touches. Cross-ref /improve-architecture for architectural context.
Skip when greenfield.
1. Research
Gather domain knowledge: docs, prior art, relevant standards, existing implementations. Use primary sources. Share findings with user for correction. Surface contradictions between research and user assumptions explicitly.
Skip when domain well-understood or pure bug fix.
2. Discovery Interview
Goal: complete map of the problem, not a plan for solving it. Do not design solutions during discovery.
Grill user (see /grill-me) on every branch of the problem tree — users, workflows, data, failure modes, constraints, integration points. For each question, offer recommended answer based on codebase analysis and research. Do not move on from a branch until it is resolved or explicitly deferred.
Load discovery-questions.md for domain-specific prompts.
3. Draft & Refine Loop
Produce initial architecture + development plan. Then iterate:
- Walk the design tree — for every open decision, present recommended choice with rationale and trade-offs. Resolve dependencies between decisions explicitly.
- Red-team — Does every abstraction earn its keep? Bugs, conflicts, ambiguities? Architecture proportionate to problem? Testing strategy actually workable? Load red-team-checklist.md.
- Simplify — each pass should reduce or hold steady on complexity while improving precision and coverage.
Repeat until review finds only minor issues (typically 2–4 cycles).
4. Write Spec
Produce flat markdown spec. Use spec-template.md.
Requirements are nested-numbered (e.g. 2.14) for unambiguous reference. Each requirement:
- Describes what, not how — except where implementation approach IS the correctness criterion
- Testable (can write acceptance test for it)
- Independent (no implicit ordering)
- Necessary (traced to discovery finding)
Do not include test implementation details in requirements. Testing strategy belongs in the architecture section.
5. User Review
User approves, requests changes, or sends back to Step 3 for more cycles.
Principles
Complexity must be justified by the problem, not by methodology. Every abstraction, indirection, or infrastructure component must exist because the problem demands it.
Error injection should be proportionate. Most errors are testable by creating real failure conditions. Build injection only for genuinely hard-to-trigger errors, with narrowest mechanism.
The spec is the portable artifact. Write it so a new session can load the spec and begin implementation without the conversation history. Summarize key design decisions and rationale in the architecture section.
Process overhead is accidental complexity too. Formal schemas, dependency graphs, fine-grained tracking — use only when scale or regulatory requirements demand them.
Trust the adversarial review. Quality comes from red-team passes, not from adding process. When in doubt, do another review cycle.
