agentleFS
david-vrbadavid-vrba/claude-power-skills/skills/heavy-review/SKILL.md

heavy-review

Skill4 starsChanged 5 months agoMITView on GitHub

Adversarially review a heavy prompt in heavy/library/ before it is run. Hunts logical gaps, missing defaults, unattended-safety violations, resumability holes, output-spec errors, guardrail gaps, and cost-realism issues. Produces a severity-tagged report (BLOCKER / SHOULD / NICE) and a PASS / NEEDS-FIX verdict, then offers to apply fixes.

  • Reads credentials

The file

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 4 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read heavy/library/_TEMPLATE.md and heavy/README.md for the rules the prompt must satisfy.

    line 17
  2. Distinctive:

    Check every item. For each problem, record: severity (BLOCKER / SHOULD / NICE), the exact location, why it breaks, and the concrete fix.

    line 20
  3. Very specific:

    Every phase has steps; every step has ONE clearly-defined output (a file + an anchored SEC:id).

    line 24
  4. Distinctive:

    No missing step between inputs and an output that later steps assume exists.

    line 25
  5. Very specific:

    No internal contradictions (frontmatter vs body: phases list matches the phases written.

    line 27
  6. Very specific:

    EVERY decision the agent must make has an encoded default + the "take default, log to assumptions" rule.

    line 31
  7. Distinctive:

    All paths are explicit/relative-from-repo-root.

    line 32
  8. Distinctive:

    No step depends on information only a human would have.

    line 33
  9. Very specific:

    If there are BLOCKER/SHOULD findings, offer to apply the fixes directly to heavy/library/NN-*.md with Edit (keep edits minimal and within …

    line 64

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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

Heavy-prompt reviewer

A smart, skeptical review of one heavy prompt before a real run wastes a long unattended window on a flawed spec. Read adversarially: assume the unattended agent will take the dumbest valid interpretation of every instruction — find where that breaks. Work in your project repo (current working directory); the library lives under heavy/.

The argument $ARGUMENTS is the heavy-prompt number NN.

Step 1 — Load

  • Resolve heavy/library/NN-*.md (the prompt under review). If none, list heavy/library/INDEX.md and stop.
  • Read heavy/library/_TEMPLATE.md and heavy/README.md for the rules the prompt must satisfy.

Step 2 — Review against the rubric

Check every item. For each problem, record: severity (BLOCKER / SHOULD / NICE), the exact location, why it breaks, and the concrete fix.

A. Completeness & logic

  • Every phase has steps; every step has ONE clearly-defined output (a file + an anchored SEC:id).
  • No missing step between inputs and an output that later steps assume exists.
  • DONE criteria are concrete and checkable (not "do a good job"). They actually cover all OUTPUT SPEC files/sections.
  • No internal contradictions (frontmatter vs body: phases list matches the phases written; subagents/websearch_max match what the body tells the agent to do).

B. Unattended-safety (the big one)

  • The body NEVER instructs AskUserQuestion / waiting for the user. (Grep for it.)
  • EVERY decision the agent must make has an encoded default + the "take default, log to assumptions" rule.
  • All paths are explicit/relative-from-repo-root; no "the relevant file" hand-waving.
  • No step depends on information only a human would have.

C. Resumability

  • A RESUME step exists (read STATUS.json, skip completed_steps).
  • Steps are idempotent; output is append-only within a run (re-running a partial step rewrites only its own anchored block).
  • Nothing assumes the whole run happens in one window.

D. Output spec

  • Sections use stable <!-- SEC:id --> anchors with dotted semantic ids and the shared vocabulary.
  • 00-summary.md opens with the yaml machine block (run, status, top_findings, key_sections).
  • MAP.md regeneration is included (at least at synthesize; ideally each phase end).
  • heavy/output/INDEX.md update is in the synthesize phase.

E. Guardrails

  • subagents: false unless the user explicitly accepted the added complexity/billing; body doesn't spawn Agent/sub-workflows when false.
  • websearch_max present and the body respects it.
  • Security (secrets in env/.env only, no bypassing hooks), low-risk-edits-only if it touches live code (risky → backlog, not blind edits), verify-external-facts where it makes claims about APIs/pricing/tools.

F. Cost realism

  • est_cost matches the scope; websearch_max isn't absurd; phases aren't so huge a single window can't make real progress (resumability mitigates, but flag if a phase is monolithic).

G. Edge cases

  • What if a WebSearch returns junk / rate-limits? a step's input file is missing? the prompt finishes mid-window? double-launch? — the prompt should degrade gracefully, not hang.

Step 3 — Verdict

Print:

  • A findings table: severity | location | problem | fix.
  • Verdict: PASS (no BLOCKERs and no SHOULDs) or NEEDS-FIX (lists what must change).
  • A one-line readiness call: is this safe to launch unattended right now?

Step 4 — Offer to fix

If there are BLOCKER/SHOULD findings, offer to apply the fixes directly to heavy/library/NN-*.md with Edit (keep edits minimal and within the template). Re-state the verdict after fixing. Don't launch the prompt — that's /heavy NN.

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.