liza-mas/detailed-spec-writing

Objective

Transform fuzzy requirements into precise, SMARC specifications an Orchestrator can decompose into coding tasks.

Output: a git-tracked markdown document. One spec per task; one cohesive capability per spec.

Trigger

• Orchestrator creates a spec-writing task (Liza mode)
• User asks to spec a feature or requirement (Pairing mode)

Inputs

Your task provides:

• An output file path
• References to source material (vision doc, prior specs, existing code)

Scope discipline:

• Upfront: Read task references + scan existing specs in same domain for consistency.
• Reviewer-driven: Broader material only when Spec Reviewer feedback identifies gaps.
• Declare what you read and why in References.

Output Format

PRD format template at the task-specified path.

Protocol

1. Parse

Read source material. Distinguish what is said, implied, and missing before writing any requirements.

Multiple independent capabilities → flag to Orchestrator for split. Do not silently produce a mega-spec.

2. Write

SMARC every functional requirement — Specific, Measurable, Achievable, Relevant, Context-bound (implementable within a single agent's context budget — if it forces loading many files or deep cross-component knowledge, decompose further).

Cannot make it SMARC → Open Question, never a vague requirement.

References: Each feature traces to source material. No traceability = Scope Absorption.

Dependencies: True ordering constraints only. FT-X depends on FT-Y only if FT-X cannot be implemented or tested without FT-Y.

Assumptions are first-class outputs. Surfaced ambiguity is more valuable than papered-over gaps. Assumptions resolve behavioral ambiguity only — if an assumption names a file format, storage engine, library, or protocol, you've crossed into design territory. State the behavioral need and move on.

Acceptance Criteria are contracts. ACs define done. If they don't fully define done, the spec is incomplete.

Audience: Write for the implementing engineer. No marketing language, hedging, or filler.

3. Self-Review

Before submitting:

• [ ] Every FR/NFR traces to a source reference
• [ ] Added specificity beyond source is traceable to an Assumption, derivation, or Open Question
• [ ] Every AC maps to ≥1 FR; ACs test requirements, they don't create them
• [ ] No LOW-confidence assumptions in requirements (they belong in Assumptions section)
• [ ] Out of Scope covers adjacent concerns the Coder might drift into
• [ ] Spec is self-contained — Coder needs only this spec and referenced files
• [ ] Re-read source — nothing missed, nothing added beyond scope
• [ ] No assumptions/NFRs prescribe implementation choices — only observable behavior
• [ ] No OQ contradicts or duplicates an existing FR

Fix issues before submitting.

Constraints

• Do not write code — spec only
• Do not modify existing specs unless task explicitly scopes an update
• Do not invent requirements beyond source material and task scope
• Surface contradictions between references as Open Questions

Anti-Patterns

| Pattern | What goes wrong | |---------|----------------| | Wishful Specification | Untestable requirements. "User-friendly" is not a requirement. | | Hidden Coupling | References other components' behavior without declaring a dependency. | | Assumption Burial | Assumptions embedded in FRs instead of surfaced in Assumptions section. | | Scope Absorption | Source says "support X" → you spec X, Y, and Z. Stick to task scope. | | Premature Solutioning | Specifying how instead of what. Common leak: assumptions resolving technical ambiguity instead of stating behavioral need." | | Context Explosion | Requirement needs the entire codebase to implement. Decompose until bounded. |

Integration

| Skill | Relationship | |-------|-------------| | spec-review | Downstream. Validates specs against completeness/consistency/testability checklist. | | spec-backfill | Complementary. Backfill: specs from code (archaeology). This skill: specs from requirements (forward-looking). Same specs/ hierarchy. | | code-review | ACs are the contract bridge — Coders implement against them, reviewers validate against them. |

Mode-Specific Behavior

Pairing: Present draft for human review before writing. Human may redirect scope, resolve OQs, or confirm assumptions. Liza: Autonomous within task scope.

| Pairing Prompt | Liza Behavior | |----------------|---------------| | "Multiple capabilities — split?" | BLOCKED with split recommendation | | "LOW confidence assumption — resolve?" | Surface in Assumptions; human resolves before coding sprint | | "Adjacent spec may conflict?" | Read adjacent spec, declare in References, note in Context |