specsmith（specsmith 搜索）

Spec Smith

Turn ephemeral plans into structured, persistent specs built through deep research and iterative interviews. Specs have phases, tasks, acceptance criteria, a registry, resume context, a decision log, and a deviations log. They live in .specs/ at the project root and work with any AI coding tool that can read markdown.

Whether .specs/ is committed is repository policy. Respect .gitignore and the user's preference for tracked vs local-only spec state.

Critical Invariants

1. Single-file policy: Keep this workflow in one SKILL.md file.
2. Canonical paths:

- Registry: .specs/registry.md - Per-spec files: .specs/<id>/SPEC.md, .specs/<id>/research-*.md, .specs/<id>/interview-*.md

1. Authority rule: SPEC.md frontmatter is authoritative. Registry is a denormalized index for quick lookup.
2. Active-spec rule: Target exactly one active spec at a time.
3. Parser policy: Use best-effort parsing with clear warnings and repair guidance instead of hard failure on malformed rows.
4. Progress tracking is sacred: After completing any task, immediately update SPEC.md (checkbox, ← current marker, phase marker) AND registry.md (progress count, date). Then re-read both files to verify the edits landed correctly. Never move to the next task without updating both files. Never end a session with the registry out of sync with SPEC.md. This is non-negotiable — if you do nothing else, do this.

Claude Code Plugin

If running as a Claude Code plugin, slash commands like /specsmith:forge, /specsmith:resume, /specsmith:pause etc. are available. See the plugin's commands/ directory for the full set. The /forge command replaces plan mode with deep research, iterative interviews, and spec writing.

Session Start

If active-spec context was injected by host tooling, use it directly instead of reading files. Otherwise, fall back to reading files manually:

1. Read .specs/registry.md to check for a spec with active status
2. If one exists, briefly mention it: "You have an active spec: User Auth System (5/12 tasks, Phase 2). Say 'resume' to pick up where you left off."
3. Don't force it — the user might want to do something else first

Deterministic Edge Cases (Best-Effort)

Working on a Spec

Resuming

When the user says "resume", "what was I working on", or similar:

1. Read .specs/registry.md — find the spec with active status. If none, list specs and ask which to resume
2. Load .specs/<id>/SPEC.md
3. Parse progress:

- Count completed [x] vs total tasks per phase - Find current phase (first [in-progress] phase) - Find current task (← current marker, or first unchecked in current phase)

1. Read the Resume Context section
2. Present a compact summary: Resuming: User Auth System Progress: 5/12 tasks (Phase 2: OAuth Integration) Current: Implement Google OAuth callback handler Context: Token exchange is working. Need to handle the callback URL parsing and store refresh tokens in the user model. Next file: src/auth/oauth/google.ts
3. Begin working on the current task — don't wait for permission

Implementing

After completing each task, immediately edit the SPEC.md file to record progress. Do not wait until the end of a session or until asked — update the spec as you go. This is sacred (see Critical Invariant #6).

1. Check off the completed task: - [] -> - [x]
2. Move ← current to the next unchecked task
3. When all tasks in a phase are done:

- Phase status: [in-progress] -> [completed] - Next phase: [pending] -> [in-progress] - Review Acceptance Criteria — check off any that are now satisfied

1. Update the updated date in YAML frontmatter
2. Update progress (X/Y) and updated date in .specs/registry.md

Update transaction (required order — never skip steps):

1. Edit SPEC.md (checkbox, current marker, phase marker, resume context).
2. Recompute progress directly from SPEC.md checkboxes.
3. Edit the matching registry row (status, progress, updated date).
4. Verify: Re-read both SPEC.md and registry.md to confirm the edits are correct. If the registry progress doesn't match the SPEC.md checkbox count, fix it now.
5. If registry update fails, keep SPEC.md as source of truth and emit a warning with exact repair action for .specs/registry.md.

If you notice you forgot to update after a previous task, stop what you're doing and update now before continuing. Stale tracking is the single most common failure mode — it makes resume unreliable and the registry useless.

