Compound Knowledge

Compound Knowledge Skill

Capture a solved problem as searchable documentation so the team never solves the same problem twice.

Only ONE file gets written — the final documentation. All sub-agents return text only.

Categories

Solutions are organized into these categories (used as subdirectories under docs/solutions/):

• build-errors
• test-failures
• runtime-errors
• performance-issues
• database-issues
• security-issues
• ui-bugs
• integration-issues
• logic-errors

Filename Convention

docs/solutions/<category>/YYYY-MM-DD-<kebab-description>.md

Example: docs/solutions/runtime-errors/2026-02-13-null-pointer-in-auth-middleware.md

Solution Template

Use the template at solution-template.md for the final file format.

Orchestration: 3 Phases

Phase 1: Parallel Information Gathering (3 sub-agents via Task tool)

Spawn these three sub-agents in parallel using the Task tool. Each receives the conversation context about the solved problem and returns structured text. None of these sub-agents write files.

1a. Context Analyzer

Prompt: Analyze the conversation and codebase context to extract:

• Problem type: Which category from the list above best fits
• Affected area: Files, modules, or systems involved
• Severity: low / medium / high / critical
• How detected: How the problem surfaced (test failure, user report, monitoring alert, etc.)
• Impact: What was broken or degraded

Return as structured text.

1b. Solution Extractor

Prompt: Analyze the conversation and code changes to extract:

• Root cause: Why the problem happened (1-2 paragraphs)
• Code references: Specific files and lines involved
• What fixed it: The actual changes made (before/after snippets)
• Key insight: The non-obvious realization that led to the fix

Return as structured text.

1c. Prevention Strategist

Prompt: Based on the problem and solution, determine:

• How to prevent recurrence: Specific practices, checks, or guards
• Tests to add: What test cases would catch this in the future
• Monitoring: Any alerts or checks that should exist
• Related patterns: Similar problems that might exist elsewhere

Return as structured text.

Phase 2: Assembly (sequential)

The orchestrator (you) assembles the outputs from all three sub-agents into a single solution file:

1. Read the solution-template.md for the expected format
2. Determine the category from the Context Analyzer's output
3. Generate the filename: YYYY-MM-DD-<kebab-description>.md using today's date
4. Create the directory docs/solutions/<category>/ if it doesn't exist
5. Write the single assembled file at docs/solutions/<category>/<filename>.md
6. Fill in all YAML frontmatter fields and all sections from the sub-agent outputs

This is the only file write in the entire skill.

Phase 3: Optional Review

After writing the file, offer the user a specialized review based on the category:

• security-issues: "Want me to check if this security fix needs broader application?"
• performance-issues: "Want me to verify the performance improvement with benchmarks?"
• build-errors / test-failures: "Want me to run the test suite to confirm the fix holds?"
• Other categories: "Want me to search for similar patterns elsewhere in the codebase?"

Only proceed with the review if the user confirms.

Rules

• Sub-agents return TEXT only — they do not write files
• Only ONE file gets written — the final documentation in docs/solutions/<category>/
• Always use the solution template format
• Always confirm the category with the user if ambiguous
• Tag with relevant keywords for future searchability