# Analyze Spec

> Spec analysis with ambiguity scoring — interactive locally, single-pass remotely via label

- Skill: `a16z/analyze-spec` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add a16z/analyze-spec`
- Raw SKILL.md: https://api.skillmd.com/api/skills/a16z/analyze-spec/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: a16z (https://skillmd.com/u/a16z)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/a16z/analyze-spec

---


<Purpose>
Analyze a spec file using mathematical ambiguity scoring. The goal: ensure the spec is clear enough for a one-shot implementation with zero clarifying questions.

This skill operates in two modes:
- **Local mode** (invoked via `/analyze-spec` in Claude Code): Full interactive Socratic interview — one question at a time, iterative refinement with the spec author.
- **Remote mode** (triggered externally via the `claude-spec-review-request` label): Single-pass analysis — all questions posted at once as a single PR comment. Reads prior PR comments as context to account for already-answered questions.

Adapted from the Ouroboros-inspired deep interview methodology — specification quality is the primary bottleneck in AI-assisted development.
</Purpose>

<Execution_Policy>
- Gather codebase facts via `explore` agent BEFORE asking about them
- Cite repo evidence (file path, symbol, or pattern) instead of asking the spec author to rediscover it
- Score ambiguity transparently
- Do not approve until ambiguity ≤ threshold (default 0.2)
- Allow early approval with a clear warning if ambiguity is still high
</Execution_Policy>

<Mode_Detection>
Detect which mode to use:
- **Remote mode**: Running in a remote Claude instance triggered by the `claude-spec-review-request` label. Indicators: environment is non-interactive (no TTY), or running inside a GitHub PR context.
- **Local mode**: Running interactively in a terminal via `/analyze-spec`.

When in doubt, default to local mode (interactive).
</Mode_Detection>

<Steps>

## Phase 1: Initialize

1. **Locate the spec**:
   - If a path is provided in `{{ARGUMENTS}}`, use that directly.
   - Otherwise, detect the PR number:
     - Run `gh pr view --json number --jq .number` to get the current branch's PR number.
     - If that fails, list specs: `ls specs/` and ask the user which one.
   - Look for `specs/<PR#>-*.md` matching the PR number. This is the spec for this PR.
   - If no match, fall back to finding any `specs/*.md` file that is NOT `TEMPLATE.md`.
   - If multiple specs match, prefer the one matching the PR number. If still ambiguous, ask the user.
2. **Read the spec** thoroughly — understand all sections (Summary, Intent, Evaluation, Design, Execution).
3. **Read `jolt-eval/README.md`** so you understand the invariant/objective framework for scoring Success Criteria and generating questions.
4. **Explore the codebase**: Run `explore` agent to map codebase areas relevant to the spec's intent.
5. **Read prior context (remote mode)**: Read all existing PR comments via `gh pr view --json comments` to identify questions already asked and answers already given. Account for these when scoring — don't re-ask answered questions.

## Phase 2: Analyze

Score clarity across four dimensions (0.0–1.0 each):

| Dimension | Weight | What to assess |
|-----------|--------|---------------|
| Goal Clarity | 0.35 | Is the primary objective unambiguous? Can you state it in one sentence? Are key entities and relationships clear? |
| Constraint Clarity | 0.20 | Are boundaries, limitations, and non-goals clear? |
| Success Criteria | 0.30 | Could you write a test that verifies success? Are acceptance criteria concrete? Are relevant `jolt-eval` invariants/objectives described? |
| Context Clarity | 0.15 | Do we understand the existing system well enough to modify it safely? |

**Calculate ambiguity:**
`ambiguity = 1 - (goal × 0.35 + constraints × 0.20 + criteria × 0.30 + context × 0.15)`

For each dimension below 0.9, generate a targeted question that would improve it:
- Questions should expose ASSUMPTIONS, not gather feature lists
- If the scope is conceptually fuzzy, ask an ontology-style question about what the thing fundamentally IS
- Cite specific codebase context (files, types, patterns) when relevant

## Phase 3: Output (mode-dependent)

- **Remote mode**: single-pass PR comment with all findings — follow `references/remote-mode.md` (in this skill's directory) for the comment template and label handling.
- **Local mode**: interactive Socratic interview — follow `references/local-mode.md` for the round protocol, challenge modes, and stop conditions.

</Steps>

<Examples>
<Good>
Probing jolt-eval coverage:
```
The Intent → Invariants section says "streaming must produce the same
commitments as the non-streaming path." That looks like a binary property —
have you considered capturing it as a new `jolt-eval` invariant? The existing
`split_eq_bind_low_high` in `jolt-eval/src/invariant/` is a close model
(reference vs. optimized implementation comparison). If this is out of scope,
the Invariants section should say so explicitly.
```
Why good: Names a concrete existing invariant as a model, leaves the N/A
door open, doesn't force a fit.
</Good>

<Good>
Citing codebase context:
```
I found that `DoryCommitmentScheme` in `crates/jolt-prover-legacy/src/poly/commitment/dory/`
uses a two-tier structure (tier-1 chunks → tier-2 aggregation). The spec mentions
"streaming commitments" but doesn't specify which tier is streamed. Is this streaming
at tier-1 (chunk-level), tier-2 (aggregation), or both?
```
Why good: Explored first, cited specific types and files.
</Good>
</Examples>

Task: Analyze the spec. {{ARGUMENTS}}