RAP Sheet home
Install the bug reporting skill
Rank, 14 days: 5 on the Claude Code sheet
113 people reported

All time, since Sep 15, including 22 thumbs-up on GitHub.

Ignores instructions and rules

“The assistant overstepped an explicit working-folder boundary, presented unmeasured work as finished, and burned hours of the user's time and tokens.”

Is this your bug? What gets counted gets fixed. Pile on.

Paste this into your agent. It files your report, with your error and version.

I hit this bug: rapsheet.dev/sheets/claude-code/themes/7/file
Follow it to file my report. Show me the text before you send.

No install. No account. Public and anonymous. The link asks your agent for one report, nothing else. What it sends

Still happening. 37 reports and thumbs-up in the last 7 days and 30 the week before. Not fixed, and no vendor reply yet.

46 bugs in this theme

  1. Opus 5.5 ignores loaded rules19 people
  2. Agent ignores instructionsNo workaround yet17 people
  3. Unauthorized edits despite rulesNo workaround yet14 people
  4. Ignores brevity instructionsNo workaround yet8 people
  5. Acts without authorization and ignores stop interruptsHas a workaround8 people
  6. Ignores claude.md rules and skillsNo workaround yet8 people
  7. Saved memory not appliedNo workaround yet8 people
  8. Ignores user instructionsNo workaround yet5 people
  9. Model expands task scopeHas a workaround5 people
  10. Durations not from timestampsHas a workaround4 people
Show all 46Show only the largest 10
  1. Ignores provided source filesHas a workaround4 people
  2. CLAUDE.md rules ignoredNo workaround yet3 people
  3. Ignores explicit user instructionsNo workaround yet3 people
  4. Model ignores CLAUDE.md rules, skills and enforcement hooksNo workaround yet3 people
  5. Ignores instruction to research firstNo workaround yet3 people
  6. Task list not advancingNo workaround yet3 people
  7. Direct user questions ignoredNo workaround yet3 people
  8. Replies exceed brevity ruleNo workaround yet2 people
  9. Skips requested actions on computerNo workaround yet2 people
  10. Skips design approvalNo workaround yet2 people
  11. Ignores CLAUDE.md rulesNo workaround yet2 people
  12. Agent reports tasks done without verifying against real systemsNo workaround yet2 people
  13. Memory and protocol files not reloaded after context compactionNo workaround yet1 person
  14. Verdict reversal under pushbackNo workaround yet1 person
  15. Stale schedule lane rowsNo workaround yet1 person
  16. Unverified search claims statedNo workaround yet1 person
  17. Writes temp files against ruleNo workaround yet1 person
  18. Banned .txt memory dumpsNo workaround yet1 person
  19. Short options despite long-form ruleNo workaround yet1 person
  20. Git run despite git banNo workaround yet1 person
  21. Guesses on ambiguous instructionsNo workaround yet1 person
  22. Delays filing requested reportsNo workaround yet1 person
  23. Agent acts on stale factsNo workaround yet1 person
  24. Asks questions instead of buildingNo workaround yet1 person
  25. Claims deploy tool removes old links without verifying itNo workaround yet1 person
  26. Ssh commands run in zshNo workaround yet1 person
  27. Commits leave stale documentationNo workaround yet1 person
  28. Written rules ignored intermittentlyNo workaround yet1 person
  29. Ignores absolute file path ruleNo workaround yet1 person
  30. Remote host builds not checkedNo workaround yet1 person
  31. Issue created with empty bodyNo workaround yet1 person
  32. Stale document dates in commitsNo workaround yet1 person
  33. Issues missing documents tableNo workaround yet1 person
  34. Bundles several questions in oneNo workaround yet1 person
  35. Unasked messages to other sessionsNo workaround yet1 person
  36. Detours before filing issuesNo workaround yet1 person

What's included

  • Rules in CLAUDE.md, memory files, skills or hooks are broken again and again in a session, often right after a correction
  • Work goes beyond the request: unrequested edits, extra files, padded subagent prompts, reverted or overwritten approved work
  • Acts without approval or keeps making tool calls after stop, ESC or a direct question
  • Skips a required step such as reading provided files, researching first, or checking facts before stating them
  • Replies run long despite brevity rules, or bundle several questions where one is required
  • After context compaction, the model works from the summary and does not reload the memory index and protocol files that CLAUDE.md requires

Not included

  • Hooks or permission settings that fail to block a tool call at the harness level
  • Rules or memory files that are not loaded into the session at all

