LLM Context Modular Docs
LLM Context Modular Docs
Stop dumping a massive llm.txt into every agent query. This skill teaches you to split project documentation into small, purpose-built context files that agents load selectively.
Token savings: 70-96% per query.
The Problem
Most AI-assisted projects have one giant documentation file. Every agent query loads the whole thing even when 2% is relevant.
- Context overload: agents miss details buried in noise
- High token costs: 500 KB per query adds up fast
- Slow responses: large context = longer processing
- Stale docs: monoliths are hard to maintain
The Solution: Memory + Reasoning Files
Replace one monolith with focused pairs of files -- one per component:
Memory files (-memory.txt): Schema, API surface, config options. Load for reference lookups.
Reasoning files (-reasoning.txt): Workflows, decision trees, examples. Load for how-to questions.
Real Results (Crawl4AI)
Instead of loading 500 KB for every query:
- "How to install?" loads 1 file, 2.5 K tokens
- "All config options?" loads 1 file, 8 K tokens
- "Extract with LLM?" loads 2 files, 12 K tokens
- "Deploy to production?" loads 3 files, 18 K tokens
Median token reduction: 96%
What the Skill Does
- Audit -- Is your current documentation overloaded?
- Design -- Identify 8-15 components; decide the memory/reasoning split
- Create -- Write each file with correct format, scope, and length
- Validate -- Test with real queries; ensure 1-3 files answer any question
- Automate -- Generate memory files from source schema on every release
- Integrate -- Wire into Hermes skill references/ or any agent framework
When to Use
- Project documentation over 50 KB
- Multiple distinct topics (install, config, API, deployment)
- Agents loading irrelevant sections frequently
- Token cost is a concern
- Multiple agents on different parts of the same project
When NOT to Use
- Small projects under 10 KB total docs
- Already well-chunked separate files
- No AI agent consumers
Setup
No external dependencies. Works with any project that has documentation.
Time investment: 2-4 hours initial setup, 15-30 minutes per release update.
FAQ
Q: Why .txt files instead of .md?
A: Forces machine-readable format. Signals LLM context not user docs.
Q: How many components?
A: 8-15 is ideal. If you need more than 3 files to answer a basic question, components are too fine-grained.
Q: How to keep memory files current?
A: Automate generation from source (TypeScript types, OpenAPI specs). Run on every release.
Q: Works with Hermes skills?
A: Yes -- apply to references/ files. SKILL.md stays lean (1-3 KB); references load on-demand.
