Skill: Copilot Self-Improvement
Activation Context
- Editing
copilot-instructions.mdor 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:
| File | Loading | Impact |
|---|---|---|
copilot-instructions.md |
Every request | Highest — competes with every interaction |
*.instructions.md |
When editing matching files | Medium — present during file-type work |
| Skills | On demand | Lowest — only when activated |
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:
| Signal | Destination |
|---|---|
File-extension-specific conventions (.tf, .py, .ts) |
*.instructions.md |
| Step-by-step procedures, "how to" sections, decision trees | Skills |
| URLs to external docs, version-specific info | copilot-references.md |
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
| File | When Loaded | Context Impact | Content |
|---|---|---|---|
copilot-instructions.md |
Every request | High (always present) | Minimal: overview, stack, principles, skill registry |
*.instructions.md |
File pattern match | Medium (during edits) | File-type conventions |
Skills (SKILL.md) |
On demand | Low (only when needed) | Procedures, workflows |
Agents (*.agent.md) |
Agent selection | Medium (per session) | Persona, handoffs, tool restrictions |
copilot-references.md |
On demand | Low | External links, stale-prone data |
.copilot-tasks.md |
Conversation start | Low | Cross-conversation state |
Placement Decision
| Question | Yes → |
|---|---|
| Always needed, fits in <20 lines? | copilot-instructions.md |
| Always needed, >20 lines? | Summarize in instructions, detail in skill |
| Triggered by file type? | *.instructions.md |
| Agent persona? | agents/*.agent.md |
| Procedure/workflow? | Skill |
| External reference data? | copilot-references.md |
| None of the above? | Probably doesn't need documenting |
Size Guidelines
| File | Guideline | Key Question |
|---|---|---|
copilot-instructions.md |
~200 lines | Could any content be conditionally loaded? |
| Skill (SKILL.md) | <500 lines, <5000 tokens | Split refs to references/? |
*.instructions.md |
Comprehensive | Is file type common enough to justify? |
| Agent | ~150 lines | Focused persona? Clear handoffs? |
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.
| Creating | Must Have |
|---|---|
| Skill | Clear activation context, registered in copilot-instructions.md |
| Agent | Distinct persona, clear boundaries, handoff definitions |
Anti-Patterns
| Anti-Pattern | Solution |
|---|---|
Everything in copilot-instructions.md |
Split per placement decision above |
| Duplicated content across files | Single source of truth |
| Procedures in always-loaded instructions | Move to skills |
Skill not registered in copilot-instructions.md |
Register it — unregistered = undiscoverable |
| Tool-enforceable rules in instructions | Remove instruction, verify tool config |
| Hardcoded external data | Use copilot-references.md |
Health Checks
On every invocation of this skill:
- Scan for content that could be conditionally loaded
- Scan for tool-delegable rules
- If found, assess and extract/remove before other work
| Periodic Check | Action |
|---|---|
| Stale references | Test 2-3 for relevance |
.copilot-tasks.md cleanup |
After branch merge, remove branch section |
Fitness Assessment
When reviewing skills (on request or during health checks):
| Criterion | Question |
|---|---|
| Purpose clarity | Immediately clear from name + description? |
| Audience fit | Written for Copilot, not humans? |
| Actionable | Specific instructions vs vague principles? |
| Domain value | Knowledge Copilot lacks from training? |
| Up-to-date | Reflects current project state? |
| Dependencies valid | Referenced files actually exist? |
Verdict: ✅ Fit / 🟡 Needs attention / ❌ Needs rewrite (fix structural issues NOW, not later)
Knowledge Gaps
| Trigger | Action |
|---|---|
| Uncertain about current syntax/API | Check copilot-references.md |
| User provides corrective link | Add to copilot-references.md with category and date |
| Repeated procedure (3x) | Create a skill |
| Instructions over size limit | Extract to skill or *.instructions.md |
Self-Registration
This skill must verify copilot-instructions.md contains:
- Trigger rule in Collaboration Guidelines: "Before editing Copilot config — read copilot-self-improvement skill"
- Skill entry in the skills table
If missing, add them before making other changes.
Source: SkaneTrails/meal-planner — distributed by TomeVault.