agentleFS
metacircu1armetacircu1ar/hive_skills/skills/microskills/microskills-review-generator/SKILL.md

microskills-review-generator

Skill4 starsChanged 2 months agoMITView on GitHub

Study a codebase's stack and architecture and generate as many narrow, project-specific review checklists ("microskills") as the evidence supports, one skill file per microskill, then report where they were written. Use to prepare exhaustive single-family review passes, such as unchecked errors, lost async failures, leaked resources, missing timeouts or authorization gaps, for microskills-review-runner. It only writes microskill files: it never runs them, never changes project code and never executes builds or tests.

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 5 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Static study only: Read code, build files, configuration, documentation and repository instructions.

    line 46
  2. Distinctive:

    Start from the catalogue: Use references/catalogue.md.

    line 74
  3. Distinctive:

    Keep a family only when the code shows it applies, and cite the evidence.

    line 75
  4. Distinctive:

    Avoid overlap: When two microskills would check the same thing, assign it to one, and mention the boundary in both.

    line 90
  5. Very specific:

    Rank by expected yield: when --max applies: the hazards most common in this code first, and those with the worst failures if missed.

    line 92
  6. Distinctive:

    Check that each file's frontmatter parses as YAML before reporting it as written.

    line 105
  7. Very specific:

    When INDEX.md already exists, add the new rows and keep the existing ones.

    line 114

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “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

Microskills review generator

An agent finds far more when it is told exactly which problems to look for and how to recognize them. General review skills ask for "problems". A microskill is a short checklist for one family of defects, written for this codebase: its APIs, wrappers, conventions and risky places. This skill writes them. microskills-review-runner runs each one as an independent, exhaustive pass, a micropass, and merges the results.

Invocation

/microskills-review-generator [--output DIR] [--max N] [<target>]
  • --output DIR: where to write the microskills. It defaults to .microskills/ at the repository root. The directory is created if missing.
  • --max N: at most N new microskills. By default there is no limit: write as many as the evidence supports.
  • <target>: the code to study, a file or directory path. It defaults to the whole repository, with . meaning the same.

An invalid value requires a corrected invocation; do not guess.

Example activations

/microskills-review-generator
/microskills-review-generator --output /reviews/hive-microskills .
/microskills-review-generator --max 20 src/runtime/

Boundaries

  • Static study only. Read code, build files, configuration, documentation and repository instructions. Do not build, test, run the project, or run microskills.
  • Write only the microskill files and INDEX.md in the output directory. Never edit project code, stage, commit or push.
  • Never overwrite an existing microskill: the user owns those files. Skip names that already exist, and propose any changes to them in the report.
  • Launch no subagents. One agent studying the whole codebase keeps the microskills consistent and free of overlap.

1. Learn the codebase

Record the facts the microskills will depend on, with evidence: paths, packages, counts of call sites.

  • Stack: languages, frameworks, runtimes and build files.
  • Concurrency model: threads, async runtime, workers, queues and schedulers.
  • Persistence: databases, transactions, files, caches and migrations.
  • External I/O: HTTP and RPC clients, sockets, subprocesses and brokers.
  • Trust boundaries: CLI, HTTP, RPC, files, environment, plugins, and model or tool input.
  • Conventions: the error type and propagation style, the logger and its fields, wrappers and helpers, lint and policy gates, and existing review checklists.
  • Domain hazards: money, time, units, identity, permissions and multi-tenancy.
  • Hot spots: large or central modules, recent churn, and known-fragile areas named in docs or issues.

2. Choose the microskills

  1. Start from the catalogue. Use references/catalogue.md. Keep a family only when the code shows it applies, and cite the evidence. Add families the catalogue lacks whenever this architecture has its own hazards, such as a custom transaction layer, a plugin ABI, a cache with invalidation rules, or a state machine.
  2. Always include errors-and-logging, adapted to the project.
  3. Make each microskill narrow. One defect family each, with 3–8 checklist items. Split broad areas by the mechanism a reviewer must trace. For example, concurrency becomes shared-state races, locks held across await, lost background failures, and cancellation and shutdown.
  4. Make every item checkable. A site either has the defect or it does not. "Errors are handled" is a theme. "A Result is assigned to _ or dropped with .ok() outside tests" is a checkable item.
  5. Be specific to this codebase. Name its functions, types, wrappers, file areas and grep anchors, and the indirect forms a grep misses. State what is not a finding here, and the preferred fix.
  6. Avoid overlap. When two microskills would check the same thing, assign it to one, and mention the boundary in both.
  7. Rank by expected yield when --max applies: the hazards most common in this code first, and those with the worst failures if missed.

There is no target count. A small service may need a handful of microskills, and a large system several dozen. Write every microskill the evidence supports.

3. Write one skill file per microskill

For each microskill, write OUTPUT/<name>/SKILL.md using references/template.md. Each file is a complete skill: name and description in the frontmatter, then the checklist. It can be read or run on its own.

Check that each file's frontmatter parses as YAML before reporting it as written.

Then write OUTPUT/INDEX.md with:

  • the repository path, commit and generation date;
  • the stack facts from step 1;
  • a table with one row per microskill: name, family, scope, item count and evidence;
  • the families considered and skipped, and why.

When INDEX.md already exists, add the new rows and keep the existing ones.

4. Report

Return:

  • the absolute path of the output directory;
  • the list of microskills written: name, one-line purpose, and file path;
  • names skipped because they already exist, with proposed changes to them;
  • skipped families, and why;
  • the next command to run, for example /microskills-review-runner --skills /abs/path/.microskills --finders 12 .. Without --finders, the runner launches no agents and runs the passes one at a time. Leave the choice of N to the user.

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.