agentleFS
saffron-healthsaffron-health/libretto/.agents/skills/spec-review/SKILL.md

spec-review

Skill883 starsChanged 3 months agoMITView on GitHub

Review a spec for under-specified areas, bugs, and adherence to the spec requirements in this skill. Use when asked to review, critique, or check a spec.

We split this skill into the rules it gives an agent

10 rules. 0 are common, 10 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 36 edited versions we read keep these.

  • 10

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the spec file the user references.

    line 10
  2. Distinctive:

    Use sub-agents to read every file listed in the spec's "Important files" section.

    line 14
  3. Distinctive:

    If the spec references external libraries or APIs, look up their documentation.

    line 16
  4. Distinctive:

    Do not flag things that are obviously implied by context.

    line 32
  5. Distinctive:

    Test assertion details — the test runner catches these

    line 47
  6. Distinctive:

    Check whether the spec follows the structure and principles listed in this skill:

    line 52
  7. Distinctive:

    Do not check whether guiding questions were asked — the reviewer does not have context about the conversation that produced the spec.

    line 62
  8. Distinctive:

    If a category has no findings, write "None" under it.

    line 88
  9. Distinctive:

    Do not suggest scope additions.

    line 95
  10. Distinctive:

    Do not comment on writing style or formatting unless it violates the structure requirements in this skill.

    line 96

“Common” is measured: kept by at least 6 in 10 of the 36 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Review the spec the user points you to. Follow these steps in order.

Step 1: Read the spec

Read the spec file the user references.

Step 2: Research the codebase

Use sub-agents to read every file listed in the spec's "Important files" section. For each file, understand its current state well enough to evaluate whether the spec's proposed changes are correct and complete.

If the spec references external libraries or APIs, look up their documentation.

Step 3: Evaluate

Assess the spec against three categories. For each finding, cite the specific spec line or section and the relevant source file and line number.

Under-specified areas

Find places where the spec leaves ambiguous decisions that should have explicit design intent from the user. Examples:

  • A phase says "update the handler" but two handlers exist and the spec does not say which one
  • A schema change is described but the migration path for existing data on disk is not addressed
  • A new dependency is introduced between two modules but the spec does not state which module owns the interface
  • Error handling or edge cases are not mentioned for a path that can clearly fail
  • The spec describes behavior but does not specify what happens when the precondition is not met

Do not flag things that are obviously implied by context. Flag things where a reasonable implementer would have to guess.

Bugs

Find design-level logical errors — things that would send an implementer down the wrong path or produce incorrect behavior that tooling won't catch. Examples:

  • A phase deletes a module but a later phase depends on the concept it provided (not just the import — the typecheck catches that)
  • The spec contradicts its own goals or non-goals (e.g., says "calendar support later" but removes the only calendar mechanism without acknowledging it)
  • A fan-out loop creates sessions but nothing deduplicates if the same event matches the same trigger twice
  • The spec assumes a data shape or behavior that doesn't match reality in the codebase

Do not flag:

  • Missing imports/exports — the typechecker catches these
  • Exact function signatures or parameter mismatches — the implementer will resolve these during development
  • Test assertion details — the test runner catches these
  • Wiring specifics that are obvious from context (e.g., "this function isn't exported yet" — that's an implementation detail, not a spec concern)

Adherence to spec requirements

Check whether the spec follows the structure and principles listed in this skill:

  • Does it have all required sections (problem overview, solution overview, goals, non-goals, important files, implementation)?
  • Is each phase commit-sized (under 100 lines of change)?
  • Does each phase have success criteria that verify the thing most likely to go wrong?
  • Does the spec avoid speccing extensibility, abstraction layers, or polish that nobody asked for?
  • Are there phases that are pure setup or refactoring with no user-facing progress?
  • Is the sanity checklist present and correct?
  • Does the spec represent the simplest version that works end-to-end?

Do not check whether guiding questions were asked — the reviewer does not have context about the conversation that produced the spec. Assume all decisions in the spec were intentional.

Step 4: Structure your response

Use this format:

# Spec Review: [spec title]

## Under-specified

- [Finding with spec section reference and source file reference]

## Bugs

- [Finding with spec section reference and source file reference]

## Adherence

- [Finding referencing specific requirement from this skill]

## Verdict

[2-3 sentences. State whether the spec is ready to implement, needs minor clarifications, or needs significant rework.]

If a category has no findings, write "None" under it.

Principles

  • Review the plan, not the implementation. A spec is a design document, not a line-by-line coding guide. Don't flag things that tooling (typechecker, linter, test runner) will catch during implementation. Focus on decisions that, if wrong, would waste significant time or produce incorrect behavior.
  • Be specific. Every finding must reference a spec section and a source file or requirement in this skill.
  • Be conservative. Only flag things that would cause an implementer to go down the wrong path, produce a design-level bug, or violate an explicit rule in this skill.
  • Do not suggest scope additions. The goal is to find holes in the existing scope, not to expand it.
  • Do not comment on writing style or formatting unless it violates the structure requirements in this skill.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.