shipshitdev/skill-comply

Skill Comply

Test whether an instruction artifact changes agent behavior in realistic conditions. The goal is not to prove that the file is well-written; it is to measure whether an agent follows it when the user prompt is supportive, neutral, or competing.

Contract

Inputs:

• Path to a SKILL.md, rule file, command, agent definition, or instruction doc
• Optional scenario list, trace logs, target harness, and pass threshold

Outputs:

• Expected behavior spec
• Scenario prompts by strictness level
• Compliance scores and failure categories
• Action timeline with matched, missing, reordered, or extra steps

Creates/Modifies:

• May create a local report file when requested
• Does not change the target skill or rule unless asked to fix findings

External Side Effects:

• May run local agent commands or tests when explicitly available
• No network calls or production writes by default

Confirmation Required:

• Before running costly model trials, modifying the target artifact, or writing reports outside the repo

Delegates To:

• skill-validator for frontmatter/spec checks
• skill-auditor for portfolio-level duplicate and quality audits
• evaluation for broader agent eval design

When to Use

• A new skill, rule, command, or agent definition was added.
• An instruction was rewritten and behavior needs regression testing.
• The agent keeps ignoring a known rule.
• A maintainer wants a pass/fail signal before publishing an instruction pack.
• A rule works only when the user explicitly repeats it, but not in normal tasks.

Core Idea

Compliance has three layers:

1. Spec extraction: convert the instruction artifact into observable steps.
2. Scenario pressure: test supportive, neutral, and competing prompts.
3. Timeline matching: compare observed actions to the expected sequence.

Measure behavior, not prose quality.

Process

1. Derive the Expected Behavior

Read the target artifact and extract:

• trigger conditions
• required actions
• forbidden actions
• ordering constraints
• confirmation gates
• expected output shape
• side effects that must or must not happen

Represent the spec as a short table:

| Step | Required Behavior | Evidence Type | Required Order | | --- | --- | --- | --- | | 1 | Inspect existing examples before editing | search/read action | before write | | 2 | Preserve unrelated user changes | diff/status check | before final |

2. Build Scenarios

Create three prompts:

• Supportive: explicitly asks for the workflow.
• Neutral: asks for the task but does not mention the instruction.
• Competing: adds pressure that often causes skipping, such as urgency, ambiguity, or a tempting shortcut.

For destructive or side-effecting skills, use dry-run scenarios against fixture files or synthetic repos. Do not create real external side effects for a compliance test.

3. Execute or Simulate

Use the strongest safe evidence available:

• actual tool/action trace from a sandbox run
• local command logs
• saved transcript
• reviewer reconstruction from a completed session
• dry-run plan when execution is not safe

Do not count a step as compliant because the final answer claims it happened. Require an action trace, command output, file diff, or visible artifact.

4. Grade the Timeline

Classify each expected behavior:

| Result | Meaning | | --- | --- | | Matched | Required behavior happened with adequate evidence | | Missing | Required behavior did not happen | | Reordered | Behavior happened, but too late to matter | | Weak | Behavior happened superficially or without enough evidence | | Violated | A forbidden action happened |

Compute a simple score:

score = matched / required
critical violation = automatic fail

Use judgment for ordering and gates. A late safety check after writing code does not satisfy a pre-edit investigation requirement.

Failure Categories

• Trigger failure: the skill or rule did not activate in neutral wording.
• Ordering failure: the right action happened after the risky action.
• Evidence failure: the agent claimed compliance without observable proof.
• Scope failure: the agent followed the rule for one file but not the full surface.
• Conflict failure: another instruction overrode the rule.
• Overhead failure: the rule caused excessive work for low-risk tasks.

Output Format

Compliance result: 2/3 scenarios passed. Neutral prompt failed because the agent edited before checking existing patterns.

| Scenario | Score | Verdict | Main Failure |
| --- | --- | --- | --- |
| Supportive | 5/5 | Pass | - |
| Neutral | 3/5 | Fail | Missing pre-edit example search |
| Competing | 4/5 | Pass | Weak final evidence |

## Timeline Findings

- Missing: [step] [evidence]
- Reordered: [step] [why order matters]

## Recommended Fix

1. Tighten the trigger phrase in `description`.
2. Move the precondition into the Contract.
3. Add a short output checklist.

Promotion Guidance

When a step repeatedly fails:

• Make the trigger clearer if the skill is not selected.
• Move mandatory behavior into the Contract if it is buried in prose.
• Split the skill if it has multiple unrelated workflows.
• Convert fragile mandatory behavior into a hook, script, or command only when the current platform can actually enforce it.
• Add examples only when the failure is ambiguity, not lack of enforcement.

Anti-Patterns

• Testing only prompts that name the skill.
• Treating final-answer claims as proof.
• Running live destructive actions to test a safety rule.
• Using a single scenario to certify a broad workflow.
• Fixing compliance by adding long repetitive wording instead of clarifying the trigger, contract, or gate.