Clarity Framework
An agentic intelligence framework for technical engagements. Gives any engineer full project context on day one and grows smarter throughout the engagement.
Built on Karpathy's LLM Wiki Pattern
Andrej Karpathy proposed that the best way to use LLMs is to build a personal wiki — structured knowledge that the model reads before answering. Clarity extends this pattern with structured operational data (expertise.yaml), behavioral memory (Claude's auto-memory), and a self-learn loop that validates observations against live state.
Three Knowledge Systems
Operational Data
Project state, API gotchas, results, configuration. Structured YAML updated by /se:* commands. The agent's persistent brain — capped at 1000 lines.
Behavioral Rules
User preferences, guardrails, process rules. Markdown with frontmatter, auto-managed by Claude. Persists across conversations.
Durable Knowledge
Synthesized patterns, decisions, people, concepts. Obsidian-compatible markdown with [[wiki links]]. Updated by /se:wiki-* commands.
9 SE Commands
| Command | What It Does |
|---|---|
| /se:create | Create a new client with config files |
| /se:discover | Phase 0 auto-generation, seeds expertise.yaml |
| /se:brief | Standup/handoff summary from expertise.yaml |
| /se:self-improve | Validate observations, integrate confirmed facts |
| /se:check | Design guidelines compliance check |
| /se:wiki-ingest | Process raw/ files into wiki pages |
| /se:wiki-file | File a conversation insight as wiki page |
| /se:wiki-lint | Health check: orphans, broken links, stale pages |
| /se:meeting | Ingest meeting notes into notes.md and expertise.yaml |
The Self-Learn Loop
Observe
Every /se:* command appends raw observations to unvalidated_observations
Validate
/se:self-improve checks each observation against current live state
Promote
Confirmed facts get promoted into the relevant expertise section
Compound
Knowledge grows and compresses. The system gets smarter with every session
Want us to set it up for you?
We implement Clarity for teams — custom client templates, wiki seeding, command tuning, and training.
Get in Touch