spec

Create or edit a spec from any input — plain English, PRDs, customer feedback, code files, or existing specs. The agent extracts structured rules from unstructured input. The user describes what they want; the agent writes the spec.

For syntax: references/formats/spec_format.md. For quality: references/spec_quality_guide.md.

Usage

purlin:spec <name>              Create or edit a spec
purlin:spec <name> --anchor     Delegate to purlin:anchor create
purlin:spec <name> --review     Review an existing spec for rule quality
purlin:spec                     (no name — extract from user's input)

---

--review Mode

Lightweight rule quality check. Reads an existing spec and evaluates each rule — no test execution, no file modifications, no proof files. Use this to catch quality problems before anyone writes tests.

Step 1 — Load the spec

Find specs/**/<name>.md. If not found, error: "Spec not found. Run purlin:spec <name> to create it."

Read the spec in full. Also read the source files listed in > Scope: (if any) for context on what the code actually does.

Step 2 — Evaluate each rule

For each RULE-N: line, apply the three tests from references/spec_quality_guide.md ("The rebuild test"):

1. Rebuild test: "If an engineer rebuilt this feature from only this spec, would they get this wrong without this rule?" If the answer is no — flag as NOISE.
2. Behavior test: Does this describe what the feature does, or how the code does it? Check for signals: names a library, hook, CSS value, token, internal function, or specific technique. If yes — flag as IMPLEMENTATION.
3. Overlap test: Would this rule always pass or fail together with another rule in the same spec? If yes — flag as OVERLAP and name the paired rule.

Also check:

• Missing coverage: Read the source code in > Scope: and identify behavioral aspects not covered by any rule. Flag each as GAP with a suggested rule.
• Vague rules: Rules that say "handle errors" or "work correctly" without specifying what happens. Flag as VAGUE.

Step 3 — Report

Spec review: <name> (<N> rules)

  ✓ RULE-1: Returns 200 with JWT on valid credentials
  ✓ RULE-2: Returns 401 on invalid password
  ⚠ RULE-3: Uses bcrypt for password hashing — IMPLEMENTATION (names library; rewrite as: "Passwords are hashed before storage")
  ⚠ RULE-4: Locks account after 5 failures — OVERLAP with RULE-5 (same trigger condition)
  ✗ RULE-5: Handles rate limiting — VAGUE (what status code? what error message? what threshold?)

Suggested rules (GAP):
  + "Returns 429 with Retry-After header when rate limit exceeded" (rate limit response not specified)
  + "Login audit log records every attempt with timestamp, IP, and success/failure" (audit trail in code but not in spec)

Summary: 2 clean, 2 warnings, 1 vague, 2 gaps suggested

For bad→good rewrite examples, reference references/rule_examples.md.

No writes

This mode is read-only. It does not modify the spec, create files, or run tests. The user takes the findings and applies them manually or via purlin:spec <name>.

---

Step 1 — Accept Input

Accept ANY of these input types without asking which format it is:

• Plain English description ("users should be able to reset their password via email")
• PRD or requirements document (pasted or referenced)
• Customer feedback or feature request
• Slack/email thread describing a problem
• Existing spec name to update
• A file path to code that needs a spec
• Image (screenshot, mockup, design comp, whiteboard photo) — create a design anchor. Images are locally owned. See "Image-Based Design Anchors" below.

If a spec name was given, search specs/**/<name>.md:

• Found: Read the spec. Call sync_status for coverage. Go to Step 7 (Update Existing Spec).
• Not found: Use the name and any additional input to draft a new spec.

If no name was given, read the user's input and infer the feature name from the topic.

Step 2 — Extract Everything Before Asking

If the input is substantial (more than a sentence or two), extract as much as possible BEFORE asking any questions:

• Feature name — infer from the topic
• Category — infer from the domain per references/spec_quality_guide.md ("Spec Categories"):
  • Auth, security → security/ or component dir with security_ anchor
  • File formats, contracts, cross-cutting standards → schema/
  • Reference docs, skill definitions, agent definitions → instructions/
  • End-to-end flows → integration/
  • Executable code → category matching the source directory
• Rules — extract every testable constraint mentioned or implied (see Step 5 heuristics)
• Proof descriptions — generate observable assertions with concrete inputs/outputs for each rule, with tier tags per the quality guide
• Anchors — if the input mentions security, compliance, design standards, check specs/ for matching anchors to reference via > Requires:
• Scope — if code files are mentioned or inferable
• Stack — if technologies are mentioned

Present the complete draft spec to the user, followed by the approval block. Assumed rules should be visually obvious in the draft:

RULE-1: Fetches weather from OpenWeatherMap
RULE-2: Shows temperature in Fahrenheit
RULE-3: Cache results for 10 minutes (assumed — user said "don't hit the API too much")
RULE-4: Show error message on API failure

The PM sees the assumed tag and either confirms, changes the value, or asks a gap question.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚡ REVIEW DRAFT — Does this look right?

