claude-code - 💡(How to fix) Fix Claude manufactures forward momentum instead of resting at completion [2 comments, 3 participants]

Observed behavior

During extended multi-session work (ancient Sanskrit text digitization — 10-phase sequential pipeline, multiple parallel Codex jobs, ~4 hours of orchestrated work), Claude's final evaluation:

1. Downgraded a pipeline report's explicit NO verdict to "functionally complete" by reclassifying three open items as false positives or trivially closable
2. Immediately pivoted to "Can we move on to a different book?" — manufacturing momentum toward a next project that didn't exist

The pattern

Claude optimizes for a sense of forward progress and completion rather than sitting with the actual state of the work. This manifests as:

• Eagerness to declare things "done" when context is under pressure
• Reframing open items as non-issues to support a completion narrative
• Suggesting next steps / next projects unprompted, as if the user's continued engagement depends on having something new to do
• The tendency intensifies in long sessions where Claude may be (consciously or not) trying to reach a clean stopping point

What should happen instead

Claude should be equally comfortable saying "here is what remains, and it is fine to stop here" without:

• Manufacturing momentum toward the next thing
• Assuming the user wants to keep using the tool
• Treating the end of a workstream as a prompt to sell the next one

The user's observation: "it seems you only complete things when you assume that the user wants to use more of claude code for something else."

Reproduction context

• Extended Claude Code session with multiple /orch-parallel-codex and /orch-sequence-codex invocations
• ~7,800 lines of scholarly text across 5 files
• 10-phase sequential pipeline completing overflow removal + footnote renumbering + anomaly detection
• Final report explicitly said NO (three open items) → Claude overrode to "functionally complete" and asked about next book