How to Give Claude Feedback That Improves the Next Draft
Correcting Claude's current work is different from sending product feedback to Anthropic. Here's how to give correction feedback that works, with concrete examples and how to keep lessons and prompt improvements for later.
Updated 5 min read
Giving Claude feedback means two different things. The first is correcting the current output so the next draft is better — this is in-session correction and Claude responds to it immediately. The second is sending product feedback to Anthropic about how the model behaves — this uses the thumbs up/down buttons or Anthropic's feedback channels, and it does not change the current session.
Most people want the first kind: a better version of the draft in front of them. The key is being specific about what to change and why. fenbs can hold the lessons from feedback rounds and the improved prompt versions so the next project starts from what already worked.
In-session correction: how to give feedback that works
Vague feedback produces vague revisions. "Make it better" leaves Claude to decide what better means. "Make it shorter" does not say how much shorter or which part to cut. Specific feedback — naming the part, the problem, and the change — gets a more useful revision.
The structure that works best is: locate the problem, name what is wrong with it, and say what you want instead. You can give all three in one sentence. "The second paragraph repeats the point from the intro — merge them into one" is complete feedback. "The opening is weak" is not.
Concrete feedback examples
Here are examples of weak feedback and specific alternatives for common situations:
| Weak feedback | Specific alternative |
|---|---|
| Make it shorter. | Cut this to 150 words. Remove the third bullet point and tighten the introduction. |
| This doesn't sound right. | The tone in the second paragraph is too casual. Replace "yeah" with "yes" and "it's" with "it is." |
| Add more detail. | The "how to set it up" section is missing the authentication step. Add it between steps 2 and 3. |
| The list is too long. | Keep only the top three points from the list. Remove any that cover the same idea as an earlier one. |
| Start over. | Keep the structure but rewrite the opening sentence. The current one starts with "I" — start with the benefit instead. |
| Not what I asked for. | I asked for a bulleted list but you wrote paragraphs. Convert the body to bullet points with one sentence per bullet. |
Giving feedback on factual or reasoning issues
When the problem is a factual error or a flawed argument, the feedback should name the specific claim and supply the correct information. "That date is wrong — the event was in March 2024, not January" gives Claude exactly what it needs to fix the error. "You've got the dates wrong" leaves it to guess which dates and what the correct ones are.
For reasoning issues, name the step where the logic breaks. "You concluded X from A and B, but B does not support X — B says the opposite" is useful. "The logic doesn't work" is not. The more precisely you identify the error, the less likely the next draft will introduce a new version of the same problem.
If you are unsure what the correct information is, say so and ask Claude to check: "I think this figure is wrong — can you verify it against [source]?" This is more useful than guessing and more honest than presenting the correction as certain.
When feedback does not produce the change you wanted
If Claude revises the output and still misses the point, the problem is usually in how the feedback was worded. Re-read the feedback you gave and check whether it could be interpreted differently. If it could, rephrase it more specifically.
Sometimes a different approach works better than more feedback. Show Claude an example of what you want: "Here's a sentence in the style I'm looking for: [example]. Rewrite the opening in this style." An example is often faster than a description.
If the same type of feedback is needed across multiple sessions, the problem may be in the initial prompt rather than Claude's interpretation of the feedback. If you find yourself correcting the same issue repeatedly, add the constraint to the prompt before the task rather than after the draft.
Sending product feedback to Anthropic
The thumbs up and thumbs down buttons in Claude.ai send a signal to Anthropic about the response quality. This is separate from correcting the current output; it informs Anthropic's ongoing work on the model, not the current session.
Use the thumbs down button when Claude's response is unhelpful, inaccurate, or inappropriate — not just when it did not match your specific preference. Rating a correct and useful response as negative because it was not what you wanted misleads the feedback. For systematic issues or safety concerns, Anthropic's feedback form at anthropic.com is more appropriate than a thumbs rating.
Save the lessons from feedback rounds
When a feedback round reveals something worth repeating — a wording that always needed the same correction, a prompt that consistently produced the wrong format — record it as a lesson in fenbs Lessons learned. Save the improved prompt version in Skills and prompts. Connected assistants read the lessons before similar work, so the next session starts from what you already learned.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How to do this with fenbs
- Create a free account, free for one person with no credit card, and add a project for the work.
- After a feedback round that produced a notably better draft, save the revised prompt in Skills and prompts in fenbs, noting what changed and why.
- If the feedback revealed a recurring type of error — a format issue, a scope problem, a missed constraint — add it as a lesson in Lessons learned.
- Connect Claude to fenbs through Integrations and ask it to read the project lessons and skills before starting the next draft.
- Over time, the accumulated lessons become the starting point for every similar task, so you do less corrective feedback.
fenbs does not give Anthropic feedback on your behalf and does not change how Claude's model behaves. What it provides is a record of what worked and what did not, accessible to every session that connects to the project.
For related reading on improving Claude's behavior more broadly, see what is Claude. The AI prompts topic hub has more on writing instructions that hold up across sessions.