Tell agents the why, not just the how: prompt with goals, not specs
rseroter · x · 2026-09-16
Sean Goedecke's post "Tell agents the why, not just the how" argues: early agents were enthusiastic idiots needing precise instructions; today's frontier models go wrong not from confusion but from incorrect assumptions about your goals and priorities. E.g. GPT-6 Astra, when it thinks code is for itself, produces minified code — it can write readable code, but you must tell it humans will read it.
His key advice: give the agent context on your priorities, not just the specific task. He shares a real starting prompt (project Deckard): stating a long-term goal (a local tool/browser extension to detect and hide AI content) and short-term goals (find an efficient detection model for his MacBook — maybe Pangram's EditLens 3B or a smaller Roberta, possibly quantized), and letting the agent work via Runpod through MCP.
More from coding & agent
- HF Jobs OCR script zoo: 28 models, one command, 7 pages for 7 cents — vanstriendaniel · 2026-09-16
- Creator's 3-step system turns 15 manual podcast steps into 8 AI skills, saving 5 hours a week — petergyang · 2026-09-16
- Friends vibe code a personal chat app with retro Windows-style UI, planning a shareable release — floguo · 2026-09-16
- Cumora: open-source team chat where AI agent teams collaborate with humans — tom_doerr · 2026-09-16
- EDSL to add activation capture, probes and model steering for LLM social simulation research — soumitrashukla9 · 2026-09-16
- NoSpoon agent autonomously generates hundreds of AI microdramas daily — Kyrannio · 2026-09-16