What you will be able to do
- Feed a specific correction back to Claude within a conversation, with the reason behind it
- Match what your evidence showed to the change that addresses it: prompt wording, examples, project knowledge or effort
- Make a repeated correction permanent with project instructions, project knowledge, a Skill or CLAUDE.md
1.Feed corrections back while the work is in progress
When a response misses the mark, the quickest fix is usually in the conversation itself. Claude's help centre lists this among its core prompting principles: "Give Claude feedback: you provide Claude with follow-up instructions, clarifications, or ask it to rewrite an answer." A useful correction is as specific as the feedback you'd want to receive yourself. It names what was wrong and what to do instead. Claude Code's documentation shows a correction as short as "That's not quite right. The issue is in the session handling." It is short, but it points at the exact place where the work went wrong.
Say why, not just what. Anthropic's prompting guidance notes that explaining the motivation behind an instruction "can help Claude better understand your goals and deliver more targeted responses", and that "Claude is smart enough to generalize from the explanation." A bare rule fixes one output. A reason helps Claude apply the fix to related cases you didn't mention.
Some feedback is about length or completeness. The help-centre article on short answers advises: "If its responses are too brief or only partially complete, consider retooling your prompts." When an answer stops partway through, it recommends "giving Claude the previous prompt and response when writing your next prompt to pick up where it left off." Building on the existing output this way is better than starting again from nothing.
Checkpoint 1 of 4· Check yourself
Claude's draft policy summary uses ellipses throughout, and the summaries are going to a text-to-speech system. Which correction is most likely to fix this case and related ones?
Giving the reason lets Claude understand the goal and apply it more widely, for example to other symbols a speech engine would stumble over. A bare rule, a blind regeneration or a vague request gives Claude nothing to generalise from.
“Claude is smart enough to generalize from the explanation.”Source: platform.claude.com
2.Choose the change that matches what the evidence showed
Once your checks show a pattern, the next decision is which lever to pull. Different findings call for different changes, and using the wrong one wastes effort.
If the format or tone keeps drifting even with clear instructions, add an example. Anthropic's prompt-engineering blog recommends examples when "Simple instructions haven't produced consistent results". Its advice is to "Start with one example (one-shot). Only add more examples (few-shot) if the output still doesn't match your needs." Anthropic's prompting guidance calls examples "one of the most reliable ways to steer Claude's output format, tone, and structure."
If outputs confidently repeat something out of date that is still in your project knowledge, rewording the prompt won't help. Anything uploaded to a project's knowledge base "will be used across all of your chats within that project", so the outdated document has to be replaced.
If the outputs contain claims nobody checked, the help centre is clear that "Users should not rely on Claude as a singular source of truth and should carefully scrutinize any high-stakes advice given by Claude." The thumbs-down button sends feedback on an unhelpful response to Anthropic. It doesn't change your own setup.
Sometimes the evidence points at reasoning depth rather than wording. For developers using the API, Anthropic's guidance for recent models is direct: "If you observe shallow reasoning on complex problems, raise effort to high or xhigh rather than prompting around it." The opposite case matters just as much. That advice covers shallow reasoning on problems that really are complex. If the prompt never said what depth, audience or format was needed, more reasoning won't fill that gap. The prompt needs to say it.
Checkpoint 2 of 4· Match them up
Match each finding from your review to the change that addresses it.
Tap a term, then the definition that fits it.
Each finding points to a different lever: examples steer format, the knowledge base controls what Claude draws on, effort controls reasoning depth, and high-stakes content needs human checking.
“If you observe shallow reasoning on complex problems, raise effort to high or xhigh rather than prompting around it.”Source: platform.claude.com
3.Close the loop: make a working fix permanent
If you give Claude the same correction every week, you're fixing the same problem over and over. Once your evidence shows a fix works, move it somewhere that applies automatically. Project instructions are the most direct option: "Claude will use these instructions for all the chats within the project." The help centre describes them as useful when you "need Claude to maintain consistent context across multiple conversations within the same project." Reference material belongs in the project knowledge base. Anthropic's usage guidance recommends projects "for anything you'll reference multiple times."
When the fix is a procedure rather than background context, such as a house format for a recurring report, a Skill packages it. Skills "teach Claude how to complete specific tasks in a repeatable way". On Team and Enterprise plans, Owners can provision them for the whole organisation. In Claude Code, the equivalent is CLAUDE.md, "A markdown file where you store project-specific instructions, conventions, and context that Claude should know every session."
| Mechanism | How it applies | Best home for |
|---|---|---|
| Projects | Static background knowledge that's always loaded when you start chats within them | Context and reference documents for one body of work |
| Skills | Task-specific; load dynamically when relevant and work everywhere across Claude | Repeatable procedures such as company formats and workflows |
| Custom instructions | Apply broadly to all your conversations | General preferences that hold for every chat |
The loop isn't closed until the fix has been checked. After updating the instructions or adding a Skill, run the same failing examples and the same criteria that exposed the problem. That's how you confirm the permanent version behaves like the one you tested in conversation.
