Chapter
Unlearn the old playbook — so stale advice stops sabotaging you
Before frontier models, prompting was a craft of coaxing. Most moves that worked then backfire now. Worth knowing, because half the advice still circulating is from this era.
| The old move | Why it worked then | What to do now |
|---|---|---|
| Magic thinking words — "think", "think hard", "ultrathink" | They mapped to literal reasoning-token budgets (ultrathink ≈ 32k) | Adaptive thinking + effort settings; Opus 4.8 defaults to high. Use /effort; reserve ultrathink for one hard turn |
| Anti-laziness pep talks — "Don't hold back. Give it your all. Go above and beyond." | Older models under-delivered without a push | Newer models over-trigger and over-engineer. Push thoroughness down; scope tight |
| Negative constraints — "Don't use X" | Models were loose enough to need fencing | Literal instruction-following makes negations backfire. Say what to do |
| Context babysitting — cram everything under 200k, trim obsessively | Context was scarce and precious | 1M is the default; manage by signal, not a percentage (rot is still real) |
| "Be conservative" reviews | Kept review noise down | 4.8 obeys and suppresses real findings. Ask for everything, severity-labeled |
| Turn-by-turn steering | Models lost the thread fast | Multi-turn costs more and the model rewards autonomy. Front-load the brief |
| Manual subagent choreography | The only way to parallelize | It can self-orchestrate parallel subagents now. Give it the goal + the bar |
The throughline: you used to push the model to do more. Now you constrain it to do exactly the right thing, and move your effort to the system around it.