Pilot Runbook

Pilot runbook

Purpose

Use this runbook to evaluate I-9 Skills against an existing collection without treating a pilot result as a release, publication, or automatic migration. The pilot proves whether the framework can identify a bounded improvement, make it reviewable, and preserve evidence for the next decision.

Scope

Choose one existing repository and one low-risk skill or small, coherent group of skills. Work only in a dedicated branch or disposable workspace. The pilot may inspect, propose, and validate changes; it must not publish, install packages globally, contact external services, or merge changes unless separately authorized.

Entry prompt

Give this prompt to an agent with the I-9 Skills collection available:

Use skill-routing to plan a read-only assessment of this repository's skills. Start with skills-audit, then return the smallest evidence-based route for one low-risk, high-value improvement. Do not edit files, install packages, publish anything, or access external services. State none if no focused improvement is justified.

Procedure

1. Route the request. Use skill-routing with the pilot goal, repository constraints, and whether independent work is permitted. Preserve its routing decision.
2. Audit the collection. Use skills-audit in read-only mode. Record package-contract gaps, mixed responsibilities, catalog drift, security concerns, and coverage limits.
3. Select one target. Prefer a clear, low-risk improvement with a concrete acceptance condition. Stop if the audit supports none.
4. Collect evidence. Use skill-evidence-collection to normalize the observations, origin, scope, and strength of the evidence. Keep weak signals marked as hypotheses.
5. Choose the smallest change path. Use skill-evolution for a bounded evidence-backed improvement. Use skills-refactoring only when responsibility boundaries need redesign. Use skill-design and skill-authoring when a new or rebuilt package is justified.
6. Prove the candidate. Run skill-evaluator and skill-security-review. Validate the package with the official Agent Skills validator and the repository's local checks.
7. Review and decide. Produce a concise pilot report that links the original route, audit, evidence packet, candidate diff, validation results, rejected alternatives, and recommendation: accept, revise, defer, or reject.

Acceptance criteria

The pilot is useful when it produces all of the following:

• one recorded routing decision, including a defensible none result when appropriate;
• one evidence-linked audit with stated coverage limits;
• at most one focused candidate change or a documented decision not to change anything;
• separate behavioral, security, and structural validation results;
• no secrets, personal records, customer data, credentials, or private source snapshots in the generated artifacts;
• a human-reviewable recommendation that does not silently publish, install, merge, or expand scope.

Interpreting the result

A successful pilot does not prove every host, model, or collection will behave identically. It demonstrates one bounded workflow under stated conditions. Capture failures as evidence: an unclear route, missing metadata, inflated responsibility, weak validation, or host incompatibility is useful input for the next focused improvement.

Related material

• Skill entry-path map
• Pilot evaluation
• Validation
• Security policy