skillify

Part of the Overpowered skill suite. Generalize and package a qualified or otherwise proven repeatable workflow into a portable Agent Skill with concise instructions, progressive disclosure, trigger-focused metadata, gotchas, and evals. Use after real execution evidence exists—especially for a Skill Academy capability ready to graduate; do not use as the hot-generation mechanism during a task or merely to save a conversation.

Overpowered / Skillify

Positioning

This skill is for productizing a proven workflow, not acquiring a missing capability during the current task. gear-up creates/stages the smallest temporary capability; the Skill Academy collects reuse evidence; skillify turns a qualified method into a durable portable Agent Skill.

Harness-native self-learning may already capture local procedures. skillify focuses on portability, minimality, trigger quality, and testability.

Core rule

Extract the invariant method, remove accidental project context, and prove the skill adds value with realistic evals.

Evidence gate

Accept one of these starting states:

• a Skill Academy capability marked qualified under the project's Academy policy; or
• equivalent evidence that the workflow succeeded in real use and is plausibly reusable beyond one accidental task.

If the method has never worked, package it only as an explicit experimental draft. Never call it proven or graduated.

Procedure

1. Inspect the execution/Academy evidence and identify what actually created value.
2. Extract:
   • reusable sequence;
   • corrections that mattered;
   • non-obvious gotchas;
   • required capabilities/tools;
   • stopping condition;
   • output contract.
3. Remove project-specific names, paths, values, and assumptions unless they intentionally define a domain skill.
4. Define one coherent responsibility. Split only when parts can usefully stand alone.
5. Write a precise description that covers both capability and trigger conditions.
6. Keep SKILL.md concise; move deep details to one-level references/ files.
7. Add 2–3 realistic evals, including a boundary/should-not-overtrigger case and regression cases for observed failures.
8. Validate Agent Skills structure and naming.
9. Test with a clean context against a no-skill or prior-version baseline when the environment permits.
10. Refactor away instructions that do not measurably improve behavior.
11. Record deployment target/version only after the human/project graduation gate is satisfied.

Output package

text

skill-name/
├── SKILL.md
├── references/       # only when needed
└── evals/
    └── evals.json

Gotchas

• A conversation summary is not a skill.
• One ephemeral Gear Up success is not enough to claim generality.
• Generic advice the model already knows is context tax.
• Do not encode one successful answer; encode the reusable procedure.
• Do not copy runtime-specific tool names unless the skill intentionally targets that runtime.
• Do not claim the skill is “tested” unless behavioral evals actually ran.
• Do not automatically install the result globally; graduation/deployment is a separate authority decision.

Stop condition

Stop when the package is structurally valid, its scope is coherent, trigger metadata is precise, evals can distinguish the desired behavior from a baseline, and its graduation/deployment status is stated truthfully.