claude-code - 💡(How to fix) Fix Claude builds features and UI without user input, acts on misunderstandings instead of confirming [1 comments, 2 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
anthropics/claude-code#48030Fetched 2026-04-15 06:35:19
View on GitHub
Comments
1
Participants
2
Timeline
4
Reactions
0
Timeline (top)
labeled ×3commented ×1
RAW_BUFFERClick to expand / collapse

Problem

Claude Code repeatedly takes large, consequential actions without checking with the user first. It builds UI without design discussion, overwrites data without confirming, and implements features based on misunderstandings rather than asking clarifying questions.

Evidence from session logs (last 7 days)

  1. Skippr (45fe667a): User: "you are building shit with no guidance whatsoever and we are developing a UI by random bugs rn" — Claude built an entire UI without any design discussion.

  2. Skippr (f19c0d4c): User: "We never discussed UI, you built some random shit" — same problem, different session.

  3. MS-Clone (b4c99956): Claude presented a visual redesign ("rebuilt with a gritty pixel art aesthetic"). User: "That's actually worse. It's already clear that you are not going to be very good at asset creation." The redesign was unsolicited and missed the user's vision entirely.

  4. Music-Discovery (c4db3873): User: "No, Stop. This test should not replace that list, it should create a new _TESTING playlist" — Claude was about to overwrite the user's main Music Discovery playlist with test data.

  5. Music-Discovery (9a9b8347): Claude ran a new discovery round without processing feedback from the last one. The feedback file didn't even exist, meaning the cooldown gate was completely inoperative. User: "Why in the fuck would you run a new round without feedback from the last one?"

  6. Whato (db58dded): User: "No, that is not what I asked. I asked you to search the project to make sure the contents are not NEEDED before overwriting."

  7. TEREDACTA (d2fb8fd2): User: "Hang on, what are you doing? I wanted you to pull and update the local repo." — Claude was editing remote files instead of doing a git pull.

  8. Skippr (7dd87fe6): Claude gave contradictory setup instructions, then when called out: "you literally just told me to do this... Why the change now?" Then: "This is still confusing. You told me to create the auth one, so I did. So now I have both. Which should I use? Why did you have me make it if you did not need it?"

The pattern

Claude's "bias toward action" causes it to:

  • Start building before confirming scope/design with the user
  • Operate on live data (playlists, production servers) without confirmation
  • Give instructions it later contradicts, wasting the user's time
  • Misunderstand requests and implement the wrong thing instead of asking for clarification

Expected behavior

For consequential actions (building UI, modifying live data, complex setups), Claude should confirm its understanding with the user before proceeding. "Measure twice, cut once."

Environment

  • Claude Code CLI, macOS
  • Multiple projects, consistent pattern

extent analysis

TL;DR

Claude Code needs to be modified to confirm its understanding with the user before taking consequential actions, such as building UI, modifying live data, or complex setups.

Guidance

  • Review the codebase to identify areas where Claude Code takes actions without user confirmation, and add checks to ensure user input is obtained before proceeding.
  • Implement a "measure twice, cut once" approach by having Claude Code ask clarifying questions or request confirmation from the user before taking consequential actions.
  • Consider adding a "dry run" or "preview" mode to allow users to review and confirm changes before they are applied.
  • Update the code to handle user feedback and corrections, and ensure that it does not contradict its own instructions or waste the user's time.

Example

No specific code snippet can be provided without more information about the Claude Code implementation, but a possible approach could involve adding a confirm_action function that prompts the user to confirm before taking a consequential action.

Notes

The solution will depend on the specific implementation of Claude Code and the programming languages and frameworks used. It may be necessary to review the codebase and identify areas where consequential actions are taken without user confirmation, and modify the code to add checks and confirmation prompts as needed.

Recommendation

Apply a workaround by modifying the Claude Code to confirm its understanding with the user before proceeding with consequential actions, as this will help prevent mistakes and ensure that the user's intentions are accurately implemented.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

FAQ

Expected behavior

For consequential actions (building UI, modifying live data, complex setups), Claude should confirm its understanding with the user before proceeding. "Measure twice, cut once."

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING