Stop Burning Your Context Cash

A giant AGENTS.md feels responsible. It is often just burning tokens. Kent walks the Claude Code 80% cut, a research paper that found +20% cost for little gain, a live prune of his own file, and homework: make the always-on file a router, not an encyclopedia. (00:00) - Cold open: you're burning context dollars (00:43) - Better with Kent intro (00:59) - Claude Code cut 80% of the system prompt (01:43) - When system prompt, skills, and user requests clash (03:05) - Context engineering best practices that became myths (04:12) - Avoid stating the obvious / progressive disclosure (04:42) - You don't need useMemo for your agent (05:36) - Performance optimizations are not free (06:17) - Research: AGENTS.md often doesn't help (+20% cost) (07:06) - The cargo-cult loop (07:24) - Overstuffed AGENTS.md walkthroughs (09:01) - If the agent can find it in the repo, cut it (09:54) - Live prune of my own AGENTS.md (14:05) - I audited 37 agents (15:30) - AGENTS.md should be a router (16:30) - Shrink always-on context (18:30) - Homework: slim the always-on file (19:54) - Close: get better together Better with Kent - durable skills for people who ship software. Stuffing AGENTS.md, CLAUDE.md, and auto-loaded skills because it feels responsible is the new useMemo-everywhere cargo cult. Context is not free. Claude Code cut about 80% of its system prompt for newer models. A research paper evaluating AGENTS.md found those files often do not improve task success while increasing inference cost by over 20%. Kent walks real overstuffed files (discoverable repo trees, package managers, test commands the agent can find itself), then prunes his own Kody AGENTS.md live: keep what the repo is for, keep landmines, keep a short index, cut the rest. He audited 37 of his own agents from one day: 0 re-read AGENTS.md, 86% pulled docs and skills on demand, and most went straight to the leaf. The durable skill: treat the always-on file as a tiny map. Progressively disclose the rest. Measure whether the extra tokens still help after the models change underneath you. Homework: take 15-30 minutes with your agent. Rip out everything discoverable from the repo. Keep the gotchas that trip it every time. Index the rest. Delete docs nobody has opened. Links When to useMemo and useCallback The new rules of context engineering (Claude blog) Evaluating AGENTS.md (ETH Zurich / arXiv) Episode slides Better with Kent