Addy Osmani: You're overprompting Claude — one goal plus a check beats prompt bloat
addyosmani · x · 2026-10-29
Google engineer Addy Osmani argues newer models like Opus 5.5 need far less prompting than most people still do.
Core workflow
- Give Claude a single prompt with one clear goal and how it should verify completion, then let it work; start at /effort medium and adjust.
- If you don't know what you want, have Claude interview you until it can write a spec, then start a fresh session from that spec.
- If the task yields pass/fail or a number to beat, reported in a readable way, Claude keeps going unattended.
- For long background tasks, /goal lets a separate model check a condition after every turn.
From the illustrations
- Over 80% of Claude Code's system prompt was removed with no measurable drop in coding evals.
- To migrate old instructions: remove some, see what it misses, add back only those — repeat per model; /doctor prompt-audit flags instructions written for older models.
- Projects let one long-running chat orchestrate parallel threads, reporting done/PR-ready/needs-you.
More from coding & agent
- OWASP releases 144-page GenAI Red Teaming Manual with 8-phase methodology for AI agents — AxSaucedo · 2026-10-29
- company-brain: open-source knowledge graph that refuses to answer without a citation — JeremyCMorgan · 2026-10-29
- MCP Servers v2.2.0 tightens cert checks; old fetch environments need a certifi refresh — caolonghua · 2026-10-29
- theo mocks devs prompting agents in the terminal, and observers guess it hints at his next product — cgarciae88 · 2026-10-29
- Robotics engineers hold a 'vibe engineering death match' for depth image compression — facontidavide · 2026-10-29
- Strix Halo-specific kernels speed up vLLM prefill by up to 31x at long context — asfbrz96 · 2026-10-29