When Skill Evolution Means Removing Instructions

Description and use

Evolving agent skills by trimming instructions and employing a deterministic workflow that separates experience, knowledge, and executable skills, improving robustness across model changes.

Problem

How to evolve agent skills without overloading them with instructions and ensuring they remain effective across model changes.

Outcome

The author’s smaller implement-execplan skill achieved similar quality with ~18% fewer tokens, and the implement-approved-plan skill demonstrated a deterministic, multi-role workflow that reduced reliance on flagship models.

Workflow

• 1. Experiment manually with the model to understand the problem.
• 2. Design a workflow that works reliably without a skill.
• 3. Capture deterministic knowledge into mechanisms (e.g., state ledger, role-based chats).
• 4. Consolidate only the necessary instructions into a skill.
• 5. Validate the skill against quality gates and real pull requests.
• 6. Iterate: remove redundant instructions, adapt to model changes, and keep uncertain experience outside the skill until proven useful.

Evidence

In the first valid comparison, both versions passed the same quality gates, while the smaller version used about 18% fewer candidate input + output tokens.

Source

The skill itself coordinates a multi-chat workflow inside ChatGPT Desktop.

Source