Self-Improve Prompt Design

Guide for writing self-improve prompts that maintain expertise file accuracy.

Core Purpose

Self-improve prompts teach agents HOW to learn. They:

• Validate expertise against actual codebase (source of truth)

• Identify drift between mental model and reality

• Update expertise files automatically

• Enforce line limits to prevent context bloat

• Run after every ACT step in Act-Learn-Reuse

When to Use

• Creating maintenance workflows for agent experts

• Designing validation logic for expertise files

• Implementing the LEARN step of Act-Learn-Reuse

• Writing self-improve prompts for new domains

• Reviewing or debugging existing self-improve prompts

• Setting up git diff conditional checks for efficiency

Self-Improve Prompt Template

description: Sync {domain} expertise with codebase reality argument-hint: [check-git-diff]

Self-Improve: {Domain} Expert

Maintain expertise accuracy by validating against codebase implementation.

Arguments

• $1: check_git_diff flag (optional, default: "true")
  • "true": Only check files changed in git diff
  • "false": Full rescan of all domain files

Configuration

• EXPERTISE_FILE: .claude/commands/experts/{domain}/expertise.yaml
• MAX_LINES: 1000
• DOMAIN_PATHS: ["path/to/domain/files/"]

Workflow

Step 1: Check Git Diff (Conditional)

If $1 is "true" or not provided:

1. Run git diff --name-only HEAD~1 to get changed files
2. Filter to files in DOMAIN_PATHS
3. If no relevant changes, report "No domain changes detected" and exit
4. Continue with changed files only

If $1 is "false":

• Skip git diff check
• Process all files in DOMAIN_PATHS

Step 2: Read Current Expertise

1. Load EXPERTISE_FILE
2. Parse YAML structure
3. Note current line count
4. Identify sections present

Step 3: Validate Against Codebase

For each section in expertise:

core_implementation:

• Verify all file paths exist
• Check line counts are approximately correct
• Confirm key exports still exist

key_operations:

• Verify function names exist
• Check signatures match
• Validate file locations

schema_structure: (if applicable)

• Compare against actual schema definitions
• Check field names and types

best_practices:

• Ensure still relevant to current patterns

known_issues:

• Check if any resolved
• Look for new issues in domain

Step 4: Identify Discrepancies

Create a list of findings:

Step 5: Update Expertise File

Apply changes:

1. Update outdated information
2. Add new discoveries
3. Remove stale entries
4. Maintain YAML structure

Step 6: Enforce Line Limit

If line count > MAX_LINES:

1. Identify lowest priority sections
2. Summarize verbose entries
3. Remove least critical items
4. Continue until under limit

Priority order (highest to lowest):

1. core_implementation
2. key_operations
3. best_practices
4. known_issues
5. patterns_and_conventions
6. testing_notes

Step 7: Validation Check

1. Parse updated YAML (ensure valid syntax)
2. Verify line count: {current}/{MAX_LINES}
3. Run quick sanity checks on paths/names

Output Report

## Self-Improve Complete: {domain}

### Summary
- Files scanned: X
- Changes detected: Y
- Updates applied: Z

### Changes Made
| Section | Change | Reason |
| --- | --- | --- |
| ... | ... | ... |

### Expertise Health
| Metric | Value |
| --- | --- |
| Line count | X/1000 |
| Files validated | X/X exist |
| Functions verified | X/X accurate |
| Schema accuracy | X% |

### Recommendations
- [Any manual review suggestions]

Notes

• Mental model is NOT source of truth - codebase is
• Run after every ACT (build, fix, modify)
• Check git diff by default to save time
• Full rescan periodically or when drift suspected

Key Design Principles

1. Conditional Git Diff Check

Always include git diff optimization:

If check_git_diff is true:

• Only process changed files
• Exit early if no relevant changes
• Saves time on routine syncs

2. Line Limit Enforcement

Must be explicit and prioritized:

MAX_LINES: 1000

If over limit:

1. Summarize verbose sections

2. Remove lowest priority items

3. Never exceed limit

4. Validation Before Update

Always validate before writing:

Before writing updated expertise:

1. Parse as YAML (catch syntax errors)

2. Count lines

3. Verify critical paths exist

4. Actionable Output

Report should enable follow-up:

Changes Made

Recommendations

• Items requiring human review
• Potential issues to investigate

Validation Patterns

File Path Validation

For each file path in expertise:

1. Check if file exists at path
2. If not, search for file by name
3. If found elsewhere, update path
4. If not found, mark for removal

Function Validation

For each function in key_operations:

1. Search file for function definition
2. Compare signature if found
3. Update or flag discrepancy

Schema Validation

For each table/entity in schema_structure:

1. Find schema definition file
2. Compare fields and types
3. Note additions/removals

Anti-Patterns

Anti-Pattern Problem Solution

No git diff option Wastes time on no-ops Always include conditional check

No line limit Context overflow Enforce MAX_LINES strictly

Silent failures Drift undetected Report all validation results

Manual edits Human time wasted Self-improve only updates

No validation Bad YAML written Always parse before save

Integration with Act-Learn-Reuse

┌─────────────────────────────────────────────┐ │ ACT: Build, Fix, or Answer │ │ (Agent performs useful work) │ └──────────────────┬──────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────┐ │ LEARN: Self-Improve Prompt │ │ • Check git diff │ │ • Validate expertise against code │ │ • Update mental model │ │ • Enforce line limits │ └──────────────────┬──────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────┐ │ REUSE: Next Execution │ │ (Agent reads updated expertise first) │ └─────────────────────────────────────────────┘

Testing Self-Improve Prompts

To validate a self-improve prompt works correctly:

• Initial state: Run against seeded expertise

• Introduce change: Modify a file in domain

• Run self-improve: Check it detects change

• Verify update: Confirm expertise updated correctly

• Line limit test: Add content until limit hit, verify truncation

Related Skills

• expertise-file-design : Structure of expertise files

• agent-expert-creation : Full agent expert workflow

• meta-agentic-creation : Self-improving at scale

Last Updated: 2025-12-15

Version History

• v1.0.0 (2025-12-26): Initial release