Knowledge Distillation Protocol for AI Agents
Still is a structured knowledge distillation framework designed for AI agents, built on the Cangjie (仓颉) deep distillation methodology. It transforms raw information — videos, articles, conversations, concepts — into structured knowledge units that agents can reason about directly.
Cangjie invented Chinese writing; Still distills knowledge.
| Traditional Notes | Still |
|---|---|
| Written for humans, grows stale and ignored | Written for AI agents, lightweight and queryable |
| Flat tag-based organization | Bidirectional knowledge graph |
| Pure summarization, depth depends on the writer | Structured validation, framework-driven depth |
| No boundary conditions | Explicit applicability matrix |
| Isolated notes | Auto-linked across entries |
Core insight: In the age of information overload, the bottleneck isn't acquisition — it's distillation. AI agents don't need more data; they need purer knowledge.
Every distilled entry has 8 structured modules + 1 optional, with typed relationships between entries:
0. 📡 Lifecycle → V1 (raw) → V2 (verified) → V3 (synthesized)
1. ⚡ One-Liner → 3-second relevance check
2. 🏗️ Core Framework → The skeleton
3. 📚 Theory Backing → Why it holds
4. ⚙️ Method E/B → Execution steps + Boundaries
5. 🤖 What AI Can't Do → Human-AI division
6. 🧭 Judgment Path → Novice → Expert roadmap
7. ✅ Triple Validation → Cross-domain / Predictive / Distinctive
8. 🔗 Linked Concepts → Typed relationship graph (8 relation types)
[9]. 🧩 Cross-Domain Grid → Mental model lattice (optional)
Full specification: SCHEMA.md (Chinese)
still/
├── README.md # This file
├── SCHEMA.md # Distillation protocol specification (Chinese)
├── AGENTS.md # Agent usage guide (Chinese)
├── templates/ # Distillation templates
│ ├── video.md # Video content
│ └── concept.md # Concepts & methodology
├── examples/ # Real-world distillation examples
│ └── ai-imagination-bottleneck.md
├── scripts/ # CLI tools
│ ├── distill.py # Distiller
│ ├── validate.py # Validator
│ └── link.py # Linker
└── workflows/ # Automation pipeline
└── distill.sh # One-shot distillation pipeline
Pick a template from templates/, follow the schema, fill in the 8 modules.
# Distill a raw article
python3 scripts/distill.py --input article.md --output distilled.md
# Validate a distilled entry
python3 scripts/validate.py examples/*.md
# Find orphan entries (no cross-links)
python3 scripts/link.py --target examples/my-distill.md --wiki ../wiki/# Batch process all raw files
bash workflows/distill.sh ./raw/ ./wiki/Place distilled .md files into your agent's knowledge base. Agents can:
- 3-second relevance check — read the "One-Liner"
- Cross-domain reasoning — traverse "Linked Concepts" for analogies
- Boundary awareness — check the "Boundary" matrix for applicability
- Output validation — use "Triple Validation" as a cross-check
- Progressive learning — follow the "Judgment Path" from novice to expert
See AGENTS.md (Chinese) for detailed agent workflow instructions.
| Principle | Meaning |
|---|---|
| Essence | Every entry retains only its irreducible core |
| Verifiable | Every claim has a validation anchor |
| Approachable | 3 seconds to decide relevance |
This framework pairs well with:
- Obsidian / any wiki — as the storage backend for distilled entries
- codebase-memory-mcp — for code knowledge graphs
- Graphify — for project-level relationship mapping
MIT