| name | craft |
| type | skill |
| category | meta |
| description | Instruction quality gate โ reviews agent instructions (task bodies, workflow steps, skill procedures, self-test protocols, agent definition files) for shallow-execution vulnerabilities before deployment. Two modes: author (pre-hoc review) and audit (trace a failure back to the instruction gap). The bar is excellence, not compliance. |
| triggers | ["craft","review these instructions","instruction quality","are these instructions good enough","raise the bar","why did the agent miss this","review this agent definition"] |
| modifies_files | true |
| needs_task | false |
| mode | conversational |
| domain | ["meta","framework","quality-assurance"] |
| allowed-tools | Read, Grep, Glob, Bash, Edit, Write, Agent |
| model | opus |
| version | 0.2.0 |
| permalink | skills-craft |
Instruction Craftsmanship
Review and audit agent-facing instructions โ task prompts, workflow steps, skill procedures, self-test protocols โ for excellence. Applies to any Claude agent system, not only this repo's framework.
First Principles
Good instructions trust a capable, improving agent to exercise judgment; they do not try to mechanically pre-solve every case.
- Trust the harness, not today's quirks. Agents and their tools improve continuously. Never write an instruction to patch a specific client's current limitation, plug a gap that will close on its own, or hard-code a workaround for how one version of an agent happens to behave. If a rule is only true "for now," it does not belong in a durable instruction.
- Specify the process, not the keystrokes. State when to invoke which capability and what outcome proves it worked. Do not spell out sub-steps, tool flags, or branching logic a competent agent already knows how to perform (opening a PR, formatting a table, running a routine lookup). Name the judgment call, not the click-path.
- One skill, one job. An instruction set is constrained to its own pure function. Naming another skill as a delegation or dispatch target is fine; explaining, restating, or summarizing that other skill's internals, procedures, or file layout is not โ that creates a hidden dependency that silently rots when the referenced skill changes shape.
- Verification must be real, not performed. "Did the step run?" is not evidence of anything. Instructions must demand direct inspection of the actual artifact โ outputs, logs, diffs โ with an eye for the failure that looks like success (silent errors, plausible-but-wrong data, a summary standing in for the thing itself).
- Every line earns its place. Brevity is a feature. Cut anything that does not change what the agent does: provenance ("on the 2026-06-25 sessionโฆ"), incident IDs, and recipes tuned to one past failure all belong in the PR/issue/memory that records why a rule exists โ not in the instruction loaded every run. Write the durable principle the incident illustrates, not the incident.
These are lenses, not a checklist to tick. If instructions feel shallow but match nothing below, trust the feeling and say why โ depth is verification specificity, not step count.
Common Defect Patterns
Instances of the principles above, worth naming because they recur:
- Compliance framing. "Did X run?" instead of "is the output correct, complete, and verified?" Require outcome-based checks, not process-completion checks.
- Evidence laundering. Accepting an agent's summary, a partial artifact-channel check (just stdout, not logs/exit-code/schema), or a green test suite as proof โ without inspecting the actual output for silent failures, corruption, or placeholders. (Principle 4.)
