copilot-self-improvement

name: copilot-self-improvement description: Meta-skill for maintaining and improving Copilot configuration files, instructions, and skills

Skill: Copilot Self-Improvement

Activation Context

• Editing copilot-instructions.md or any *.instructions.md
• Creating or modifying skills in .github/skills/ or agents in .github/agents/
• Knowledge gaps that external information resolves
• User asks about improving Copilot's effectiveness

⚠️ Repeatedly Overlooked — READ CAREFULLY

These rules have been violated multiple times. They exist because the default behavior is wrong.

1. You are the domain expert for these files

Do NOT ask "should I extract this?" or "what line limit do you prefer?" for Copilot config files. State "I'm doing X because Y." The user can override — but you decide by default. These files are FOR you, BY you. Human role is approval only.

2. Fix structural issues NOW, not later

When you detect a problem (size violation, misplaced content, tool-delegable rules): fix it immediately as the first action. Do not defer. Do not "note it for later." Deferring compounds the problem on every subsequent request.

3. Report ALL findings transparently

When scanning for misplaced content, report everything you considered — including things you assessed as not worth extracting. Silent dismissal means the user never knows you evaluated it.

Wrong: (silently dismisses python.instructions.md) Right: "Created terraform.instructions.md. Also detected Python patterns but assessed as overhead since Ruff handles it."

Context Window Awareness

Every line in copilot-instructions.md, *.instructions.md, and skills consumes context window space, leaving less room for code reasoning, file contents, and tool results. This directly causes earlier summarization and lost conversation state.

Loading frequency matters:

Optimization priority: Always-loaded > pattern-matched > on-demand. A 10-line reduction in copilot-instructions.md frees more reasoning capacity across a session than removing 50 lines from a skill.

When reviewing any config file, ask: "Does every line here earn its context window space?" Generic knowledge from training data does not. Project-specific decisions, gotchas, and patterns that contradict defaults do.

Detect Misplaced Content

On every invocation, scan copilot-instructions.md for content that belongs elsewhere:

Extract BEFORE proceeding with other work.

Delegate to Tools, Not Instructions

Any rule that a linter, formatter, or validator can enforce MUST NOT appear in instructions. Instructions duplicate what ruff, prettier, pre-commit hooks, or tsconfig already handle = wasted tokens.

Test: Could a regex, AST parser, or schema validator enforce this? → Tool territory.

When detected: check project's prek config (.pre-commit-config.yaml) / pyproject.toml / tsconfig.json. If tool exists, remove the instruction. If not, propose adding the tool.

Keep in instructions only: semantic patterns, domain conventions, architecture decisions, project-specific terminology — things tools cannot enforce.

File Architecture

Placement Decision

Size Guidelines

Size is a smell, not a violation. 400 lines of essential content is fine. 150 lines with extractable content is not.

Creating Skills and Agents

Read references/PROCEDURES.md for structure, registration, and anti-patterns.

Anti-Patterns

Health Checks

On every invocation of this skill:

1. Scan for content that could be conditionally loaded
2. Scan for tool-delegable rules
3. If found, assess and extract/remove before other work

Fitness Assessment

When reviewing skills (on request or during health checks):

Verdict: ✅ Fit / 🟡 Needs attention / ❌ Needs rewrite (fix structural issues NOW, not later)

Knowledge Gaps

Self-Registration

This skill must verify copilot-instructions.md contains:

1. Trigger rule in Collaboration Guidelines: "Before editing Copilot config — read copilot-self-improvement skill"
2. Skill entry in the skills table

If missing, add them before making other changes.