7 Common Prompt Mistakes and How to Fix Them

Seven common prompt mistakes and their fixes, framed before/after from Anthropic's official prompt engineering principles — vagueness, too few examples, and more.

When a prompt underperforms, it's usually not because you're missing a clever trick — it's because of a few common mistakes repeated over and over. This article inverts the principles in Anthropic's official prompt engineering guide to lay out frequent anti-patterns and their fixes, in before/after form. If you'd rather see the principles first, read how to write good prompts; if you're brand new, start with prompting basics.

Vague → specific: one line changes the result ✗ Before (vague) "Clean this up" → length/format/audience unclear → Claude guesses everything → different result each time ✓ After (specific) "Summarize these notes into a 3-line recap + action items for a new hire" → audience/format/length clear → consistent result

Anti-pattern 1 — Being vague and hoping Claude figures it out

The most common mistake. The official guide compares Claude to "a brilliant but very new employee with no context." The more precisely you explain what you want, the better the response.

Fix: Give context (what the output is for, who the audience is, what a successful result looks like) and be specific about what you want. The guide's "golden rule": show your prompt to a colleague with minimal context on the task and ask them to follow it — if they're confused, Claude likely will be too.

Anti-pattern 2 — No examples, or too few

Describing your desired output format in words alone tends to produce inconsistent formatting.

Fix: Provide input/output examples. The official guide recommends including 3–5 examples for best results. Examples should be relevant to your actual use case and diverse enough to cover edge cases. If you don't have good ones, you can ask Claude to evaluate your examples for relevance and diversity, or to generate more based on your initial set.

Anti-pattern 3 — Mixing instructions, context, and data into one blob

Cramming long instructions, reference material, and the actual input into one undifferentiated paragraph makes it hard for Claude to tell instructions from data.

Fix: Structure with XML tags. When working with multiple documents, wrap each in <document> tags with <document_content> and <source> subtags to mark content and origin clearly.

No role given → assign a role via the system prompt Question only at the top → put it at the end for long context Relying on "tricks" → use principled, direct techniques No steps → break into sequential steps

Anti-pattern 4 — Not giving a role

Asking with no context tends to yield generic answers.

Fix: Assign a role via the system message. A single line like "You are a helpful coding assistant specializing in Python" shapes the tone and depth of the answer.

Anti-pattern 5 — Putting the question at the top of a long document

Pasting several documents with the question up front can bury the key ask under a long body.

Fix: In long context, place the query at the end. Per the official guide, this can improve response quality by up to 30% in tests, especially with complex multi-document inputs. For long-document tasks, also ask Claude to quote the relevant parts first before doing the task — it helps cut through the noise.

Anti-pattern 6 — Trying to trick the model

Forcing odd formats or adding irrelevant role-play to boost quality may stop working as models evolve.

Fix: Use principled, direct techniques. If you're adding convoluted instructions "just to see if it works," check first whether there's a clearer way — it saves time.

Anti-pattern 7 — Cramming everything into one prompt

Trying to finish a complex task in one giant prompt makes it easy to lose the thread mid-task.

Fix: Break instructions into sequential steps (numbered lists or bullets), and for genuinely complex work, chain the prompt across multiple stages. The clearer each step, the more reliably Claude carries out the task as intended.

Summary

  • Vague → specific context, goal, and format (golden rule: if a colleague is confused, so is Claude)
  • Too few examples → 3–5 relevant, diverse examples
  • Mixed input → separate instructions from data with XML tags
  • No role → assign one via the system prompt
  • Long context → put the question at the end; quote relevant parts first
  • Relying on tricks → principled, direct techniques
  • One giant prompt → break into steps, chain when needed

The recommendations here recast the principles in Anthropic's official prompt engineering guide (platform.claude.com/docs) into a "common mistake → fix" frame. Models and techniques change, so check the official docs for the latest.

Keep reading

Have a question or want to share how you use Claude?

Join the community to share tips with other users, or explore more guides.