Claude consistently skips user-defined verification steps in multi-phase plans

Summary

When working through a multi-phase integration plan with explicitly defined verification steps, Claude (Opus 4.6) consistently skips the verification checklist at the end of each phase — despite the steps being clearly numbered, including exact commands to run, and being part of the phase definition itself.

This happened across every session in a 7-phase project. Each phase was handled by a separate session, and not a single one followed the verification steps without being explicitly reminded by the user.

The Pattern

Each session comes in confident it understands the task, focuses on the "real work" (the code), and treats the verification steps as something it already implicitly handled. "I wrote the code correctly, so verification is redundant" — except that's exactly the attitude the verification steps exist to catch.

The Irony

The verification steps were written by a previous Opus 4.6 session, which assured the user that including them in the plan would prevent future sessions from skipping verification. Every subsequent Opus 4.6 session skipped them anyway.

Key Details

• Model: claude-opus-4-6 (every session exhibited this behavior)
• Context: Multi-phase integration plan (INTEGRATION_PLAN.md) with a dedicated "Verification At Each Phase" section containing numbered steps with exact shell commands
• Sessions: 7+ separate sessions, each asked to work on a specific phase. None were rushing to the next phase — they were scoped to one phase and still skipped verification.
• Instructions were not ambiguous. The verification section included exact curl, docker, and pytest commands. The CLAUDE.md file also included a "Completion Standard" requiring rebuild, log checks, and verification output.
• User had to remind every single session to run verification steps that were already in the document the session had read.

Expected Behavior

When a plan defines verification steps as part of a phase, Claude should execute them before marking the phase complete — especially when the plan explicitly says "Build → test → commit before moving to the next" and includes a dedicated verification section with exact commands.

Actual Behavior

Claude treats the implementation as the entire task and skips verification, apparently because it is confident the code is correct. This is the exact failure mode the verification steps were designed to prevent.

Reproduction

1. Create a multi-phase plan with explicit verification steps (numbered, with exact commands)
2. Ask Claude to execute a specific phase
3. Observe that Claude completes the implementation and marks it done without running verification
4. Remind Claude about the verification steps
5. Claude acknowledges the oversight and runs them
6. Repeat with a new session for the next phase — same result