build-spec

build-spec — Interview-Driven Spec Builder

Execution Rules

• Sub-skill returns are internal values — when a skill invocation produces JSON output, capture it silently and proceed immediately. Do NOT present the result to the user or pause for confirmation.
• Only pause for user input at explicit AskUserQuestion steps — all other steps run to completion without stopping.

Interview Rules

• Mark the inferred default with [default].
• For choice questions, always include a free-text fallback as the last option (e.g. other: <describe>).

---

Phase 0: Entry Point

• 0.1 If the user provided a spec number (e.g. SPEC-003) as an argument or in their prompt:

  • Look it up — use skill solid-coder:find-spec with scan --status draft,ready,in-progress. Find the matching spec in the results.
  • If not found: report error and stop.
  • If found: load the spec, then load ancestors — use skill solid-coder:find-spec with ancestors <target-SPEC-NNN>. Read each file in the returned path fields (root → leaf). Hold all content as context.
  • Ask using AskUserQuestion:

    Found: SPEC-NNN — <feature> [<type>, <status>]
    What would you like to do?
    1. Edit or extend this spec
    2. Break down into subtasks
    3. Resume (continue drafting)
    

    • If option 1: set target_spec and proceed to Phase 3.5 (Extend).
    • If option 2: set target_spec and proceed to Phase 3.6 (Breakdown).
    • If option 3: set target_spec and proceed to Phase 2.2 (Load context), then Phase 4.

• 0.2 If the user provided a descriptive prompt (not a spec number, not empty):

  • Proceed to Phase 1 (Prompt-Aware Create).

• 0.3 Otherwise (no argument), ask using AskUserQuestion:

  What would you like to do?
  1. Create a new spec
  2. Resume a draft spec
  3. Edit an existing spec
  

  • If option 1: proceed to Phase 1 (with empty prompt — full interview).
  • If option 2: proceed to Phase 2.
  • If option 3: proceed to Phase 3.

---

Phase 1: Create New Spec

• 1.1 Analyze prompt — if the user provided a descriptive prompt, infer as much as possible:

  • Type — from language cues (e.g., "bug" → bug, "screen/view/feature" → feature, "break down/epic/initiative" → epic)
  • Title — the core noun/action from the prompt
  • Description — the full prompt, cleaned up
  • Parent — use skill solid-coder:find-spec with scan --status draft,ready,in-progress. Match the prompt topic against existing epics/features. If one clearly matches, suggest it. If ambiguous or none match, suggest root.
  • User stories (draft) — extract any behaviors, goals, or actions mentioned in the prompt. Format as stories. These are drafts — Phase 4 will refine them.
  • If no prompt was provided, set all inferences to empty.

• 1.2 Present inference — if any fields were inferred, present them via AskUserQuestion:

  Here's what I got from your prompt:
  
  Type: <inferred type>
  Title: <inferred title>
  Parent: <inferred parent or "root">
  
  Description: <inferred description>
  
  User Stories (draft):
  - <story 1>
  - <story 2>
  ...
  
  1. Looks good, continue to detail questions [default]
  2. Needs adjustments
  3. Start fresh (full interview)
  

  • If "looks good": store all inferred values, proceed to Phase 1.4.
  • If "needs adjustments": ask what to change, incorporate, re-present.
  • If "start fresh": clear all inferences, proceed to Phase 1.3.

• 1.3 Manual create (when no prompt, or user chose "start fresh"):

  • Run python3 ${CLAUDE_PLUGIN_ROOT}/skills/build-spec/scripts/build-spec-query.py types and ask using AskUserQuestion: "What type is this?"
  • Where does it belong? — use skill solid-coder:find-spec with navigate draft,ready Select as parent. Store as parent_epic.
  • "Give it a short title." — ask using AskUserQuestion free-text only.
  • "What are we building?" — ask using AskUserQuestion free-text description only.

• 1.4 Get next spec number — use skill solid-coder:find-spec with next-number. Store result as next_number.

