agentleFS
mattdweigand-sketchmattdweigand-sketch/agent-skills/skills/60-30-10/SKILL.md

60-30-10

Skill1 starsChanged 40 days agoMITView on GitHub

Evaluate how a project allocates judgment across owned data, deterministic code, and prompt-model work. Invocation-only: use this skill only when the user explicitly invokes `$60-30-10` or `/60-30-10`, or explicitly asks to use or run the 60-30-10 skill. Do not infer its use from related audit, architecture, prompt, agent, harness, workflow, or judgment-allocation requests.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    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. Very specific:

    Run this workflow only when the user directly invokes $60-30-10 or /60-30-10, or explicitly asks to use or run the 60-30-10 skill.

    line 10
  2. Distinctive:

    If the target has no reusable judgment, say so in one sentence and stop.

    line 24
  3. Very specific:

    Render the verdict: Use the template and field specification in references/output-template.md.

    line 34

≈ “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

Composition Audit

Invocation Gate

Run this workflow only when the user directly invokes $60-30-10 or /60-30-10, or explicitly asks to use or run the 60-30-10 skill. Do not trigger it from topic similarity or from phrases such as composition audit, structure audit, prompt allocation, or judgment allocation alone.

Evaluate whether a project puts each kind of judgment in its most durable home: facts and chosen policy in owned data, consequence-worthy checks and fixed execution in deterministic code, and only genuine interpretation and steering in prompt-model work.

In this skill, 60/30/10 always means owned data / deterministic code / prompt-model work. These are local operational buckets; see source framings for the course's revised data-and-thinking / existing-tools / AI framing and older code/rules/AI heuristics.

The principle this enforces: every necessary piece of judgment has a right home, and the default home, prose in a prompt, is usually the wrong one. 60/30/10 is a memorable directional guideline for that principle, not a quota, score, compliance threshold, codebase-composition claim, or required numerical result. A healthy system routes each piece to the bucket that fits; its actual ratio may differ substantially for good domain-specific reasons. The failure mode is not missing an exact percentage. It is manufacturing live decisions for fixed sequences or trapping durable and consequence-worthy judgment in perishable model context.

Produce an inline verdict. Do not write a file unless asked. When the user asks for a saved verdict, validate its structure before delivery.

Applicability Gate

Audit the target only if it holds reusable judgment: rules, policies, or criteria that decide outputs across runs. This includes agent harnesses, skills, AGENTS.md or CLAUDE.md systems, MCP workflows, eval pipelines, scoring or recommendation engines, RAG apps, sales or intelligence tools, and wiki or process systems that agents operate against.

If the target has no reusable judgment, say so in one sentence and stop. Do not force the rubric.

Procedure

  1. Inventory. Read the user-supplied target and its declared direct dependencies. List owned data, deterministic code, and prompt-model work. Correct for declared off-repo data and reference documents that actually load into model context. Do not query connectors, credentials, or unrelated repositories to discover more data. If a declared dependency cannot be read, mark that bucket unknown/unverified, state the resulting confidence limit, and do not infer its contents. Identify any model decision that is really a fixed pipeline.
  2. Classify the load-bearing judgment. Load references/composition-model.md. When the target has fewer than eight load-bearing rules, name all of them; otherwise name the top eight to twelve. Record where each currently lives, where it belongs or whether it should be deleted, its consequence if wrong once, and who writes it.
  3. Estimate the shape. Anchor the estimate in that rule set, not a vibe. Give an approximate current split and state confidence. Treat it as a diagnostic aid, never a score.
  4. Name the biggest misallocation. Identify the single highest-leverage move.
  5. Check connected targets for duplicated policy. Inspect only the repos, workflows, or execution surfaces the target directly names, loads, generates, or depends on. Do not use connectors or credentials to discover targets. Name one duplicated chosen policy and its owner, or state the bounded scope and that none was found. Do not roam unrelated systems.
  6. Give a punch list. Order concrete moves by leverage and lowest risk first.
  7. Render the verdict. Use the template and field specification in references/output-template.md. When the verdict is available as a local text file, run python3 scripts/validate_composition_audit_report.py <report-file> and correct structural failures before delivery. The validator checks format only, never placement judgment.

This skill produces the verdict and punch list only. It does not execute the moves. Apply-mode requests route to the target project's own tooling and authority boundaries.

Reference Loading

  • references/composition-model.md — required for step 2. Owns the buckets, routing sort, counting traps, and good or bad shapes.
  • references/output-template.md — required for step 7. Owns the verdict template and field specification.
  • references/audit-worksheet.md — load when the target is large enough that the estimate would otherwise become a vibe.
  • references/worked-example.md — load when the target's shape is unfamiliar and you need to calibrate depth. Treat its dated project facts as illustrative, not current.
  • scripts/validate_composition_audit_report.py — run only against a saved verdict. Owns the mechanically checkable output contract.

Sources

Ported from Matt Weigand's earlier 60-30-10 skill. It operationalizes the concepts judgment-routing, prompt-technical-debt, storage-vs-enforcement, and context-as-moat. Matt applied the shorthand in his audits; the course's teaching framings and this skill's operational adaptation are distinguished in source framings. Read provenance only when checking source history.

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.