The first layer. A prompt is the words a model is asked to act on — and most failures at every higher layer trace back to an instruction that was never clear in the first place. Nine lessons — the classic toolkit, then the frontier-model playbook: the effort dial, why-not-what prompting, and the fence.
The single highest-leverage move in prompting is stating the task the way you'd brief a smart new colleague: what to do, what the output should look like, what to do when unsure. Vague asks produce confident guesses; explicit asks produce checkable work. Add the audience, the format, and the failure rule (“say I don't know rather than invent”) and half your quality problems disappear.
Models pattern-match. One or three worked examples of input → ideal output teach format and judgment faster than paragraphs of description. Pick examples that cover an edge case, not just the happy path — the model learns from the differences between your examples.
For anything with steps — math, analysis, tradeoffs — asking the model to reason before answering measurably improves accuracy. Structure it: “First list the facts, then the options, then decide.” The trap: for trivial lookups, forced reasoning just burns tokens. Match the technique to the task.
The moment a prompt's output feeds anything downstream — code, a spreadsheet, another prompt — free text becomes a bug. Demand a schema: JSON with named fields, a table, XML tags. This is the bridge to every higher layer: a node contract in a graph is just a structured-output prompt with enforcement.
Prompting cannot fix missing information, stale knowledge, or multi-step work that exceeds one context window. When you hit those walls you don't need a better prompt — you need the next layers: context (what the model sees), harness (what it can do), loops (how it iterates). Recognizing which wall you hit is the skill.
Frontier models now expose an effort control — low, medium, high, xhigh — and it is the most misunderstood dial in prompting. Effort buys reasoning depth before the answer, not response length. Lowering it to get shorter replies gives you worse answers at the same length; raising it on easy tasks buys nothing but latency and cost. And the settings that were right for last year’s model are wrong for this one: newer models are so much stronger at low and medium that carried-over defaults quietly overpay.
Old models needed step-by-step instructions; frontier models need context and goal — they work out the steps, usually better than you would specify them. The formula: “I’m working on [larger task] for [who it’s for]. They need [what the output enables]. Request: [one clear sentence]. Output format: [structure]. Constraints: [what must not happen].” The why-first prompt is shorter than the checklist it replaces, and the output is better — the model fills in the context you forgot to specify. This softens Lesson 02’s instinct: examples still teach format, but over-specified steps now cap the model at your imagination.
“Double-check your answer.” “Include a final verification step.” “Use a subagent to verify.” On older models these helped; on frontier models they stack on top of the self-verification the model already performs — wasted tokens, longer responses, no quality gain. Anthropic removed all of them from Claude Code’s own prompts and measured no loss and a significant cost drop. The deeper lesson: when a model misbehaves, most people add instructions. The engineers who built it removed instructions until the model could breathe. The guardrails were the problem.
A proactive model will expand scope because it thinks it is helping. For agentic work the highest-value prompt is the fence: “Deliver what was asked, at the scope intended. If a better approach exists, say so in a sentence and continue with the task as asked.” Pair it with narration control — describe the communication style you want, positively, instead of listing don’ts — a subagent rule (“delegate only genuinely parallel work; don’t delegate what a few tool calls can finish”), and a checkpoint rule for autonomous runs: “pause only for destructive actions, real scope changes, or things only I can provide; otherwise keep going and report when done.”