analyze

Spec 024 introduces this skill as jig's cross-artifact consistency auditor. It is the seventh non-stub active jig skill that ships without a .py helper — analyze is fundamentally a judgment skill, and the determinism it needs (locate the spec, walk siblings, read the principles list, sort findings by severity) Claude can run inline via Read + Glob. The skill slots between READY_FOR_REVIEW (spec body ready for first review) and IN_PROGRESS (slice already in flight), giving the spec author one more pass to catch cross-artifact drift before the first review verdict.

Per user direction on 2026-05-18, jig's analyze ships as a standalone baseline, not a deferral surface. Power users who want spec-kit's richer /speckit.analyze install spec-kit and invoke it explicitly under /speckit.*. There is no category-based deferral hint in this skill's description.

What this skill does

Produces a non-destructive cross-artifact consistency report for a single spec across the six-category taxonomy at CRITICAL/HIGH/MEDIUM/LOW severity. Reporter only — no file writes. The skill reads one spec.md plus its sibling slice-NN-*.md files, plus a small set of read-only cross-reference docs (docs/product-vision.md for principles, docs/decisions/*.md for ADR resolution, docs/memory/glossary.md for terminology, docs/architecture.md for architecture-drift checks), and emits a markdown report to stdout: a findings table sorted by severity, a per-category coverage summary, and actionable next steps.

The skill is breadth over depth: surface the obvious cross-artifact drift within a few minutes, leave deep manual auditing to the spec author's judgment or to a richer downstream tool. The 50-findings ceiling exists to keep the scan scannable; if a spec generates more than 50 findings, the spec needs structural surgery, not a longer report.

When to use vs. when to defer

There are four sibling skills people often confuse with this one. Pick the right one:

• /jig:clarify — sibling skill for pre-DRAFT ambiguity scans. Clarify asks questions to the user and appends a ## Clarifications section to the target doc. It is prospective: "what's unanswered in this spec?" Analyze is retrospective: "where do these artifacts disagree?" Reach for /jig:clarify before the spec hits READY_FOR_REVIEW; reach for this skill once the spec body is stable enough to cross-check against ADRs, architecture.md, and the principles list.
• spec_lint.py — structural linter (frontmatter shape, slice numbering, file naming). Lint is structural; this skill is semantic. Lint catches "slice file missing status: frontmatter"; this skill catches "slice contradicts ADR-0003" or "spec proposes a fourth subagent type". The two layers are complementary — run python3 "${CLAUDE_PLUGIN_ROOT}/scripts/spec_lint.py" first to fix structural issues, then run this skill to find semantic drift.
• /jig:independent-review — sibling skill for spec-vs-implementation reviews. Independent-review reads a finished slice's deliverables and verifies they meet the ACs. This skill reads the spec body and ADRs/architecture and surfaces intra-spec drift. Different axis: independent-review evaluates implementation against spec; this skill evaluates spec against the project's other artifacts. Reach for /jig:independent-review after the slice is implemented; reach for this skill while the spec body is being finalized.
• /jig:pr-review — sibling skill for diff-shape reviews. PR review reads a git diff and surfaces blockers / nits / strengths. This skill reads a spec doc, not a diff. Different artifact shape: PR review evaluates a code change; this skill evaluates a spec document. Reach for /jig:pr-review after the slice's PR is open; reach for this skill while the spec body is being authored or refined.

Rule of thumb: draft a spec → /jig:clarify. Structural lint → spec_lint.py. Cross-artifact drift → this skill. Review the implementation → /jig:independent-review. Review the PR diff → /jig:pr-review.

Inputs

The MVP scans one spec at a time.

Primary input:

• One spec.md (e.g., docs/specs/024-analyze/spec.md) plus all sibling slice-NN-*.md files in the same directory. The skill walks the spec directory and treats every slice-*.md file as part of the audit scope.

