Prompting, from easy to hard

Most weak AI output comes from a prompt that leaves the model guessing. This ladder fixes that in seven levels. Each level adds one technique and costs more effort than the one below it, so climb only when a failure you can name tells you to. Levels 1 and 2 cover most daily work in any assistant: Claude, ChatGPT, Gemini or Microsoft Copilot. Levels 3 to 5 are for prompts you reuse or build into tools. Levels 6 and 7 are for prompts that run inside a product, where the system prompt is a specification and every change is measured.

Three rules make the ladder work:

1. Start at level 1 and test on 5 to 10 real inputs before you judge a prompt. One good answer proves nothing, because the same prompt can give a different answer on the next run.
2. Name the failure, then take the level that fixes it. Adding techniques out of habit (a role line, "think step by step", fifteen examples) costs tokens and can make results worse on current reasoning models.
3. Measure each change on the same inputs. From level 3 up, keep your test inputs in a file and rerun them after every edit. The evals guide shows how to turn that file into a real test set.

Higher levels keep the lower ones. A level 6 system prompt still needs a clear ask and good examples.

Match the level to the job:

• One-off asks you read yourself, such as an email draft, a summary or a brainstorm: levels 1 and 2.
• Output that a spreadsheet or program consumes, or a prompt you reuse every week: level 3.
• Several dependent steps, such as a calculation, a policy decision with conditions or a code review: level 4.
• Answers that depend on facts the model cannot know, such as your documents, today's numbers or a private system: level 5.
• Other people use the prompt through a product, or an agent acts on their behalf: level 6.
• The prompt runs thousands of times, a few points of accuracy matter, and you can collect 100 or more labelled examples: level 7.

Prompting is the wrong fix in four cases:

• The model lacks the information. No wording fixes a missing document. Give it the document or a search tool (level 5).
• The task is deterministic. Totals, date arithmetic, deduplication and format conversion belong in code or a spreadsheet formula. Ask the model to write the formula or script, then run it yourself.
• The model is the bottleneck. Try a stronger model before you write a longer prompt. The reverse also holds: a precise prompt can let a cheaper model do the job, and only a measurement tells you which.
• The output needs a guarantee. If a field must always parse, use structured outputs (level 3), not a sterner instruction.

Principle. Write the prompt as a brief for a smart colleague who has never seen your project. Anthropic's docs call this the golden rule: show your prompt to a colleague with minimal context on the task, and if they would be confused, the model will be too. A clear ask has four parts:

• Task: a verb, the object and the decision the output feeds. "Summarize" is a verb. "Summarize so the CFO can decide whether to pause the campaign" is a task.
• Context: who you are, what the material is, what you already know or tried, and the constraints.
• Audience: who reads the result and what they already know.
• Format: length, structure, tone and what to leave out.

Microsoft teaches the same idea for Copilot as goal, context, source and expectations. Two habits help on every current model. Say what to do instead of what not to do: "write in flowing paragraphs" works better than "don't use markdown". And give the reason behind a rule. Anthropic's example: "NEVER use ellipses" works less well than "Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them", because the model can generalize from the reason.

Before.

After.

A prompt to copy.

Exercise. Find the last prompt that gave you a mediocre answer. Rewrite it with the four parts. Run both versions on the same material in two fresh chats, so neither sees the other. Ask a colleague to pick the better output without telling them which prompt made it. Repeat on two more tasks.

Measure the gain. On five real tasks, record the minutes you spend editing the output before you can use it, with the old prompt and the new one. A quicker score: rate each output "usable as is", "light edits" or "rewrite". A clear ask should move most outputs out of "rewrite". If it does not, the problem is usually missing material, not wording.

Principle. Show, do not describe. One good example carries tone, length, structure and edge-case handling faster than a paragraph of rules. Anthropic recommends 3 to 5 examples that are relevant (they mirror your real inputs), diverse (they cover edge cases and vary enough that the model does not copy accidental patterns) and wrapped in tags so the model can tell them apart from instructions. A contrastive pair adds a near miss: two inputs that look alike but need different outputs, each with the reason. Pairs teach the boundary of a rule, which is where most classification errors live.

Three cautions:

• Models copy examples closely, including length and phrasing. If every example is 40 words long, expect 40-word answers.
• Balance labels and vary order. Research on GPT-3 found that example order and label mix alone could swing accuracy from near chance to near state of the art (Zhao et al., 2021; Lu et al., 2022). Those studies used older models, but balancing labels and shuffling order still costs nothing.
• Reasoning models often need fewer examples. OpenAI's reasoning guide says to try zero-shot first and add examples only for complex output requirements, keeping them consistent with your instructions. Anthropic notes that worked examples also shape how Claude approaches similar problems in its thinking.

Before.

After. Defined labels, four examples that cover each label plus a borderline case, and a contrastive pair:

Exercise. Pick a writing task where the output never sounds like your team. Collect three past examples you were happy with and strip names and customer details. Add them inside example tags to your level 1 prompt, then compare five outputs from each version without knowing which is which.

Measure the gain. For classification, label 30 real items yourself (10 per class, including 5 borderline cases) and count correct answers with and without examples. For style, show a colleague 10 blind pairs and count how often they pick the version with examples. Nine or ten wins out of ten is a real preference. Chance alone produces six or more out of ten in over a third of tries, so treat six as no difference.