PRs from humans and agents are welcome. Before submitting, understand the bar:
Every line must either fill a verified LLM blind spot OR be something the agent needs to teach the human.
If a stock LLM already knows something and the human doesn't need to hear it, it doesn't belong here. But some content exists because the agent is a teacher — the human needs to learn about gas costs, security patterns, or why Ethereum matters, and the agent needs accurate material to teach from. Both are valid reasons for a line to exist.
Before adding or modifying content, run the triage process:
- Spawn a fresh LLM — no tools, no skills, no web access. Pure training data.
- Give it a realistic task that exercises the content you're proposing. Don't ask "do you know X?" — ask it to build something and examine what it produces.
- Classify each item in your proposed change:
- 🔴 LLM blind spot — consistently gets this wrong → keep
- 🟣 Human needs to learn this — the agent knows it, but needs to teach it accurately → keep
- 🟡 Knows but skips — knows the concept, won't do it unprompted → compress to one line
- 🟢 Does this naturally — any competent model does this already AND human doesn't need teaching → cut
Full methodology with worked examples: research/triage-methodology.md
- Don't trust intuition. "I think agents get this wrong" is not evidence. Run the test.
- Don't ask leading questions. The LLM will say "yes I know that." Make it demonstrate knowledge by building.
- Don't keep content because it's correct. Correct ≠ necessary. A lot of correct information is already in training data.
- Don't pad skills with training-data-tier knowledge. A 95-line skill that's all blind spots is more effective than a 205-line skill where half is noise.
- Ran a baseline test against a stock LLM (no tools/skills)
- Every item classified as 🔴, 🟡, or 🟢
- 🟢 items removed
- 🟡 items compressed
- Content verified against onchain reality (not LLM-generated "facts")