Convergence
When an AI agent fails in production, four operators independently say the same thing: don't change the model, change the context. Codify tribal knowledge into greppable artifacts (wikis, .md, .claude/, shorthand jargon files); the team's quality is gated by how complete and curated this substrate is.
Operators
- Sherwin Wu (OpenAI) — When the agent isn't doing what you want, fix the context — not the model. The dominant failure mode is missing context; encode tribal knowledge as greppable files.
- Andrej Karpathy — An LLM should maintain a wiki, not re-derive knowledge per query. Replace stateless RAG with a persistent wiki the LLM owns and edits.
- Simon Willison — Hoard a personal repository of things that worked — coding agents will recombine them and Encode jargon shorthand once, save tokens forever. Hoard a personal repo of things that worked; encode jargon shorthand once to save tokens forever.
- Pawel Huryn — Treat `.claude/` as a deployable artifact with versioning and rollback. Treat
.claude/as a deployable artifact with versioning and rollback. - Cat Wu — When a new model lands, re-read the system prompt and remove crutches. The reciprocal: when the model improves, remove context that's now redundant.
Variation
- Sherwin from inside a model lab argues capability isn't the gate.
- Karpathy proposes the architectural shape (wiki vs RAG).
- Willison frames it personally (hoarding) and at the team level (shorthand).
- Huryn productises the artifact (treat .claude/ like code).
- Cat Wu adds the curation discipline (prune as models improve).
- Convergence: the contextDB is a first-class deliverable, not a config file.
Implication
Your AI quality budget should fund context curation more than it funds model upgrades. Maintain a .claude/ or equivalent as a deployable artifact: versioned, reviewed, pruned. When agents fail, the first investigation is the context, not the prompt or model. Treat the wiki as the unit of compounding leverage.
Sources
- ins_bottleneck-is-context-not-capability — Sherwin Wu
- ins_llm-wiki-pattern — Andrej Karpathy
- ins_personal-pattern-hoarding — Simon Willison
- ins_red-green-tdd-shorthand-for-agents — Simon Willison
- ins_dotclaude-as-deployable-artifact — Pawel Huryn
- ins_remove-features-as-models-improve — Cat Wu