Read-only secondary inputs (for cross-referencing):

• docs/product-vision.md — the seven principles for the Principle Violations finding category.
• docs/decisions/*.md — accepted ADRs for resolving ADR-NNNN references in the spec body.
• docs/memory/glossary.md — canonical glossary terms for the Terminology Drift finding category.
• docs/architecture.md — declared architectural decisions for architecture-drift checks.

Cross-spec input is explicitly NOT supported by the MVP. Auditing all of docs/specs/ for drift (e.g. "spec 010's ADR-0003 reference resolves to an ADR that's been superseded") is the territory of a future slice 024-02 if signal emerges. If the same friction surfaces three times across real usage, that slice can land. Until then, the user re-runs the skill per spec.

Six finding categories

The skill scans the target spec across six categories. Each finding gets a severity (CRITICAL / HIGH / MEDIUM / LOW). All findings are collected, sorted by severity, and rendered in the output. If a category produces zero findings, it still appears in the coverage summary table with 0 so the reader can verify the scan ran.

Duplication

Two or more places in the spec body (or across sibling slice files) say the same thing in subtly different ways. Watch for near-duplicate ACs, repeated constraints with conflicting phrasing, and ADR-NNNN restatements that re-declare a decision already captured in an ADR.

What triggers a finding:

• Two ACs in the same slice whose normalized text overlaps >70% (likely duplicate or near-duplicate).
• The same constraint stated in both spec.md and a sibling slice-NN-*.md file with conflicting phrasing (one says "exit 2", the other "exit 3").
• An ADR-NNNN restated as an AC in the spec body (the decision should live in the ADR; the AC should reference it, not duplicate it).
• A non-goal stated in ## Non-goals that is also implicit in a Goals bullet (e.g., "Goal: ship feature X" + "Non-goal: do not ship feature X immediately").

Ambiguity

Vague terms, placeholders, and unresolved TODOs that make the spec non-actionable. Watch for unbounded qualifiers and copy-pasted "TBD" markers that should have been resolved before READY_FOR_REVIEW.

What triggers a finding:

• Vague qualifiers without bounds: "fast", "scalable", "lightweight", "configurable" (without naming what knobs).
• Literal TBD / _TBD_ / <TBD> placeholders in non-Overview sections (Overview can stay TBD during DRAFT; ACs and Goals cannot).
• Unresolved TODO bullets in ACs or DoD checklists.
• Pronouns or demonstratives with no clear antecedent ("the helper", "this slice", "the user" — when multiple candidate referents exist).
• A ## Clarifications section exists but one Q entry has no answer yet (the question was asked but unanswered).

Underspecification

ACs without measurable outcomes; slices without dependencies declared; clarifications missing when the spec is non-trivial. Watch for the case where the spec gestures at a deliverable without saying how "done" is observed.

What triggers a finding:

• An AC without a measurable outcome verb — no "returns", "writes", "refuses", "logs", "exits", "creates", "raises". Aspirational ACs like "the helper should work well" trigger this.
• A slice with empty dependencies: [] frontmatter when prose references a precedent slice (e.g., "follows 022-02's pattern" but no dependencies: [022-02]).
• A spec with ≥5 ACs but no ## Clarifications section (suggests the spec didn't go through /jig:clarify despite non-trivial scope).
• A DoD checkbox referencing an artifact never produced anywhere in the spec body or slice files.

Principle Violations

The spec contradicts one or more of the seven principles in docs/product-vision.md § Design principles. This is the constitution-gate category — it codifies the same check independent-review now runs on every slice review (per slice 024-01 AC #6).

What triggers a finding:

• Spec proposes a fourth subagent type or a new role beyond implementer / reviewer / architect (principle 3 — three subagents, no more).
• Spec describes a new skill that duplicates an existing skill's responsibility without an ADR justification (principle 6 — no shims, also principle 4 — dogfooding).
• Spec assumes a context-window above 40% fill is fine (principle 2 — stay below the dumb zone).
• Spec describes a workflow rule that should be a hook but ships as a skill, or vice versa (principle 1 — hooks deterministic, skills judgment).
• Spec ships scaffolded files as plugin-internal-only with no scaffolding-mode equivalent (principle 7 — owning scaffolding beats renting).