Checkpoint 3 of 4· Check yourself
A team's members all paste the same formatting rules into Claude whenever they draft client proposals, in many different chats and projects. Which change best captures this as a reusable team practice?
A Skill captures a repeatable procedure once and loads whenever it's relevant, across the team, so nobody has to remember to paste the rules. One person's custom instructions only affect that person.
“Organizational knowledge capture: Package your company's workflows, best practices, and institutional knowledge for Claude to use consistently across your team.”Source: support.claude.com
Checkpoint 4 of 4· Exam question
An assistant notices that Claude's drafts in a shared project have become weaker since a teammate edited the project instructions last week. What should she do first?
Correct answer: C — Look at what changed in the instructions and retry a recent task with the old wording restored.
- A. Abandoning the project discards the shared context instead of fixing an edit that can be reverted.
- B. Claude cannot know which wording produced last month's better drafts, so this is guesswork.
- C. Correct. A change in results that follows a known change in configuration is best tested by reverting that change on a real task, which confirms or rules out the cause in one step.
- D. Layering instructions on top may mask the conflict while leaving the contradictory wording in place.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Once you correct Claude in one chat of a shared project, every other chat in that project picks up the correction.Why is that wrong?
Chats in a project don't share context with each other. A correction only carries over once it's in the project instructions or knowledge base.
Covered in Close the loop: make a working fix permanent
2.When Claude reasons shallowly on a really complex API task, the fix is to keep adding prompt instructions.Why is that wrong?
Anthropic's guidance says to raise the effort setting in that case. Prompt changes are for gaps in what you asked, not for reasoning depth the model wasn't given room for.
Covered in Choose the change that matches what the evidence showed
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://support.claude.com/en/articles/7996857-my-prompt-isn-t-giving-me-a-helpful-answerOfficial docs
“Give Claude feedback: you provide Claude with follow-up instructions, clarifications, or ask it to rewrite an answer.”
↩︎ Feed corrections back while the work is in progress - 2.
“That's not quite right. The issue is in the session handling.”
↩︎ Feed corrections back while the work is in progress“Your CLAUDE.md. A markdown file where you store project-specific instructions, conventions, and context that Claude should know every session.”
↩︎ Close the loop: make a working fix permanent - 3.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“can help Claude better understand your goals and deliver more targeted responses”
↩︎ Feed corrections back while the work is in progress“Examples are one of the most reliable ways to steer Claude's output format, tone, and structure.”
↩︎ Choose the change that matches what the evidence showed“Claude is smart enough to generalize from the explanation.”
↩︎ Checkpoint - 4.https://support.claude.com/en/articles/8114518-claude-s-response-to-my-prompt-is-too-briefOfficial docs
“If its responses are too brief or only partially complete, consider retooling your prompts.”
↩︎ Feed corrections back while the work is in progress“giving Claude the previous prompt and response when writing your next prompt to pick up where it left off”
↩︎ Feed corrections back while the work is in progress - 5.https://claude.com/blog/best-practices-for-prompt-engineeringSecondary source
“Pro tip: Start with one example (one-shot). Only add more examples (few-shot) if the output still doesn't match your needs.”
↩︎ Choose the change that matches what the evidence showed - 6.
“Anything you upload to this space will be used across all of your chats within that project.”
↩︎ Choose the change that matches what the evidence showed“Claude will use these instructions for all the chats within the project.”
↩︎ Close the loop: make a working fix permanent“Note: Context is not shared across chats within a project unless the information is added into the project knowledge base.”
↩︎ Exam trap 1“Note: Context is not shared across chats within a project unless the information is added into the project knowledge base.”
↩︎ Prediction - 7.https://support.claude.com/en/articles/8525154-claude-is-providing-incorrect-or-misleading-responses-what-s-going-onOfficial docs
“Users should not rely on Claude as a singular source of truth and should carefully scrutinize any high-stakes advice given by Claude.”
↩︎ Choose the change that matches what the evidence showed - 8.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-sonnet-5Official docs
“If you observe shallow reasoning on complex problems, raise effort to high or xhigh rather than prompting around it.”
↩︎ Choose the change that matches what the evidence showed“If you observe shallow reasoning on complex problems, raise effort to high or xhigh rather than prompting around it.”
↩︎ Exam trap 2 - 9.https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-featuresOfficial docs
“need Claude to maintain consistent context across multiple conversations within the same project”
↩︎ Close the loop: make a working fix permanent - 10.
“Use projects for anything you'll reference multiple times.”
↩︎ Close the loop: make a working fix permanent - 11.
“Skills teach Claude how to complete specific tasks in a repeatable way”
↩︎ Close the loop: make a working fix permanent“Organizational knowledge capture: Package your company's workflows, best practices, and institutional knowledge for Claude to use consistently across your team.”
↩︎ Checkpoint