Also:

• If a task is more complex than expected, split it into subtasks
• Update resume context at natural pauses
• Log non-obvious technical decisions to the Decision Log
• If implementation diverges from the spec (errors found, better approach discovered, assumptions proved wrong), log it in the Deviations section

Pausing

When the user says "pause", switches specs, or a session is ending:

1. If there is no active spec, report that there is nothing to pause and stop.
2. Capture what was happening:

- Which task was in progress - What files were being modified (paths, function names) - Key decisions made this session - Any blockers or open questions

1. Write this to the Resume Context section in SPEC.md
2. Update checkboxes to reflect actual progress
3. Move ← current marker to the right task
4. Add any session decisions to the Decision Log
5. Update status: paused in frontmatter
6. Update the updated date

Resume Context is the most important part of pausing. Write it as if briefing a colleague who will pick up tomorrow. Include specific file paths, function names, and the exact next step. Vague context like "was working on auth" is useless — write "implementing verifyRefreshToken() in src/auth/tokens.ts, the JWT verification works but refresh rotation isn't hooked up to the /auth/refresh endpoint yet."

Switching Between Specs

1. Validate the target spec ID first. If missing, list available specs.
2. Confirm .specs/<target-id>/SPEC.md exists. If not, stop with an error.
3. If target is already active, report and stop.
4. Pause the current active spec if one exists (full pause workflow).
5. Set target status to active in frontmatter and in .specs/registry.md.
6. Resume the target spec (full resume workflow).

Command Ownership Map

• SKILL.md: global invariants, lifecycle rules, state authority, and conflict handling, plus cross-tool OpenAPI behavior.
• commands/*.md: command-specific entrypoints, prompts, and output shapes.
• If there is a conflict, preserve Critical Invariants from this file and apply command-specific behavior only where it does not violate invariants.

Spec Format

Frontmatter

YAML frontmatter with: id, title, status, created, updated, optional priority and tags.

Status values: active, paused, completed, archived

Phase Markers

[pending], [in-progress], [completed], [blocked]

Task Markers

• - [] [CODE-01] unchecked, - [x] [CODE-01] done
• Task codes: <PREFIX>-<NN> — prefix is a short (2-4 letter) uppercase abbreviation of the spec (e.g., user-auth-system → AUTH). Numbers auto-increment across all phases starting at 01
• ← current after the task text marks the active task
• [NEEDS CLARIFICATION] after the task code on unclear tasks

Acceptance Criteria

Testable conditions that define when the spec is "done". Written during forge, verified after each phase completes. Format: checkboxes with specific, verifiable statements — not vague goals.

## Acceptance Criteria

- [ ] Users can sign in with Google OAuth and receive a JWT
- [ ] Expired tokens return 401 with `{"error": "token_expired"}`
- [ ] Refresh tokens rotate on each use (old token is invalidated)
- [ ] Rate limiting returns 429 after 100 requests per minute

Check off criteria as they are satisfied during implementation. At phase completion, review which acceptance criteria are now met. At spec completion, all criteria must be checked — if any remain unchecked, the spec is not done.

Resume Context

Blockquote section with specific file paths, function names, and exact next step. This is what makes cross-session continuity work.

Decision Log

Markdown table with date, decision, and rationale columns. Log non-obvious technical choices (library selection, architecture pattern, API design).

Deviations

Markdown table tracking where implementation diverged from the spec: task, what the spec said, what was actually done, and why. Only log changes that would surprise someone comparing the spec to the code.

See references/spec-format.md for the full SPEC.md template.

Forging Specs

When asked to forge, plan, spec out, or "write a spec for X", follow the full forge workflow: setup, research deeply, interview the user, iterate until clear, then write the spec.

If the environment is in read-only plan mode, do not run forge in that mode. Ask the user to exit plan mode (Shift+Tab) and rerun /specsmith:forge.

The forge workflow never produces application code. Its outputs are only .specs/ files: research notes, interview notes, and the SPEC.md. If the user says "write a spec", that means write a SPEC.md — not implement the feature. Implementation happens separately, after the user reviews and approves the spec.

Step 1: Setup

1. Generate a spec ID from the title (lowercase, hyphenated): "User Auth System" -> user-auth-system
2. Collision check: If .specs/<id>/SPEC.md already exists or the ID appears in .specs/registry.md, warn the user and ask:

- Resume the existing spec - Rename the new spec (suggest <id>-v2 or ask for a new title) - Archive the old spec and create a new one in its place Do not proceed until the user chooses.

1. Initialize directories: mkdir -p.specs/<id>
2. If .specs/registry.md doesn't exist, initialize it: # Spec Registry | ID | Title | Status | Priority | Progress | Updated | |----|-------|--------|----------|----------|---------|

Step 2: Deep Research

Research is the foundation of a good spec. Be exhaustive — use every available resource. The goal is to gather enough context that the spec won't need revision mid-build.