Coverage Gaps

ACs without corresponding tests in the repo (or no test plan documented); slices that claim a hard decision without an ADR back-link; APIs introduced without a contracts-skill recommendation. Watch for ACs that ship without verification surface.

What triggers a finding:

• An AC mentions a behavior with no corresponding test path in the spec's test plan, and no test file matching the implied path exists in tests/ or skills/<name>/test_*.py.
• A slice claims an architectural decision ("we chose X over Y because Z") with no adr-NNNN-*.md back-link.
• An AC mentions a public API surface (HTTP endpoint, JSON schema, CLI output) without a /jig:contracts recommendation or a declared ## Contract surfaces entry in docs/architecture.md.
• A DoD checkbox names a verification step (e.g., "dogfood runs clean") with no instructions for how the dogfood is performed.

Terminology Drift

Glossary terms used inconsistently with docs/memory/glossary.md; ADR-NNNN references that don't resolve; dependency references naming non-existent slices. Watch for the same concept under two names, or two concepts collapsing under one.

What triggers a finding:

• A glossary term defined in docs/memory/glossary.md is used with a synonym interchangeably in the spec (e.g., glossary says "slice" but spec also uses "task" for the same concept).
• An ADR-NNNN reference (ADR-0003, adr-0003, decision 0003) doesn't resolve to an existing docs/decisions/adr-NNNN-*.md file.
• A dependency reference like dependencies: [050-01] names a slice that doesn't exist (no docs/specs/050-*/ directory).
• A skill name spelled inconsistently across the spec body (e.g., /jig:slice-land, slice-land, and "the land helper" all in one doc).
• A tier label spelled inconsistently (Tier 0 vs tier-0 vs "always-on" all in the same doc).

Severity scoring

Findings are scored against a four-level severity ladder. The model judges each finding case-by-case using these definitions:

• CRITICAL — the spec is internally contradictory or violates a load-bearing principle in a way that blocks READY_FOR_IMPLEMENTATION. Example: spec proposes a fourth subagent type (principle 3 violation). Example: an AC and a Non-goal directly contradict.
• HIGH — significant gap that should be resolved or explicitly accepted before merge. Example: an AC has no measurable outcome verb. Example: an ADR-NNNN reference doesn't resolve.
• MEDIUM — material drift that ships if explicitly accepted in the spec body or tracked in docs/inbox.md. Example: glossary term inconsistency. Example: ambiguous qualifier without bounds.
• LOW — minor stylistic or near-duplicate finding; can ship silently if the rest of the spec is clean. Example: same constraint restated in two adjacent bullets.

Principle-violation severity defaults: principles 1-3 (Hooks/skills, context economy, three subagents) map to HIGH by default — violating these undermines jig's core architecture. Principles 4-7 (dogfooding, deferral, no-shims, scaffolding-not-renting) map to MEDIUM by default — these are load-bearing but allow more case-by-case judgment. The model can override either default based on context (e.g., a principle-7 violation that strands all migrating users gets bumped to HIGH).

Output format

The skill emits a markdown report to stdout. No file writes. The exact shape:

# Analyze: <spec-id> — <spec-slug>
_Scanned 2026-MM-DD; <N> findings._

## Findings (sorted by severity)

| # | Severity | Category | Location | Finding |
|---|---|---|---|---|
| 1 | CRITICAL | Principle Violations | spec.md:42 | Spec proposes a 4th subagent type, violating principle 3. |
| 2 | HIGH | Coverage Gaps | slice-01-foo.md:18 | AC #2 has no corresponding test path. |
| ... | ... | ... | ... | ... |

## Coverage summary

| Category | Findings |
|---|---|
| Duplication | 0 |
| Ambiguity | 3 |
| Underspecification | 2 |
| Principle Violations | 1 |
| Coverage Gaps | 1 |
| Terminology Drift | 0 |

## Next steps

- Address the CRITICAL finding before READY_FOR_IMPLEMENTATION.
- HIGH findings should be resolved or explicitly accepted in
  the spec body before merge.
- MEDIUM/LOW findings can ship if tracked in `docs/inbox.md`.