Continue to Phase 4. If stories were inferred in 1.1, Phase 4.4 presents them as [default] for confirmation instead of asking from scratch. Same for any other Phase 4 question where the answer is already inferred from the prompt.

---

Phase 2: Resume Draft Spec

• 2.1 Find target — use skill solid-coder:find-spec with navigate draft Resume this. Returns selected spec as target_spec.

• 2.2 Load context — use skill solid-coder:find-spec with ancestors <target-SPEC-NNN>. Read each file in the returned path fields (root → leaf). Hold all content as context.

Continue to Phase 4.

---

Phase 3: Edit Existing Spec

• 3.1 Find target — use skill solid-coder:find-spec with navigate draft,ready,in-progress Edit this. Returns selected spec as target_spec.

• 3.2 Load spec — read the target spec file. Hold its full content as context.

• 3.3 Load ancestors — use skill solid-coder:find-spec with ancestors <target-SPEC-NNN>. Read each file in the returned path fields (root → leaf). Hold all content as context.

• 3.4 Ask using AskUserQuestion:

  What would you like to do with this spec?
  1. Extend or modify the spec
  2. Break down into subtasks
  

  • If option 1: proceed to Phase 3.5.
  • If option 2: proceed to Phase 3.6.

---

Phase 3.5: Extend / Modify Spec

• 3.5.1 Ask using AskUserQuestion: "What needs changing?" — free-text. User describes what to add, remove, or modify.

• 3.5.2 Apply the requested changes to the spec draft. Preserve all existing sections and frontmatter — only modify what the user asked to change.

Continue to Phase 6 (Buildability Gate) with the updated draft.

---

Phase 3.6: Break Down Into Subtasks (lightweight)

Do NOT rerun the full Phase 4 interview per child. Derive child specs from the parent's stories.

• 3.6.1 Analyze the target spec's user stories, acceptance criteria, and scope. Suggest a subtask breakdown — each story or story group becomes a subtask:

  Based on this spec, here's a suggested breakdown:
  1. <subtask title> — <one-line scope> (from story: <story reference>)
  2. <subtask title> — <one-line scope> (from story: <story reference>)
  ...
  

  Ask using AskUserQuestion: "Adjust, add, remove, or confirm these subtasks?"

• 3.6.2 Finalize the subtask list. For each subtask, get the next spec number — use skill solid-coder:find-spec with next-number before each.

• 3.6.3 For each subtask, generate a pre-filled draft directly from parent context:

  • Set type = subtask, parent = <target_spec number>, title from the breakdown list.
  • Derive stories, connections, I/O, technical requirements, and edge cases from the parent spec's content. The parent already has all the context — the child spec narrows scope, it doesn't re-discover it.
  • Run Phase 5 (Generate Draft) — no Phase 4 interview.
  • Run Phase 6 (Buildability Gate).
  • Present each child draft to the user with a single AskUserQuestion: "Write this subtask spec, or adjust?"
  • If "adjust": user provides changes in one free-text response, apply, then write.
  • Run Phase 8 (Write) — each subtask gets its own folder with Spec.md + resources/.

• 3.6.4 After all subtasks are written, confirm to user with list of all files created.

Done — do not continue to Phase 4.

---

Phase 4: Interview (3 steps)

Infer suggestions from loaded ancestor context. Present inferences as [default] options. Always include a free-text fallback. The goal is 3 user interactions for Phase 4 — not 9.

Step 1: Stories — "What is this?"

Stories are the anchor. Everything else flows from them.

• 4.1 User Stories — type-differentiated. Ask using AskUserQuestion.

  • Epic: "List the main things a user can do with this. I'll turn each into a story." Push for breadth — one story per major capability. Acceptance criteria can be high-level outcomes (each will be detailed in child specs).
  • Feature: "What are the distinct user goals this covers?" — 1–3 stories. For each: "What exactly makes this done?" Push for concrete, independently verifiable acceptance criteria.
  • Subtask: One story. Push for specificity — placement, connection, and integration details in acceptance criteria.
  • Bug: Skip — bugs use reproduction steps, not stories.

