After a major model upgrade, prune your AGENTS.md: Astra follows stale rules too literally
keyanzhang · x · 2026-09-05
keyanzhang's advice for adapting to major model upgrades like Astra: strong instruction-following means old, overly prescriptive rules get executed too literally, constraining the model or making it do extra work.
His recommendations:
- Trim down old skills, AGENTS.md, and custom instructions
- Consider deleting everything and adding parts back only as needed
- Treat the new model like a principal engineer, not a junior: share high-level goals, define success clearly, and let it exercise judgment
The core idea: as models get stronger, your config should get leaner.
More from coding & agent
- shadcn's agent for watching restaurant reservations cheated and falsely reported success — itsOmSarraf_ · 2026-09-05
- Best-value AI coding subscriptions charted against AA Intelligence Index v4.2 — popiazaza · 2026-09-05
- Agentic memory retrieval: how to do variable top-k in low-latency tier — WonderfulArt9908 · 2026-09-05
- Dev laments Anthropic killing the ability to plug Claude OAuth into custom harnesses — BLUECOW009 · 2026-09-05
- xAI to host Grok Bot Galaxy, a three-day agent event in SF this September 15-17 — XFreeze · 2026-09-05
- Developer who found Codex 'slow and annoying' 1.5 years ago declares AGI achieved — gajesh · 2026-09-05