Getting better results

Not tricks, and not magic words. Almost everything that reliably improves an answer comes down to two decisions: what the model can see, and how precisely you have said what "good" means.

Published 12 August 2026 · Last verified 12 August 2026 · Each technique here is traced to the mechanism that makes it work, so it should outlast any particular tool

The idea underneath all of it

These systems are far better at reading than at remembering. Nearly every technique below is the same move in a different coat: take something the model would have to recall, and put it in front of it instead. If you remember only that, you will out-prompt most people who have memorised a list of tips.

What actually works

1. Give it the material rather than describing the material

The single largest jump in quality available to most people, and the one most often skipped because it feels like cheating. Paste the actual document, the real error message, the two emails you liked, the data. A model can use anything in front of it and can only guess at anything it has to recall. → Context windows

This also sidesteps the whole category of things it cannot know: your files, your business, and everything after its knowledge cutoff. → Training cutoff

2. Show an example instead of reaching for adjectives

"Professional but warm" means almost nothing. Two emails you thought were professional but warm mean a great deal. Anthropic's own guidance is blunt about this: examples are "one of the most reliable ways to steer Claude's output format, tone, and structure," and it recommends making them relevant to your real case and varied enough that the model doesn't latch onto an accidental pattern.

This is worth understanding as the cheap version of something expensive: demonstrating what you want is the same lever as fine-tuning, minus the cost, minus the commitment, and editable in ten seconds.

3. Say what the output has to be

Format, length, audience, and what to leave out. Every constraint you add removes a large space of plausible-but-unwanted answers, which is exactly how a system that predicts likely continuations gets steered.

The test I like best comes from the same guidance, which frames the model as "a brilliant but new employee who lacks context on your norms and workflows," and offers this as the golden rule: "Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too." That test costs nothing and catches most bad prompts.

4. Hand it the facts it has no way to know

Today's date if timing matters. Which version you are on. Your constraints, your audience, the decision already made that you don't want reopened. A model has no clock and no access to your situation, and it will not ask before assuming.

5. One task per message

Compound requests get partially completed, and the omission is silent rather than flagged. Ask for the summary, then ask for the rewrite. It feels slower and is usually faster.

6. Ask for sources, then actually look at them

Requesting citations improves a fluent answer in one specific way: it makes the answer checkable. It does not make invention impossible, and a model without a search tool can produce a reference that looks perfect and does not exist. → Hallucination

A model that can actually search is a different proposition, because it can go and read the current page rather than recall one. That converts a memory problem into a reading problem, which is a much better problem to have. → Tool use · Is what you're reading out of date?

7. Ask it to check its own work against something specific

Not "are you sure," which is close to useless, but a named standard: verify this against the test cases, check every figure against the table above, confirm each claim appears in the source I gave you. Anthropic recommends exactly this, appending an instruction to verify the answer against your criteria before finishing, and reports it catches errors particularly well on code and maths. It works because it turns a vague request for confidence into a concrete task with a right answer.

8. Start again rather than argue

When a long session has drifted, the instinct is to correct it repeatedly. That usually makes things worse: every correction is more text competing with your original instructions, and the earliest material is what gets squeezed out first. Open a fresh session and paste in the good context. → Context windows

What doesn't work, or isn't proven

Incantations about expertise. "You are a world-class expert" does far less than the internet believes. What people are reaching for when they write it is usually a real instruction: who the audience is, what depth to pitch at, which conventions to follow. Say that instead and you get the benefit people are hoping the phrase delivers.

Politeness, threats, and offers of payment. These are popular claims, and the honest position is that there is no dependable, documented lever here. Mechanically, any change to your text changes the input, so effects are not impossible; what is missing is evidence that they are controllable or worth your attention. Be polite because you'd rather be, not because it is a technique.

"Are you sure?" as a verification step. A challenge makes agreement the plausible continuation, so you often get a revision regardless of whether the original was right. Use a named standard, as above, or check it yourself.

Longer prompts as a proxy for better ones. Padding with material that isn't relevant dilutes what matters, and on long inputs a model asked to find one important line inside a large pile does worse than one handed the few passages that count. Relevance beats volume.

Secret templates. If a prompt is being sold as a proprietary formula, what you are buying is the thing this page just gave away. The techniques are not scarce, and nobody has a private one.

When no prompt will fix it

Knowing which failures are prompting problems saves more time than any technique here, because the rest are the wrong tool for the job.

It needs to be exact. Arithmetic, tallies, precise counts. Ask it to write the calculation or the query and run that, rather than asking for the number. → What AI is actually bad at
It needs to know your material. No wording retrieves a document the model was never given. Supply it, or build retrieval. → RAG and retrieval
It needs to be identical every time. Output varies by design. Reproducibility has to be engineered around, not requested.
It needs to be current. Give it search, or paste the current page in. → Training cutoff

Why this page has no prompt library

Copyable templates are what most prompting guides sell, and they are the fastest-rotting thing in the field: they are written for particular models, they stop matching how the products behave, and they teach nothing that transfers. The eight habits above are the whole trick, and they have not changed across several model generations.

Sources

The quoted lines on examples being a reliable way to steer output, the "brilliant but new employee" framing, the colleague test, and the recommendation to have the model verify its answer against stated criteria all come from Anthropic's prompting best practices, fetched 12 August 2026. That documentation is written for one vendor's models and contains model-specific advice not reproduced here. The section on what doesn't work is deliberately weaker in its claims: it reports an absence of dependable evidence rather than proof of no effect, because those are different things and only one of them is honest.

Next → Is what you're reading out of date? · Vocabulary → Glossary · All → Guides