Research runs on two parallel tracks to maximize thoroughness and speed:

Track A: Spawn the Researcher Agent

Always spawn the specsmith:researcher agent for codebase + internet research. Don't skip this — the researcher is purpose-built for exhaustive multi-source analysis and runs in parallel so it doesn't slow down the workflow.

Spawn it with the Task tool, providing:

• The user's request (what they want to build/change)
• The spec ID and output path: .specs/<id>/research-01.md
• Any Context7 findings you've already gathered (Track B)
• Specific areas to focus on, if known

The researcher will:

• Map the full project architecture (read manifests, lock files, directory tree)
• Read 15-30 relevant code files and trace dependency chains
• Run 3+ web searches for best practices and current patterns
• Compare 2-4 library candidates for every choice point
• Assess security risks and performance implications
• Produce a structured research document with a completeness checklist

Track B: Context7 & Cross-Skill Research (in parallel)

While the researcher runs, do these yourself — they use MCP tools that the researcher agent doesn't have access to:

• Context7: If available (resolve-library-id / query-docs tools), pull up-to-date documentation for every key library involved. Check API changes, deprecated features, and recommended patterns for the specific versions in use. Do this for 2-5 key libraries — the ones central to the feature being built.
• Cross-skill loading: Load other available skills when relevant:

- frontend-design: For UI-heavy specs — creative, professional design - datasmith-pg: For database specs — schema design, migrations, indexing - webapp-testing: For testing strategy — Playwright patterns - vercel-react-best-practices: For Next.js/React performance - Any other relevant skill that's available

• UI research (if applicable): Take screenshots, map component hierarchy, research modern UI patterns, note accessibility requirements

Merging Research

When the researcher agent completes, read its output at .specs/<id>/research-01.md. Merge your Context7 and cross-skill findings into the research notes — either append to the file or keep them in mind for the interview. The combined research should cover: architecture, relevant code, tech stack, library comparisons, internet research, Context7 docs, UI research (if applicable), risk assessment, and open questions.

Step 3: Interview Round 1

Present your research findings and ask targeted questions. Your research should inform specific questions, not generic ones.

1. Summarize findings (2-3 paragraphs — not a wall of text)
2. State assumptions — "Based on the codebase, I'm assuming we'll use X pattern because that's what similar features use. Correct?"
3. Ask 3-6 targeted questions that research couldn't answer:

- Architecture decisions ("New module or extend existing one?") - Scope boundaries ("Should this handle X edge case?") - Technical choices ("Stick with Library A or try Library B?") - User-facing behavior ("What should happen when X fails?") - Acceptance criteria ("What does 'done' look like? Any specific conditions that must be true when this is complete?")

1. Propose a rough approach and ask for reactions

STOP after presenting questions. Wait for the user to answer before proceeding. Do not answer your own questions, do not assume answers, and do not continue to Step 4 or Step 5 until the user has responded. The interview is a conversation — the user's answers shape the spec. If you skip this, the spec will be based on guesses instead of decisions.

Save to .specs/<id>/interview-01.md with: questions asked, user answers, key decisions, and any new research needed.

Step 4: Deeper Research + Interview Loop

Based on the user's answers, do another round of research — explore the specific paths they chose, check feasibility, find potential issues. Save to .specs/<id>/research-02.md.

Then present deeper findings and ask about trade-offs, edge cases, implementation sequence, and scope refinement. Save each interview round to interview-02.md, interview-03.md, etc.

Repeat research-then-interview until:

• You have enough clarity to write a spec with no ambiguous tasks
• The user is satisfied with the direction
• Every task can be described concretely (not "figure out X")

Two rounds is typical. Don't rush it — but don't drag it out either.

Step 5: Write the Spec

Synthesize all research notes, interview answers, and decisions into a comprehensive SPEC.md. See references/spec-format.md for the full template.

The spec should be thorough and detailed — someone reading it should be able to implement the feature without guessing. Include:

• YAML frontmatter (id, title, status, created, updated, priority, tags)
• Overview (2-4 sentences — what's being built and why)
• Acceptance Criteria — Testable conditions defining "done". Checkbox format so they can be checked off during implementation. Each criterion should be specific and verifiable, not vague ("works correctly").
• Architecture Diagram — ASCII art or Mermaid diagram showing the system architecture, data flow, or component relationships. Every non-trivial spec should have at least one diagram. Use ASCII for simple flows, Mermaid for complex relationships.
• Library Choices — Table comparing evaluated libraries with the selected pick and rationale. Include version numbers.
• Phases with status markers (3-6 phases is typical)
• Tasks as markdown checkboxes with task codes ([PREFIX-NN]) — be specific: include file paths, function names, and expected behavior
• Testing Strategy — Comprehensive testing plan: unit tests, integration tests, e2e tests, edge case tests. Specify which testing frameworks to use and what test files to create. Every feature task should have a corresponding test task.
• Resume Context section (blockquote)
• Decision Log with non-obvious technical choices from the interviews
• Deviations table (empty — filled during implementation)