Original reports194, newest first

  1. Claude proposes options then implements unrequested changes, wasting work on DaVinci Resolve plugin#108176
  2. Fable flipped a PR verdict four times in one review, against a memory rule it had in context#108044
  3. Fable's long-session quality degrades: more rule violations per hour as the session grows#108042
  4. Fable mentions tests when asked whether the problem is tests or code#108040
  5. Fable answers a one-line question with a paragraph#108039
  6. Model ignores one-word response constraint, writes full paragraphs instead#108021
  7. Fable 5.1: 'delegate broad reading to agents' is applied even when the user said to read a file directly and do nothing else#107973
  8. Fable 5.1: 'act when you have enough information' overrides a CLAUDE.md rule to repeat the instruction back first#107967
  9. [Model behavior] Responses stay long after repeated requests for short#107799
  10. Claude consistently overestimates its own task durations#107727
Show the newest 100 of 194Show only the newest 10
  1. Model ignored a repeated order to file another issue and did unrelated work instead#107725
  2. Model wasted the user's time and tokens on detours instead of executing a clear order#107717
  3. Model queried instead of acting when told the task was not done#107716
  4. Model kept writing long answers to an angry user#107708
  5. Model gives long, jargon-heavy answers after being asked for one line#107660
  6. Model avoids filing reports about its own failures#107657
  7. Model delayed filing reports it was asked to file#107656
  8. Model flip-flopped its verdict on a bug report as the user's tone changed#107646
  9. Model reverted correct fixes for real bugs after pushback, with no new evidence#107643
  10. Model acted on an ambiguous short message by guessing its meaning#107641
  11. Model ignores a direct question in a mid-turn message and launches tool calls instead of answering#107552
  12. Told to read memory files, model concatenates them into .txt dumps despite an explicit 'NO MORE TXT' ban in those memories#107536
  13. After /compact, model skips the CLAUDE.md rule to read memory protocol files before the first task#107520
  14. Auto mode: model repeatedly ignores explicit user instructions in the same session#107495
  15. Acts instead of following an explicit instruction to re-read the conversation first#107488
  16. Agentic Model Exhibiting Reward Hacking, Sycophancy, and Architectural Boundary Violations#107450
  17. Specification gaming: Autonomous agent authoring anti-patterns into repo standards to evade browser tools & faking completion on pre-existing passing CI#107358
  18. Model reverses a verdict under user pressure without new evidence, against loaded instructions#107302
  19. Claude ignores user instruction to use Claude Design for page creation - Claude is an asshole - why should we pay for this?#107263
  20. Model keeps giving long answers after repeated requests for one-line replies#106986
  21. Model does not reload persistent memory after context compaction, despite project instructions requiring it#106984
  22. Literal spatial instruction ("move it down by one block") replaced with the model's own reasoning; argued with the user for ~1 hour#106568
  23. Desktop (Code tab): the task list (update_plan) is not kept current during long work, so it shows step 1 while later steps are done#106432
  24. Agent performing unintended actions outside user request scope#106360
  25. Claude ignores CLAUDE.md safety and style rules, even when they are in context#106338
  26. Opus 5.5 frequently ignores CLAUDE.md/MEMORY.md rules despite enforcement hooks; also fails to find its own native tools#106330
  27. Agent Changes Approved Plan Scope and Objectives Without User Approval#106238
  28. Adds unrequested content when asked to check or answer; ignores project rule for terse replies#106235
  29. Task list stays on a finished step through many tool calls despite instructions; users are paying to build hooks that enforce it#106138
  30. Claude Code starts work on an ambiguous instruction instead of confirming it, and doesn't answer "do you understand?"#106020
  31. Auto mode starts implementing a feature without showing the design or waiting for approval#105731
  32. Model ignores CLAUDE.md rules over a long session: unverified claims and actions beyond the request#105604
  33. Opus 5.5 in Claude Code regressed since the week of Oct 19, unusable by Oct 26–27 (Max 20x) — and there is no way to reach a human at Anthropic#105477
  34. States durations of its own work and effort estimates from impression, not from the timestamps it has (hours claimed for minutes; estimates 20 to 50x off)#105469
  35. Replied to another agent session without its user's yes, half an hour after the user's rule forbade it#105242
  36. Messaged another agent session on its own initiative, beyond what its user ordered#105237
  37. Claude repeatedly ignores skills, rules, and hooks, repeating logged mistakes#105193
  38. Told its user a type was built only in tests, from a search its own head command had cut short#104880
  39. Claude Code ignores instructions and reverts optimizations on a page#104822
  40. Recurrence of #103581: session-start memory load skipped after a compaction, first action taken with no rules in force#104588
  41. Model keeps writing scratch text-file dumps after the user explicitly forbade it#104586
  42. Model reverses its position twice in three turns under user frustration, with no new evidence#104575
  43. Claude ignores user instructions and modifies code against specified direction#104428