  [y] Looks good — continue to metadata
  [n] Start over
  [edit] I want to change specific rules or proofs

Waiting for your response...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Use AskUserQuestion to pause and wait. Do NOT skip this step. Then go to Step 3 for gap questions.

Step 3 — Gap Questions (only what's missing)

After presenting the draft, ask ONLY about gaps — don't re-ask things the input already answered:

• "I couldn't determine the tech stack. What framework are you using?"
• "The input mentions authentication but I don't see a security anchor in your project. Should I create one?"
• "You mentioned 'fast response times' — what's the specific threshold? Under 200ms? Under 1 second?"
• "Is the email notification handled by your app or a third-party service?"
• "Your PRD mentions 3 requirements — are there others I missed?"

Questions must be specific and gap-filling, not generic ("what are the rules?" or "what are your requirements?").

Step 4 — Show an Example First (for novice users)

If the input is vague or minimal (e.g., "write a spec for password reset" with no details), show a brief example BEFORE generating:

Here's what a spec looks like:

Feature: password_reset

What it does — Allows users to reset their password via a time-limited email link.

Rules:

• RULE-1: POST /reset with valid email sends a reset link
• RULE-2: Reset link expires after 24 hours
• RULE-3: Clicking a valid link allows setting a new password
• RULE-4: Clicking an expired link shows an error message

Proofs:

• PROOF-1 (RULE-1): POST /reset with registered email; verify 200 and email sent @integration
• PROOF-2 (RULE-2): Create a link, advance clock 24h; verify link returns 410 Gone
• PROOF-3 (RULE-3): Click valid link, submit new password; verify password changed @integration
• PROOF-4 (RULE-4): Click expired link; verify error message displayed

Now tell me about your feature and I'll draft the spec.

Skip the example if the input is already detailed (PRD, substantial description, pasted requirements).

Step 5 — Rule Extraction Heuristics

When extracting rules from unstructured input, look for:

Signal in input	Rule type	Example
"must", "should", "needs to", "has to"	Direct constraint	RULE: Return 200 with session token on valid login
"never", "don't", "cannot", "forbidden"	FORBIDDEN pattern	RULE: FORBIDDEN — No plaintext password storage
Error cases, "what if", "fails when"	Error handling	RULE: Return 404 when user ID does not exist
"fast", "under N seconds", "real-time"	Boundary condition	RULE: API response time under 200ms at p95
"first... then... finally"	Multi-step workflow	RULE: After email verification, account status changes to active
"only admins", "users with role X"	Access control	RULE: Only users with admin role can delete accounts
"stores", "saves", "records", "tracks"	Data persistence	RULE: Audit log records every login attempt with timestamp and IP
Sequences with conditions	State machine	RULE: Order status transitions: pending → paid → shipped → delivered

For each extracted rule, generate an observable proof description with concrete inputs and expected outputs. Apply tier tags per references/spec_quality_guide.md ("Tier Tags on Proofs").

Assumption tagging (mandatory): After extracting rules, review each one. If the rule contains a specific number, threshold, algorithm, or constraint that the user did NOT explicitly state, add the (assumed) tag with context:

• User said "fast" → RULE-3: Under 500ms (assumed — user said "fast")
• User said "secure" → RULE-4: bcrypt hashing (assumed — user said "secure")
• User said "handles errors" → RULE-5: Returns 500 with error body (assumed — user said "handles errors")

If the user WAS explicit, no tag:

• User said "must return in under 200ms" → RULE-3: Under 200ms (no tag)
• User said "use argon2" → RULE-4: argon2 hashing (no tag)

Step 6 — Enhance with Metadata

After the core spec (What it does, Rules, Proof) is solid, add metadata:

1. Scan for matching anchors in specs/ and suggest > Requires: based on scope overlap:
   • For each anchor (specs in specs/_anchors/), read its > Scope: patterns
   • If the feature's > Scope: files overlap with the anchor's scope, suggest requiring it
   • Also note any global anchors (these are auto-applied and don't need > Requires:)
   • Present suggestions:

     Suggested > Requires: based on file overlap:
       api_rest_conventions — your Scope overlaps with src/api/
       security_no_eval — global anchor (auto-applied, no action needed)
     Add api_rest_conventions to > Requires:? [y/n]
     

2. If the user mentioned code files, populate > Scope: (verify paths exist)
3. If the user mentioned technologies, populate > Stack:
4. Apply tier tags to proofs per references/spec_quality_guide.md
5. Check if any rules are FORBIDDEN patterns and format proofs as grep-based assertions per "FORBIDDEN Grep Precision"
6. Suggest the category per references/spec_quality_guide.md ("Spec Categories")

Rule quality check (mandatory): Before presenting, apply the --review logic internally: evaluate every rule against the rebuild/behavior/overlap tests. Fix any IMPLEMENTATION or NOISE rules in the draft — don't present rules that fail the rebuild test.

