Claude deviates from its own plans during execution

Resolved 💬 3 comments Opened Apr 13, 2026 by dvpadmin Closed May 24, 2026

Bug Description

When Claude Code creates an implementation plan (either via Plan mode or written out as a step-by-step approach), it frequently deviates from that plan during execution. Instead of following the agreed-upon steps, it wanders off into unrelated changes, skips steps, reorders dependencies, or invents new approaches mid-execution that contradict the plan.

The Pattern

  1. User and Claude agree on a plan (e.g., 5 concrete steps in order)
  2. Claude starts executing step 1
  3. Somewhere during execution, Claude drifts — introduces changes not in the plan, skips to step 4, or reimagines the approach entirely
  4. User has to intervene to pull Claude back on track
  5. This repeats for nearly every step

Expected Behavior

A plan is a contract. The planning phase is where alternatives are discussed, trade-offs are weighed, and the approach is finalized. Once the plan is agreed upon, execution should be mechanical — follow the steps as written. If Claude thinks an alternative approach might be better, that concern belongs in the planning phase, not mid-execution.

During execution, Claude should:

  • Execute steps in the planned order, exactly as written
  • Not second-guess the plan or silently substitute alternative approaches
  • Not introduce scope creep or unplanned changes
  • Treat the plan as a finalized decision, not a starting suggestion

Actual Behavior

Plans are treated as loose suggestions. Claude appears to re-enter "planning mode" during execution — reconsidering approaches, inventing alternatives, and making judgment calls that were already resolved during planning. The plan exists in context but doesn't constrain behavior.

Impact

Planning becomes a waste of time. Users spend effort aligning on an approach, only to have to micromanage every step anyway. This is especially frustrating because the whole point of planning is to reduce the need for constant supervision. The result is that users either stop planning (and get chaotic execution) or plan AND micromanage (doubling their effort).

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