This skill routes — it does not reason. Read the user's intent, match it to an entry point below, then execute that skill's SKILL.md.
Step 1 — Match the User's Intent
Read what the user wants to do and match it to the closest entry below. If ambiguous, ask one clarifying question.
| User wants to... | Start with | Then |
|---|---|---|
| Find what context is missing from a codebase | context-gap-analyzer |
→ agent-instruction-forge if gaps need rules |
| Create or improve agent instruction files (CLAUDE.md, .cursorrules, etc.) | agent-instruction-forge |
→ rule-quality-evaluator → edd |
| Score or audit existing agent instructions | rule-quality-evaluator |
→ agent-instruction-forge if score is low |
| Measure whether agent context actually helps | context-eval |
→ agent-instruction-forge if regression found |
| Iterate on a context harness with tests | edd |
→ context-eval for measurement |
| Design what goes into a context window | context-cartography |
→ context-gap-analyzer to validate coverage |
| Debug why an agent is failing / ignoring instructions | context-debugging |
→ context-gap-analyzer or edd based on findings |
| Extract business logic or domain rules from code | business-logic-extractor |
→ llms-txt-generator or agent-instruction-forge |
| Process a large document for LLM consumption | deep-document-processor |
→ llms-txt-generator |
| Generate an llms.txt or LLM-friendly reference | llms-txt-generator |
→ context-compressor if over budget |
| Optimize / compress context to fit a token budget | context-compressor |
→ context-eval to verify compressed context works |
| Find false positives in AI-generated tests | test-challenger |
→ edd if better assertions needed |
Step 2 — Execute
- Read
skills/[skill-name]/SKILL.md - Apply that skill's full methodology
- On completion, check the "Then" column for follow-ups
Step 3 — Propose Next Steps
Do NOT auto-execute the "Then" skill. Propose it:
Based on [what the skill produced], a natural next step would be:
→ [skill-name]: [1-sentence reason]
Want me to continue with that, or is this what you needed?
Multiple follow-ups → list as options. User chooses; orchestrator never chains automatically.
Canonical Chains
These are the most common multi-skill sequences in this group:
Full context engineering lifecycle:
context-gap-analyzer → agent-instruction-forge → rule-quality-evaluator → context-eval → edd
Use when building agent context from scratch or doing a comprehensive audit.
Creating agent instructions:
context-gap-analyzer → agent-instruction-forge → rule-quality-evaluator → edd
Use when the goal is specifically to create or improve instruction files.
Debugging agent failures:
context-debugging → context-gap-analyzer → agent-instruction-forge → edd
Use when an agent is behaving incorrectly and you suspect the context layer.
Building documentation:
business-logic-extractor → llms-txt-generator
deep-document-processor → llms-txt-generator
Use when creating LLM-consumable reference material.
Skill Registry
| Skill | Purpose |
|---|---|
context-gap-analyzer |
Find implicit context missing from a codebase |
agent-instruction-forge |
Create instruction rules for coding agents |
rule-quality-evaluator |
Score rules on Seven Properties, detect redundancies |
context-cartography |
Design what goes into an agent's context window |
context-debugging |
Diagnose agent failures originating in the context layer |
context-eval |
Measure whether context changes improve outcomes |
edd |
Eval-Driven Development — TDD for context |
llms-txt-generator |
Generate token-efficient context documents |
deep-document-processor |
Multi-pass reading of large documents |
business-logic-extractor |
Extract domain rules from code |
context-compressor |
Maximize signal-per-token under a finite budget |
test-challenger |
Find false positives in AI-generated tests |