Skill Lifecycle

Skill Creator

Use this skill when the user wants to work on content skills themselves — creating, evaluating, improving, or benchmarking them.

Trigger

• Requests to create a new content skill
• Requests to improve or benchmark an existing skill
• Requests to reduce false positives or false negatives in skill triggering
• Requests to compare the current skill against a revised candidate

Workflow

1. Read references/workflow.md to choose the right Skill Lab sequence.
2. Read references/eval-guidelines.md before creating or editing trigger evals.
3. If the skill does not exist yet, scaffold a new skill package (e.g. frontagent skill scaffold <skill-name>).
4. If the skill does not yet have evals, initialize trigger evals (e.g. frontagent skill init-evals <skill-name>).
5. If behavior quality matters, initialize behavior evals (e.g. frontagent skill init-behavior-evals <skill-name>).
6. Run a benchmark before making changes (e.g. frontagent skill benchmark <skill-name>). Use --behavior when behavior evals are available.
7. When improvement is requested, generate a candidate and compare it with baseline (e.g. frontagent skill improve <skill-name>). Use --behavior to include behavior scoring.
8. Only apply a candidate when the benchmark clearly improves and the user wants promotion (e.g. frontagent skill promote <skill-name> <candidate-id> or --apply-if-better).

Platform-specific commands listed above use the frontagent skill CLI. See ADAPTATION.md for how to map these steps to a different platform.

Output Contract

• Keep the user informed of:
  • where eval files live
  • where candidate skills were written
  • whether benchmark scores improved
• Prefer benchmark-backed recommendations over intuition.
• Treat the skill package itself as the artifact under iteration:
  • SKILL.md
  • agents/openai.yaml
  • existing references/ and assets/

Guardrails

• Do not trust starter evals blindly. Encourage editing them toward real prompts before strong conclusions.
• Do not auto-apply candidates unless the user requested it or the command explicitly says to do so.
• Do not silently broaden a skill's scope just to improve trigger rate.
• Prefer preserving existing references/assets over inventing new file paths.