TL;DR
Most 'Claude got worse' is actually a habit problem. Three quiet mistakes degrade every answer: you prompt instead of brief, so it guesses; you run one endless thread, so it contradicts itself; you accept the first answer, so bad output goes unchecked. Sharper input, sharper output - here is the fix for each.
Mistake 1: you prompt instead of brief
A prompt is a wish. A brief is context plus a target. When you type one line and hope, the model fills the gaps with guesses - and the guesses are the part you end up fixing. Give it the same brief you would give a competent freelancer: what you are trying to do, what you have tried, the constraints, and what a good result looks like.
Mistake 2: one endless thread
The longer a single conversation runs, the more competing, half-finished ideas pile up in its context - and the model starts contradicting itself and dragging in stale decisions. Start a fresh conversation for each distinct task. Clear the context between unrelated jobs so old threads stop poisoning new answers.
Mistake 3: you accept the first answer
The first answer is a draft, not a verdict. Accepting it unchecked is how a confident, wrong output ships. Make the model earn it: ask it to critique its own answer, list what could be wrong, or produce a second approach and compare. The best output usually comes from the second pass, not the first.
textBefore I use this, critique your own answer. What are the two
weakest assumptions here, what would break them, and what is a
different approach worth comparing? Then tell me which you would
ship and why.The pattern behind all three
- Feed it enough to not guess (brief, do not prompt).
- Keep the context clean (fresh thread per task).
- Never take the first draft as final (make it check itself).
Common questions
How much context is too much in a prompt?
You want enough to remove guessing and no more. Include the goal, what you tried, the constraints, and what good looks like. Skip the backstory that does not change the answer. If a detail would change what a smart stranger builds, include it.
When should I start a new conversation versus keep going?
Start fresh whenever the task changes. Keep the thread only while you are iterating on the same piece of work. The moment you switch to something unrelated, clear it - stale context is where contradictions come from.
Does asking Claude to critique itself actually help?
Yes. A second pass that names weak assumptions and offers an alternative reliably surfaces problems the first draft glossed over. It is the cheapest quality check you can run, and it takes one extra prompt.
Is this about the model or about me?
Mostly you - and that is good news, because it is the part you control. The same model gives far better output to a clear brief, a clean thread, and a second look. Fix the habits and the tool gets 'smarter' overnight.
Want the prompting guide that fixes all three?
Get 650+ plug-and-play skills, MCPs & prompts, plus 5,000+ builders - $9/mo, cancel anytime.
Join the Club