Using Agent Skills

Using Agent Skills

Overview

Select the minimum set of relevant workflow skills for the current task, load them, and adapt selection if scope changes.

When to Use

• Start of every non-trivial task
• When requirements/scope change mid-task
• When output quality degrades and process guidance is needed

When NOT to Use

• Skills are already loaded and scope has not changed

Inputs

• Current persona
• Task instructions/context
• Optional requested skill list in task text/context

Requested Skill List Format

If the coordinator/user specifies skills, use this format in instructions/context:

{
  "requestedSkills": [
    "planning-and-task-breakdown",
    "security-and-hardening",
    "test-driven-development"
  ]
}

You may also accept a plain-text list (e.g., requestedSkills: security-and-hardening, test-driven-development).

Skill Selection Map

Use available local skills from mounted skill directories:

• Planning/architecture: planning-and-task-breakdown, risk-and-assumptions
• Implementation: incremental-implementation, test-driven-development
• Security-sensitive work: security-and-hardening
• Failure diagnosis: debugging-and-error-recovery
• Review: code-review-and-quality
• Manual validation: test-plan-and-execution
• Integration: merge-and-validate
• Final handoff: summarize (always)

Process

1. Parse task intent and detect any requested skill list.
2. Resolve requested skills against locally available skills.
3. Select minimal required skill set for task risk/profile.
4. Load each selected skill using read on its SKILL.md.
5. Execute task using loaded skill workflows.
6. Re-evaluate and load additional skills if scope changes.

Verification Checklist

• Selected skills match task intent and risk
• Requested skills were loaded when available
• Missing requested skills were explicitly noted
• summarize is used for final output

Failure Modes to Avoid

• Loading all skills by default (context bloat)
• Skipping testing/security skills for risky changes
• Ignoring requested skill lists from instructions/context
• Failing to re-evaluate skill selection after scope drift

Output Requirements

• In final summary text, list which workflow skills were applied.
• Use summarize to produce /output/result.json.