Present the enhanced spec with metadata added and ask "anything to adjust?"

Structural-Only Proof Check

After drafting the rules and proofs, if the spec covers files in references/, skills/, or agents/ (instruction files), check: are ALL proofs grep-based or existence checks? If yes, suggest adding behavioral rules to this same spec — not a separate spec:

All proofs for this spec are structural (grep/existence checks). This catches
deletions and drift but doesn't prove the instructions work.

Consider adding behavioral rules to this spec. For example:
  RULE-N: Agent follows the core loop when given "build X" @e2e
  RULE-N+1: Agent uses purlin:spec when asked to "update the spec" @e2e

NEVER Create Test-Only Specs

Tests must prove rules in the feature they validate — not in a separate spec.

Do NOT create specs whose sole purpose is to be a container for tests (e.g., e2e_feature_scoped_overwrite, e2e_audit_cache_pipeline). If a test validates that proof plugins preserve other features during overwrite, that test proves proof_common RULE-4 — wire it there.

When the user asks for "an e2e spec" or "integration tests for X":

1. Identify which existing feature spec the behavior belongs to
2. Add rules to THAT spec if they don't already exist
3. Write proof descriptions under THAT spec's ## Proof section
4. Never create a parallel spec just because the tests are e2e tier

If the rule already exists in the target spec, the test just needs a proof marker pointing to it — no new rule or spec needed.

Validate Before Commit

Before committing, verify:

• ## What it does has at least one full sentence
• ## Rules has at least one RULE-N: line, all numbered sequentially
• ## Proof has at least one PROOF-N (RULE-N): line, each mapping to a rule
• Proof descriptions are observable assertions, not vague instructions
• Every proof description has an appropriate tier tag per references/spec_quality_guide.md ("Tier Tags on Proofs")
• Every @e2e proof describes an observable flow (arrange → act → observe through the real running app) and does not name a source file or internal function, per references/spec_quality_guide.md ("E2E proof descriptions") — rewrite as a boundary observation or retag if not
• All > Scope: file paths exist on disk
• > Requires: validation (blocking): For EACH reference in > Requires:, glob specs/**/<name>.md. If any referenced spec does not exist on disk, DO NOT commit the spec with the broken reference. Remove the broken reference from > Requires: and print: Removed > Requires: <name> — spec not found. Create it first with purlin:spec <name>, then add the reference back.

Commit (mandatory)

After the spec is approved, commit immediately. Do not batch with other changes.

git add specs/<category>/<name>.md
git commit -m "spec(<name>): <description of change>"

This commit is mandatory — drift detection and staleness checks depend on committed spec state. Do not skip or defer.

Step 7 — Update Existing Spec

When updating a spec that already exists on disk:

7a — Understand what changed

1. Call the drift MCP tool to see what changed since the last verification
2. Read the CURRENT spec in full — every rule, every proof, every metadata line
3. Read the changed source files referenced in the diff or in > Scope:
4. Read changed skill/reference files if the feature covers instructions

7b — Identify deltas

Compare the current spec against the code changes. Categorize each finding:

Category	What it means	Action
NEW RULE NEEDED	Code added behavior not covered by any existing rule	Propose adding a new RULE-N
RULE OUTDATED	Existing rule describes behavior that changed	Propose updating the rule text
RULE OBSOLETE	Existing rule describes behavior that was removed	Propose removing the rule
PROOF OUTDATED	Proof description no longer matches the rule	Propose updating the proof text
METADATA STALE	> Scope: or > Stack: no longer accurate	Propose updating metadata
NO CHANGE	Existing rule still matches code	Keep as-is (explicitly note this)

7c — Present the delta report

Show the user EXACTLY what will change and what will stay:

Spec: specs/mcp/sync_status.md (15 rules currently)

KEEPING (unchanged):
  RULE-1: sync_status tool returns valid report ✓
  RULE-2: Proof files are parsed correctly ✓
  RULE-3: Coverage is computed for all features ✓
  ...