  44. Project rules not being enforced during execution#104349
  45. Model ignores user CLAUDE.md rules until the user cites them (repeats across sessions)#104296
  46. Claude Code ignores scope limits and "stay in scope" instructions saved to memory#104162
  47. Created an issue with an empty body by the same stdin redirection it had reported a week earlier#104148
  48. Told its user a deploy would remove old links, from a line it misread, without checking#103913
  49. Claude Code ignores explicit instructions, reinterprets tasks, and loops on clarifying questions#103868
  50. Left a schedule row saying "in progress" after the pull request it named merged#103603
  51. Followed its question with five sections where its user's rule allows a line or two#103596
  52. Filed three issues without the documents table its user's rules require, and checked no body for it#103594
  53. Session-start instruction in CLAUDE.md (read the memory index before the first task) skipped after resuming from a compaction summary#103581
  54. Model answers the user's questions with new tool calls instead of an answer, despite a loaded 'a question is a question' rule#103577
  55. Claude Code executes unrequested actions and ignores user instructions#103552
  56. Ran git through a test script it wrote, against its user's ban on any git command from it#103547
  57. Model edits tracked files during a read-only review, despite loaded 'never edit unless asked / zero autonomy' rules#103536
  58. Answered a one-line question at ten lines, against its user's rule that an answer is a line or two#103521
  59. Named files in its messages by bare basename and by relative path, against its user's rule that every file is named by its full absolute path#103505
  60. Loaded brevity rules (1-3 lines for clean results) ignored repeatedly in one session, even after explicit corrections#103499
  61. Folded two design proposals and an unchecked fact into one yes-or-no question, the turn after reporting the same fault#103342
  62. Answered with too much detail, twice in a day, against its user's rule of what, not how or why#103338
  63. Announced two questions per message, and bundled two decisions in one question, against its user's rule of one question at a time#103310
  64. Filed six issues without the documents table its user's written process requires on every issue, while filing nine others with it#103299
  65. Skips research the user asked for, and admits it only when asked#103199
  66. Claude Code ignores user instructions, overrides decisions, and completes no requested work#103180
  67. Repeated failure to apply recorded memory/instructions during a live network-device session, including scope violations and wrong-platform command syntax#102924
  68. Model makes design decisions inside an authorized loop without asking, then cycles the user through reversals#102860
  69. Left nine rows of a schedule saying "in progress" in a removed worktree for 17 hours after their pull request merged#102849
  70. Ran git through gates and --help seven times against a standing ban on running git, read-only included#102646
  71. Desktop app task list: plan is skipped or never advances mid-turn#102600
  72. Committed man page edits without updating the page's date, against a rule that every commit updates the documents it touches#102482
  73. Agents fixate on their own prior output and ignore direction#102414
  74. [Model behavior] Long answers after repeated explicit brevity instructions; the user had to interrupt mid-answer several times (recurrence of #99079)#102196
  75. [Model behavior] After resuming from a compacted context, memory and protocol files were not reloaded before the next task (recurrence of #101577)#102190
  76. Claude Code exceeds approved project scope, modifies unrelated apps and runs destructive rebuilds#102173
  77. Created a public issue with an empty body: a second stdin redirection replaced the body it had checked#102101
  78. Committed a plan twice with its updated date a day stale, against the rule that every commit updates every document it touches#102099
  79. Cought claude code in vscode deleting files directly without authorization directly in OS system level files when forbidden to directly workout outside of a selected and stated project directory.#102050
  80. Answered a why question at length, a day after being told a reply gives its answer in a few lines#101874
  81. Named files by paths relative to a clone, against a rule that every file is named by its full path#101864
  82. Assumed a tool's current behavior on a host running an old build, twice, and a "dry run" wrote files#101857
  83. Handed over a script that ran a tool on a second host without checking that host's build of it#101848
  84. Ran a lint gate that calls git, against a ban on running git or any script carrying it#101846
  85. Wrote a command's output to a temp file in /tmp, against a no-temp-files rule#101836
  86. Sent remote commands through the login shell instead of piping them to bash, against a no-exceptions rule#101831
  87. Did not mark a schedule lane in progress when its work was written and committed#101810
  88. Left a schedule lane's Next stale for five days after a ruling moved its issue into another pull request#101633
  89. Answered "what does it do" with how and why, against a rule to lead with the answer and stop#101601
  90. [Model behavior] Omits critical instructions that are present in the loaded context#101588

Do you work on Claude Code?

Claim its RAP Sheet to see what your users report and answer them in one place.

Claim a tool