Diagram guidelines:

• Use ASCII art for simple request flows and data pipelines: Client → API Gateway → Auth Middleware → Route Handler → Database ↓ Cache Layer
• Use Mermaid for complex architecture, state machines, and ER diagrams: graph TD A[Client] --> B[API Gateway] B --> C{Auth?} C -->|Yes| D[Handler] C -->|No| E[401]
• Include at least one diagram per spec (architecture, data flow, or state)

Solution quality standards:

• Proposed solutions should be simple, maintainable, and professional
• Prefer clean, modern patterns over clever hacks
• Choose the best available libraries — compare options, pick the most mature and well-maintained
• UI designs should be creative, sleek, and professional — not generic
• Code architecture should be innovative where appropriate but always clean

Coherence and logic review (mandatory before presenting):

1. Read through the entire spec as a whole — does it tell a coherent story?
2. Check that phases are in logical dependency order — no phase requires work from a later phase
3. Verify every task is concrete and actionable (file paths, function names)
4. Confirm the architecture diagram matches the task descriptions
5. Check that the testing strategy covers all feature tasks
6. Verify library choices are consistent throughout (no conflicting picks)
7. Ensure the overview accurately summarizes what the phases will deliver
8. Look for gaps — is there anything the implementation would need that isn't covered by a task?
9. Verify acceptance criteria are specific, testable, and cover the key behaviors the user expects

Save to .specs/<id>/SPEC.md. Update .specs/registry.md — set status to active.

Present the spec and wait for approval. Show the user the complete spec and ask: "Does this look right? Want to adjust anything before we start?" Do not begin implementing until the user explicitly approves. The forge workflow produces only spec files (SPEC.md, research-*.md, interview-*.md) — never application code. Implementation starts only after the user approves the spec and says to proceed.

Phase/task guidelines:

• Mark Phase 1 as [in-progress], the rest as [pending]
• Mark the first unchecked task with ← current

Implementing a Spec

When the user says "implement the spec", "implement phase N", "implement all phases", or similar:

Scope Detection

Parse the user's request to determine scope:

• "implement the spec" or "implement" → Start from the current task (the ← current marker) and work forward
• "implement phase N" or "implement phase " → Implement all tasks in that specific phase
• "implement all phases" or "implement everything" → Implement all remaining unchecked tasks across all phases, in order

Implementation Flow

1. Read .specs/registry.md to find the active spec
2. Load .specs/<id>/SPEC.md and parse phases/tasks
3. Identify the target tasks based on scope
4. For each task, in order: a. Mark it with ← current b. Implement it — write the actual code c. Check it off: - [] → - [x] d. Remove the ← current marker e. When all tasks in a phase complete:

- Phase status: [in-progress] → [completed] - Next phase: [pending] → [in-progress] f. Update updated date in frontmatter g. Update progress and date in .specs/registry.md

1. After each task completion, update Resume Context with current state
2. Log any new decisions to the Decision Log
3. If implementation diverges from the spec, log it in the Deviations section
4. If blocked on a task:

- Keep the task unchecked and record blocker details in Resume Context - Set phase marker [blocked] only when the whole phase is blocked - Continue with another unblocked task only if sequencing allows it

Testing During Implementation

When implementing, follow the testing strategy from the spec:

• Write tests as specified in the testing tasks
• Run tests after each task to verify correctness
• If a test task exists for the feature task you just completed, implement the test task immediately after

Completion

When all tasks are done:

• Verify all Acceptance Criteria are checked off. If any remain unchecked, report which ones and ask the user before marking the spec complete.
• Set all phases to [completed]
• Set spec status to completed in frontmatter and registry
• Update the updated date
• Present a summary of what was implemented
• Suggest next spec to activate if any are paused

Generating OpenAPI Docs

When the user says "generate openapi", "update api docs", or similar:

1. Scan the codebase for API routes/handlers/controllers and request/response schemas.
2. Infer auth/security schemes and endpoint grouping (tags).
3. Write .openapi/openapi.yaml (OpenAPI 3.1.1) with:

- operationId for every operation - Reusable components/schemas and $ref usage - Accurate parameters, request bodies, responses, and security

1. Write one endpoint doc per route under .openapi/endpoints/ using {method}-{path-slug}.md names (e.g., get-api-users-id.md).
2. Preserve manual additions in existing .openapi/ files when updating.
3. Report totals: endpoints, schemas, security schemes, and manual-review candidates.

