After auditing 6 repos, a developer deleted 71% of their agent instructions
JuniorCustard4931 · reddit · 2026-08-04
- The author audited agent instruction files across 6 repos and deleted 71% of them after realizing they had accumulated a year of stale, duplicated, and overly prescriptive rules.
- In the worst session, Claude Code had to read 25,112 words / 33k tokens before seeing any code; after pruning, that dropped to 8,245 words.
- The biggest wins came from removing stale bug workarounds, advice the model now handles on its own, duplicated rules, dead references, and reference data that belonged in docs rather than always-on instructions.
- Their rubric was simple: if the model could know it without being told, delete it; keep only environment facts, real workarounds that still reproduce, and other request-time essentials.
- Everything removed was archived verbatim in a non-auto-loaded file in the same commit, so any mistake could be restored line-by-line.
- The main lesson: stale instructions don’t just add noise — they can become confidently wrong, and that is often worse than having no instruction at all.
More from coding & agent
- Codex edited a full three-camera episode end to end in under 3 hours — danshipper · 2026-08-04
- CTO says an engineer automated 60% of his job with an AI agent and got promoted — sloppenheimer · 2026-08-04
- Google Cloud says GKE Agent Sandbox lifts agent density from 61 to 274 per node — rseroter · 2026-08-04
- A tool now exports React components for agent-driven integration — evilrabbit_ · 2026-08-04
- Excalidraw shares a short guide to setting up MCP in ChatGPT Plus — Vjeux · 2026-08-04
- Merge Agent Handler adds six connectors for broader business workflow access — shensi · 2026-08-04