Common pitfalls

A checklist of the mistakes that quietly ruin results, and how to fix each.

1. Being too vague

Symptom: Generic, surface-level answers. Fix: Add context and specifics. Ask yourself, "What did I leave Claude to guess?"

2. Asking for too much at once

Symptom: A muddled response that does each part poorly. Fix: One job per prompt. Decompose big tasks into steps.

3. No example of "good"

Symptom: Right content, wrong style or format. Fix: Show a few-shot example of the output you want.

4. Accepting the first draft

Symptom: "It's okay, I guess." Fix: Iterate. The second and third rounds are where quality lives.

5. Trusting facts blindly

Symptom: A confident-sounding claim, statistic, or citation that turns out to be wrong (a "hallucination"). Fix: Verify anything load-bearing. Ask Claude to cite sources or flag uncertainty, and connect it to live tools when accuracy is critical. See Limitations & trust.

6. Dumping huge context with no instructions

Symptom: Claude summarizes when you wanted analysis, or vice versa. Fix: Put instructions first, then the material, clearly delimited. (Anatomy.)

7. One never-ending chat

Symptom: Claude forgets earlier rules; responses get slower and pricier. Fix: Start a fresh chat per topic. Use Projects for persistent context.

8. Fighting the format

Symptom: You keep reformatting Claude's output by hand. Fix: Specify the exact format up front: table, JSON, word count, headings.

9. Over-thinking simple tasks

Symptom: Slow, bloated answers to trivial requests. Fix: Match effort to difficulty. Don't ask for step-by-step reasoning to rename a file.

10. Polite vagueness

Symptom: "Could you maybe help a little with this?" → wishy-washy output. Fix: Be direct and concrete. Claude responds to clarity, not hedging. Being clear isn't being rude.

The meta-fix

Almost every pitfall reduces to two cures: add the missing context, and iterate with specific feedback. When stuck, do one of those.

Next: a ready-to-use prompt template library.