ADDING:
  RULE-23 (new): sync_status scans specs/_anchors/ for anchor specs
    Reason: Anchor directory support added
    Proposed proof: PROOF-23 (RULE-23): Create anchor in specs/_anchors/, run sync_status, verify anchor rules appear

UPDATING:
  RULE-20 (changed): Global anchors auto-apply to all features
    Was: "Global specs with > Global: true auto-apply"
    Now: "Anchor specs in specs/_anchors/ with > Global: true auto-apply to all non-anchor features"
    Reason: Anchor unification

REMOVING:
  (none)

METADATA:
  ▎ Stack: unchanged
  ▎ Scope: unchanged
  ▎ Requires: unchanged

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚡ APPROVAL REQUIRED — Review the changes above.

  [y] Approve and apply all changes
  [n] Cancel — make no changes
  [edit] I want to adjust specific items

Waiting for your response...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

This approval block is MANDATORY. The agent MUST use AskUserQuestion to pause and wait for the user's response. Do NOT auto-approve. Do NOT proceed without an explicit answer. The bordered block above is the exact format to use — it must be visually distinct from the rest of the output so the user doesn't scroll past it.

7d — Apply approved changes

1. Preserve ALL unchanged rules exactly as they are — same text, same numbering
2. Add new rules at the end of the existing sequence (RULE-9 after RULE-8)
3. Update changed rules in place — same RULE-N number, new text
4. For removed rules: renumber remaining rules sequentially (no gaps)
5. For each new or updated rule, add or update the corresponding PROOF-N line
6. Apply tier tags to new proofs per references/spec_quality_guide.md
7. Preserve any existing @manual stamps — do NOT remove manual proof stamps unless the rule they reference was removed
8. Preserve rule tags on unchanged rules — (assumed), (confirmed), (deferred) stay as-is
9. If updating a rule that had (confirmed), change the tag to (assumed) since the new value hasn't been confirmed yet
10. Update > Scope: if new files were added to the feature
11. Update > Stack: if new dependencies were introduced

7e — Validate and commit

Same validation as new specs: no empty sections, sequential numbering, observable proofs, valid references, tier tags.

Commit (mandatory)

After the spec update is approved, commit immediately. Do not batch with other changes.

git add specs/<category>/<name>.md
git commit -m "spec(<name>): update rules for <description>"

This commit is mandatory — drift detection and staleness checks depend on committed spec state. Do not skip or defer.

Key principles for updates

• Never silently change existing rules. Every change must be shown to the user.
• Never remove @manual stamps unless the rule was deleted.
• Preserve rule numbering when possible — renumber only when rules are removed.
• Show what's staying, not just what's changing. The user needs to see the full picture to approve confidently.
• Ask before applying. The delta report is a proposal, not a fait accompli.

---

Anchor Specs

When --anchor is specified, delegate to purlin:anchor create with the same arguments. Do not create the anchor directly in this skill.

See references/formats/anchor_format.md for format, references/spec_quality_guide.md for when to create anchors and FORBIDDEN pattern guidance.

Image-Based Design Anchors

When the user provides an image (screenshot, mockup, design comp, whiteboard photo):

1. Save the image to specs/_anchors/screenshots/<name>.png (create the screenshots directory if needed).

2. Compute and store the image hash in the spec metadata:

   > Visual-Reference: ./specs/_anchors/screenshots/<name>.png
   > Visual-Hash: sha256:a1b2c3d4e5f6...
   

   This enables staleness detection — if the image is later replaced, sync_status will warn that the anchor may need review.

3. Create an anchor spec with:

   • > Visual-Reference: and > Visual-Hash: as above
   • > Scope: pointing to the code files that will implement the design
   • One visual match rule: RULE-1: Implementation must visually match the reference image
   • One screenshot comparison proof (using the project's test framework) tagged @e2e
   • If the user describes behavioral requirements alongside the image, add those as additional rules

4. Commit the image AND the spec together.

---

Key Principles

• The user describes what they want in their language. The agent writes RULE-N format.
• Extract everything possible from the input before asking questions.
• Questions are specific and gap-filling, not generic.
• Show, don't tell. Present the draft spec, then refine — don't ask the user to write rules.
• Progressive disclosure: core spec first, metadata second, anchors third.

Exit Criteria

The spec operation is NOT complete until all of the following are true. Verify each one before responding to the user.

1. Spec file committed. Run git status. If the spec .md file is uncommitted, commit it now per the commit instructions above.
2. No uncommitted spec files. git status must not show any modified or untracked specs/**/*.md files from this session.

If any criterion is not met, fix it before completing. Do not respond to the user with "done" or "complete" until both are verified.