agentleFS
Fadi88Fadi88/one-man-startup/.claude/skills/spec-audit/SKILL.md

spec-audit

Skill16 starsChanged 17 days agoMITView on GitHub

Check whether a feature or codebase is specified well enough to hand to a developer agent - five layers (what it does, its interfaces, its architecture, how it ages, what good looks like here) - and name the decisions an agent would otherwise fill in silently. Use before the readiness gate on any feature that touches existing code, and at setup when a codebase already exists.

7 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 7 edited versions we read keep these.

  • 2

    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 only; change nothing.

    line 29
  2. Distinctive:

    Delegate: every layer COMPLETE, or THIN only where a wrong guess is cheap to undo.

    line 45

“Common” is measured: kept by at least 6 in 10 of the 7 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
<!-- Generated by tools/sync_adapters.py from skills/spec-audit/SKILL.md. Edit the source, not this file. -->

Spec audit

An agent given a thin spec does not stop and ask; it fills the gaps with whatever is most common in the code it learned from. This audit finds those gaps before the work is delegated, so the founder decides them instead of the agent's habits.

The five layers follow Hamidreza Saghir's framing of under-specified coding agents, as adapted in Florian Bruniaux's Claude Code guide; this version is written for this template's files.

The five layers, and where this template keeps each

LayerThe questionWhere it lives hereIf missing, the agent will
1. BehaviourWhat must it do, and how do we know it does?spec.md, the requirement rowsguess the behaviour
2. InterfacesWhat goes in and out, what errors, what must always hold?contract.md, types and validation in the codehandle errors optimistically, loosen types
3. ArchitectureWhat exists already, where are the boundaries, what must not be created?design.md, docs/adr/, AGENTS.md, skills/single-source/duplicate an existing module, cross a boundary
4. LifecycleWhat is deliberately unfinished, what is deferred and why?docs/HISTORY.md, tools/known_red.json, p1-open-questions.md, p1-closed-decisions.md"fix" a deliberate shortcut, or add a band-aid
5. CultureWhat does good look like in this project?p1-standing-instructions.md, project.toml [style], linters, the guidelines skillswrite average code in its own style

How to run it

Read only; change nothing.

  1. Name the scope: one feature (its work folder) or the whole codebase.
  2. For each layer, look at the files in the table and at a few real code files in the area the work will touch. Grep before reading (skills/context-budget/).
  3. Grade each layer: COMPLETE (an agent could work without guessing), THIN (some decisions left open) or MISSING. A grade cites the file, or the absence of one.
  4. For every THIN or MISSING layer, list the silent fills: the two or three concrete decisions an agent would make on its own. Examples: "creates a second auth helper because nothing says one exists"; "returns success when one of three writes failed because no error contract says otherwise".
  5. Mark the most dangerous fill: the one hardest to undo later (a data model, a boundary, a public interface).

Verdict

  • Delegate: every layer COMPLETE, or THIN only where a wrong guess is cheap to undo.
  • Delegate with review: THIN layers whose fills are listed in the brief, so the reviewer checks exactly those decisions.
  • Do not delegate yet: any MISSING layer whose most dangerous fill is hard to undo. Turn the fills into questions for the founder (skills/grilling/) or work for architect, then audit again.

Report

docs/work/<id>/spec-audit.md for a feature, docs/spec-audit-<date>.md for a codebase:

LayerGradeEvidenceSilent fills

then the verdict, the most dangerous fill, and the fixes in order of how much risk each removes for how little work. No score out of 100: the grades and the named fills are what the founder acts on.

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.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

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.