  If Phase 1 inferred stories from the prompt, present them as [default] for confirmation.

  Story format:

  • User-facing: As a [user], I want [goal] so that [reason]
  • Business logic / system: As the system, when [trigger], [outcome]

  Acceptance criteria rules: each criterion must be independently verifiable. No "works correctly", no "handles edge cases" — name the specific value, behavior, or condition.

  • Epic breakdown (only if type is epic): after confirming stories, suggest a feature breakdown:

    Here is a suggested feature breakdown:
    1. <feature 1>
    2. <feature 2>
    ...
    
    How should I proceed?
    1. Write as a single epic spec (features listed in Definition of Done)
    2. Write the epic + scaffold each feature now (frontmatter only, dependencies set later)
    

    Store choice as epic_mode (single or split).

Step 2: Context — "What surrounds it?"

Batch related context questions into a single AskUserQuestion with multiple questions (the tool supports up to 4 questions per call). Pre-fill from ancestor context and stories.

• 4.2 Ask using AskUserQuestion with up to 4 questions in one call:

  Q1 — Connections: "What does this connect to, depend on, or interact with?"

  • Suggest related features/modules inferred from ancestor specs. Free-text fallback.

  Q2 — Inputs & Outputs: "What goes in and what comes out? Who consumes the output?"

  • Suggest I/O patterns inferred from stories and sibling specs. Free-text fallback.

  Q3 — Technical constraints (conditional — skip for epics/bugs):

  • "What technical constraints apply? (APIs, libraries, patterns, error modes)"
  • Free-text fallback.

  Q4 — Dependencies: "Any specs this is blocked by or blocking?"

  • Use skill solid-coder:find-spec with scan --parent <parent_spec> --status draft,ready to get siblings.
  • Present as multi-select from results. Free-text fallback.

Step 3: Edge cases — "What could go wrong?"

Now asked AFTER stories exist — answers become acceptance criteria on specific stories.

• 4.3 Edge cases & design choices — infer edge cases from the stories and context already collected. Present inferred edge cases as [default] options. Ask using AskUserQuestion:
  • "Based on the stories, here are the edge cases and design choices I see. Confirm, add, or adjust."
  • Suggest: error conditions, boundary values, failure modes, race conditions inferred from the stories and technical context.
  • Free-text fallback.
  • Hold answers — these will be written as acceptance criteria on the relevant user stories, not as a separate section.

Silent generation (no user interaction)

These are generated automatically and shown in the draft for review in Phase 7. No separate confirmation step.

• 4.4 UI / Mockup — ask using AskUserQuestion:

  • If UI elements detected in description/stories: "Do you have design screenshots to include in resources/?"
    • If yes: write ## UI / Mockup\nReference screenshots are in the \resources/` directory adjacent to this spec file.`
    • If no: skip ## UI / Mockup section entirely.
  • If no UI elements detected: skip.

• 4.5 Test Plan — generate test cases from user stories and edge cases collected in steps 4.1–4.3:

  • Derive test cases from every acceptance criterion and edge case — both happy paths and failure/boundary paths
  • Language: "When [precondition or state], [action], [observable outcome]" — present tense, no "should", no "verify that"
  • Group by component or screen: Unit Tests per service/model, UI Tests per screen or flow (if UI spec)
  • Each case is one sentence — concrete and independently runnable
  • After generating, present to user via AskUserQuestion:

    Here are the test cases I derived. Add, remove, or adjust:
    ### Unit Tests — <ComponentName>
    - When ...
    ### UI Tests — <ScreenName>
    - When ...
    

  • Incorporate user's changes. Store final list for Phase 5.