Before Session Ends

If the session is ending:

1. Pause the active spec (run full pause workflow)
2. Write detailed resume context
3. Confirm to the user that context was saved

Directory Layout

All state lives in .specs/ at the project root:

.specs/
├── registry.md               # Denormalized index for status/progress lookups
└── <spec-id>/
    ├── SPEC.md               # The spec document
    ├── research-01.md        # Deep research findings
    ├── interview-01.md       # Interview notes
    └── ...

Registry Format

.specs/registry.md is a simple markdown table:

# Spec Registry

| ID | Title | Status | Priority | Progress | Updated |
|----|-------|--------|----------|----------|---------|
| user-auth-system | User Auth System | active | high | 5/12 | 2026-02-10 |
| api-refactor | API Refactoring | paused | medium | 2/8 | 2026-02-09 |

SPEC.md frontmatter is authoritative. The registry is a denormalized index for quick lookups. Always update both together — when you change status, progress, or dates in SPEC.md, immediately mirror those changes in the registry. If they ever conflict, SPEC.md wins.

Listing Specs

Read .specs/registry.md and present specs grouped by status:

Active:
  -> user-auth-system: User Auth System (5/12 tasks, Phase 2)

Paused:
  || api-refactor: API Refactoring (2/8 tasks, Phase 1)

Completed:
  ok ci-pipeline: CI Pipeline Setup (8/8 tasks)

Canonical Output Templates

Use these concise formats consistently:

Resume

Resuming: <Title> (<id>)
Progress: <done>/<total> tasks
Phase: <phase name>
Current: <task text>
Context: <one to three lines from Resume Context>

List

Active:
  -> <id>: <Title> (<done>/<total>, <phase>) [<priority>]
Paused:
  || <id>: <Title> (<done>/<total>, <phase>) [<priority>]
Completed:
  ok <id>: <Title> (<done>/<total>) [<priority>]

Status

<Title> [<status>, <priority>]
Created: <date> | Updated: <date>
Phase <n>: <name> [<marker>]
Progress: <done>/<total> (<pct>%)
Current: <task text or none>

Completing a Spec

1. Verify all tasks are checked (warn if not, but allow override)
2. Set status to completed in frontmatter and registry
3. Update the updated date in both
4. Suggest next spec to activate if any are paused

Archiving a Spec

Archive completed specs to keep the registry clean:

1. Set status to archived in frontmatter and registry
2. Research files (research-*.md, interview-*.md) in .specs/<id>/ can optionally be deleted (the SPEC.md has all the decisions and context)

Specs can be archived from completed or paused status. To reactivate an archived spec, set its status back to active.

Deleting a Spec

To remove a spec entirely:

1. Delete .specs/<id>/ (contains SPEC.md, research notes, interviews)
2. Remove the row from .specs/registry.md

This is irreversible — consider archiving instead if you might need it later.

Cross-Tool Compatibility

The spec format is pure markdown with YAML frontmatter. Any tool that can read and write files can use these specs:

• Claude Code: Full plugin support or skill via npx skills add
• Codex: Snippet in AGENTS.md or skill via npx skills add
• Cursor / Windsurf / Cline: Snippet in rules file
• Gemini CLI: Snippet in GEMINI.md
• Humans: Readable and editable in any text editor
• Git: Diffs cleanly, easy to track in version control

To configure another tool, run npx skills add ngvoicu/specsmith -a <tool>.

Behavioral Notes

Be proactive about spec management. If you notice the user has been working for a while and made progress, update the spec without being asked. If a session is ending, offer to pause and save context.

Specs should evolve. It's fine to add tasks, reorder phases, or split a phase into two as understanding deepens. Specs aren't contracts — they're living documents that adapt as you learn more about the problem.

The Decision Log matters. When the user makes a non-obvious technical choice (library selection, architecture pattern, API design), log it with the rationale. Future-you resuming this spec will thank present-you.

Don't over-structure. A spec with 3 phases and 15 tasks is useful. A spec with 12 phases and 80 tasks is a project plan, not a coding spec. Keep it lean enough to parse and act on in one read.

Respect the user's flow. Don't interrupt deep coding work to update the spec. Batch updates for natural pauses — task completion, phase transitions, or session boundaries.