Shape rules:

• Findings sorted CRITICAL → HIGH → MEDIUM → LOW. Within a severity band, ordering is by category (the six-category order in this document).
• If a category produces zero findings, it still appears in the Coverage summary table with 0 so the reader can verify the scan ran across all six.
• Max 50 findings per run. If the scan exceeds this, the findings table emits "(truncated at 50)" as a final row and the Coverage summary reflects the truncated count plus a parenthetical note ("Coverage shown; underlying scan was truncated at 50 findings — the spec needs structural surgery, not a longer report.").

Surfaced, never auto-resolved (issue #125). The report reports drift and hands off — it names each finding for the spec author to resolve. When two canon artifacts disagree, analyze does not adjudicate which one wins or write the resolution: picking a winner and amending a closed record is a separate, owner-authorised act (see the spec-workflow reconciliation checklist's closed-spec-drift gate). This posture is stronger than "no file writes", which only says the skill's own process is read-only; the point here is that a discovered conflict is escalated, not decided.

Gotchas

• Non-destructive — the skill never writes to disk. Even when the scan finds a CRITICAL drift, the skill does not edit any spec, ADR, or architecture file. It emits findings; the spec author resolves them. Same shape as /jig:pr-review and /jig:arch-review: reporter only.
• One-spec-at-a-time scope. The MVP scans one spec per invocation. If the same friction surfaces three times across real usage, slice 024-02 (deferred) can ship a cross-spec auditor. Until then, the user re-runs the skill per spec.
• Principle-violation severity is subject to judgment. The defaults (principles 1-3 → HIGH, principles 4-7 → MEDIUM) are starting points. The model can and should override based on context — a principle-7 violation that strands all migrating users gets bumped to CRITICAL; a principle-2 violation in a deliberately context-heavy dev tool gets demoted to LOW. The severity column is not a deterministic table lookup.
• No .py helper for analyze itself (no helper, judgment-only skill). Spec 024-01 explicitly ships SKILL.md only. The reviewer-prompt principles-check (AC #6) is a separate code change in skills/independent-review/review.py (a new _principles_check_block() helper appended unconditionally to both prompt builders), not part of analyze's runtime. If determinism friction surfaces three times — the model misses systematic drift across runs — slice 024-03 (deferred) can ship analyze.py gather to aggregate findings.

Relationship to other skills

• /jig:clarify — sibling (spec 023, sequenced before this skill). Clarify is prospective: it asks questions to fill gaps before the spec hits READY_FOR_REVIEW. Analyze is retrospective: it surfaces drift across artifacts that should already be aligned. The two compose: run /jig:clarify early in DRAFT, run this skill before or during READY_FOR_REVIEW.
• spec_lint.py — structural lint (frontmatter shape, slice numbering, file naming). Complementary, not overlapping. Lint catches "slice file missing status: frontmatter"; this skill catches "slice contradicts ADR-0003". Run lint first to fix structural issues, then run this skill to find semantic drift.
• /jig:independent-review — sibling, orthogonal axis. Independent-review evaluates implementation against spec; this skill evaluates spec against the project's other artifacts. They are not redundant: independent-review assumes the spec is internally consistent, this skill verifies it is.
• /jig:pr-review — sibling, orthogonal artifact. PR review reads a diff and surfaces blockers / nits / strengths. This skill reads a spec doc, not a diff. Different shape, different reading material.
• /jig:adr-workflow — downstream. Some analyze findings ("this AC restates an ADR — move it to the ADR body") will route the spec author to /jig:adr-workflow new or /jig:adr-workflow resolve-todo. Analyze surfaces the gap; adr-workflow fills it.
• /jig:contracts — downstream for Coverage Gaps findings. When analyze flags an AC that mentions an external API surface with no contract recommendation, the spec author runs /jig:contracts to add an OpenAPI / JSON Schema / etc. declaration.