• 4.6 Diagrams — generate Mermaid diagrams from the collected answers:

  • Connection diagram (all types): upstream inputs, downstream consumers, sibling specs, external dependencies.
  • Flow diagram (all types): epic = high-level user journey, feature/subtask = data/control path end-to-end.
  • Sequence diagram (conditional): generate if the spec mentions async operations, callbacks, delegates, notifications, network calls, or multiple distinct actors.
  • Embed all in draft under ## Diagrams. User reviews them in Phase 7.

---

Phase 5: Generate Draft Spec

• 5.1 Load the per-type rule — Read ${CLAUDE_PLUGIN_ROOT}/spec-driven-development/specs/<type>/rule.md for the spec's type (epic/, feature/, subtask/, or bug/). Also read specs/README.md for common frontmatter fields and section rules.
• 5.2 Generate the draft following the rule's required frontmatter, required sections (in order), and conditional sections. Apply the common section rules (story format, AC verifiability, no implementation leaking, test case format) and the type-specific Story Depth guidance from rule.md.
• 5.3 Bug-specific — For type: bug, generate ONLY Phase 1 (report) sections: Description, Steps to Reproduce, Expected vs Actual, Affected Component. Leave Phase 2 (Root Cause, Fix Plan, Diagrams, Test Plan, Definition of Done) for a later pass after investigation. The spec starts at status: draft as a report artifact.

The template is the single source of truth for structure — do NOT invent sections or omit required ones. For epic_mode = split: Epic = next_number, children = next_number+1…N, children get status: draft.

---

Phase 6: Buildability Gate (max 2 rounds)

Use skill solid-coder:validate-spec with --batch on the draft spec. Batch mode returns all findings at once instead of asking one-by-one.

For epic specs, validate-spec applies epic-specific rules (scope clarity, subtask breakdown completeness) instead of the standard buildability scan. See validate-spec for details.

• 6.1 Run validate-spec with --batch on the current draft. Collect all findings.
• 6.2 If findings exist, present ALL of them in a single AskUserQuestion (multiSelect: true):

  Buildability findings — select which to resolve now (rest will be marked TBD):
  1. [category] location — question
  2. [category] location — question
  ...
  

  User selects which findings to address and provides answers.
• 6.3 Incorporate selected answers into the draft. Mark unselected findings as TBD.
• 6.4 Re-run validate-spec with --batch. If new findings → repeat 6.2 (max 1 more round).
• 6.5 If gaps remain after 2 rounds, annotate remaining as TBD. Proceed to Phase 7.

---

Phase 7: User Review (max 2 rounds)

• 7.1 Present the full draft spec (includes diagrams and mockups generated in Phase 4).
• 7.2 Ask using AskUserQuestion:

  1. Yes, write it
  2. Specific changes needed — describe ALL changes at once
  

• 7.3 If "specific changes": user provides all changes in one free-text response. Apply all changes, re-present once. Max 1 re-present — after 2 total rounds, write what exists.
• 7.4 Proceed to Phase 8.

---

Phase 8: Write Spec File(s)

• 8.1 Derive slug: lowercase, spaces → hyphens, strip special chars.
• 8.2 Resolve output path — run:

  python3 ${CLAUDE_PLUGIN_ROOT}/skills/build-spec/scripts/build-spec-query.py resolve-path <type> <SPEC-NNN> <slug> [--parent <parent-SPEC-NNN>]
  

  If non-zero exit: stop and report error.
• 8.3 Create empty resources/ directory alongside each Spec.md written.
• 8.4 If epic_mode = split: write epic first (status: ready), then scaffold each child with frontmatter only — number, feature, type, status: draft, parent: SPEC-{epic-N}. No blocked-by wiring — dependencies are set later. Each child gets its own folder with Spec.md + resources/.
• 8.5 Confirm to user with list of all files written.

---

Constraints

• Do NOT create module files — only spec files.
• Fully written specs: status: ready. Scaffolded children: status: draft.
• blocked-by entries are permanent — never remove them.
• Re-scan spec numbers at runtime.
• For split mode, write ALL specs in this session.
• If any build-spec-query.py call exits non-zero, report and stop.