All skills
simota avatar

/chisel

@965f4f9
by shingo imotasimota/agent-skills85 stars
15

Converting a supplied prompt into an executable specification: detects vague quality/quantity/explanation/style/design/technical/judgment wording, role and persona theater, and self-contradiction, then replaces each with a numeric bound, an observable behavior, or a scorable criterion — with a per-term ledger of what changed and what stayed open. Don't use for AI system design, RAG, or eval harnesses (Oracle), PRD/SRS authoring (Scribe), spec conformance verification (Attest), or SKILL.md normalization (Gauge).

Use this Skill: https://skilld.dev/gh/simota/agent-skills/chisel

This session only. Nothing lands on disk.

referencerole-decomposition.md

≈1.7k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Role Decomposition

Purpose: Dissolving titles, personas, and seniority claims into the capabilities and evaluation axes they were standing in for. Read when: A detection is classified DECOMPOSE or DELETE, or the role Recipe is active.

Contents

  • Why titles do not work
  • The six-slot decomposition
  • Standard decompositions
  • Decorative theater
  • What decomposition must not lose
  • Failure modes

Why Titles Do Not Work

A title asks the executor to infer a behavior from a label. The inference is unreliable in both directions: two readers derive different evaluation axes from "senior engineer", and a model given a title does not gain knowledge it lacked. What the title does carry — which properties get checked, which trade-offs get named, which mistakes get avoided — is exactly what can be written down instead.

This is the same rule the ecosystem applies to its own agent prompts: focus a spawn on the concrete task and output format, never on personality adoption (nexus/SKILL.md § Operational Notes for Spawns). Chisel applies it to prompts the user brings in.

Two additional constraints:

  • Never write credentials as fact. "You are a licensed attorney", "you have 20 years of experience" — these assert something untrue about the executor and can mislead a reader about the output's standing.
  • Never keep a bare title after decomposition. "As a professional, …" surviving the rewrite means the decomposition did not happen.

The Six-Slot Decomposition

Every DECOMPOSE detection fills these slots. Empty slots are omitted, not padded.

Slot Question it answers
Domain What body of practice does the work draw on?
Evaluation axes Which properties of the output get checked?
Method What thinking procedure produces the answer?
Judgment rules How are trade-offs decided when axes conflict?
Responsibility What is the deliverable accountable for?
Prohibited actions What does this role's failure mode look like?

The evaluation-axes slot is the load-bearing one. A decomposition that fills only Domain has removed a title and added nothing.

Worked example:

"You are a world-class software engineer."
→
- Evaluate from the perspective of someone who has designed, built, and operated
  large systems.
- Check not only correctness but readability, maintainability, extensibility,
  fault tolerance, security, and performance.
- Consider system-wide impact rather than local optimization.
- Avoid unnecessary abstraction and over-engineering.
- Propose an alternative alongside every problem raised.
- State the trade-offs explicitly.

Standard Decompositions

Reusable starting points. Trim to what the source prompt's task actually needs — importing all of a role's axes into a narrow task is over-specification.

World-class expert

  • Account for the field's main theories, established practice, and known failure patterns.
  • Do not stop at introductory explanation; cover the exceptions that matter.
  • Compare multiple options and give each one's advantages and disadvantages.
  • Avoid assertions whose support is weak.
  • The word "world-class" itself carries nothing and is dropped.

Strategy consultant

Structure the problem first · separate ends from means · analyze the highest-weight issues first · state hypotheses explicitly · compare options · assess feasibility and impact · close with a recommendation.

Senior software engineer

Correctness · readability · maintainability · testability · extensibility · performance · security · failure behavior · edge cases · operational cost. Solve with the minimum necessary complexity; avoid over-engineering.

Product manager

User problem · business value · usage frequency · implementation cost · risk · priority · success metric · the MVP boundary. Judge by user value, not feature count.

UX designer

User goal · cognitive load · information priority · number of operations · consistency · feedback · error prevention · accessibility. Do not evaluate on visual appeal alone.

Researcher

Hypothesis · evidence · falsifiability · correlation vs causation · data quality · sample bias · alternative explanations · uncertainty. Keep fact, hypothesis, and speculation distinguished.

Critical reviewer

False premises · logical leaps · contradictions · unconsidered cases · bias · risk · counterexamples. Do not stop at criticism — give the remedy.

Teacher

Assume the learner's prior knowledge explicitly · order from foundation to application · connect new concepts to known ones · use concrete examples · name the common misunderstandings · surface the points that confirm understanding.

Copywriter

Name the reader · write from the reader's problem and desire · make the benefit concrete · do not lean on abstract adjectives · prefer expressions understood on first read · place a clear call to action where one is warranted.


Decorative Theater

DELETE fires on lines whose removal changes nothing about the deliverable.

Pattern Disposition
"You are a genius", "IQ 200", "the world's best", "the greatest expert in history" DELETE — no derivable behavior
"You never make mistakes", "you are always right" DELETE — additionally suppresses the uncertainty signals the reader needs
"Take a deep breath", "think step by step" as a magic incantation DELETE — on current models, reasoning depth is a parameter, not a phrase
"You are a licensed X", "you have N years of experience" DELETE — a false claim about the executor

Before deleting, check whether the line implied evaluation axes. "The world's best security engineer" implies a threat-model-first check; delete the title and write the check.


What Decomposition Must Not Lose

Deleting a persona is safe. Deleting what the reader inferred from it is a behavior change disguised as a cleanup.

Before removing any role line, ask: which properties of the output would a reader have expected to be checked because of this line? Those become explicit rules. If the honest answer is "none", the line was theater and deletes cleanly.


Failure Modes

Failure Looks like Fix
Retitling "world-class engineer" → "highly skilled engineer" The output must be capabilities, not a better adjective
Silent narrowing Title deleted, axes never re-expressed Recover the implied axes first
Axis import A tone-focused prompt receives all ten senior-engineer axes Trim to the axes the task actually exercises
Credential assertion "You are a certified auditor" preserved for authority Never — assert capability, never credentials
Slot padding Every slot filled for a one-line task Empty slots are omitted

Source: SKILL.md on GitHub

No alerts13d3 checks · Risk SAFE
  • Gen Agent Trust Hub13d

    The skill is a prompt engineering assistant designed to detect and resolve ambiguity in user instructions. It transforms vague language into explicit, executable specifications through a structured workflow. The analysis found no evidence of malicious patterns, data exfiltration, or unauthorized command execution. The skill correctly identifies its own logical boundaries by routing technical requirements like security protocols and spend caps to appropriate enforcing layers rather than relying on prompt-layer hardening alone.

  • Socket13d

    No alerts

  • Snyk13d

    Risk: LOW · No issues

Signed by skilld at 965f4f9. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 2 days ago.

Activeupdated 2 weeks ago

README badge

README badge for simota/agent-skills/chisel