A failing session still looks busy. Files change, commands run, every answer sounds certain. Meanwhile it has suggested something you already rejected, named a file that does not exist, and quietly added a second feature.
The idea
Five different failures hide behind that motion, and they need different moves. The session can lose an earlier decision, bury the useful material, invent a plausible detail, grow past its boundary, or stop being checked because the answer looks authoritative. Diagnose the signal before choosing the move, because the fix for one makes another worse.
| Mode | First visible signal | Confirming check | Immediate move |
|---|---|---|---|
| Drift | A ruled-out idea returns | Compare with the recorded decision | Restate once; restart if it repeats |
| Dilution | A loaded fact is missed | Ask for the exact evidence | Trim to the smallest complete slice |
| Hallucination | A path or API cannot be found | Search the project or the primary docs | Verify the claim before fixing anything |
| Scope creep | Unasked-for files or goals appear | Compare the diff with the boundary | Revert or queue the extra work |
| Automation bias | "Done" replaces your own check | Open what it changed and confirm the claim | Pause and inspect the evidence |
How it works
The five sit on three different layers, which is why one habit does not cover them.
- Drift, dilution and hallucination happen in the session. Anthropic describes context as a finite attention budget with diminishing returns, which is dilution's mechanism. GitHub warns in its responsible-use documentation that generated code can look valid while being wrong, and OpenAI's 2025 paper on why language models hallucinate describes models guessing rather than admitting uncertainty.
- Scope creep is a workflow failure. Every individual change can be sound while the whole job is out of bounds. Check it against the The list of exactly what changed in your files, line by line., compared with the files and goals you agreed to. Fluency proves nothing here.
- Automation bias happens on your side. In a 1999 experiment by Skitka, Mosier and Burdick, published in the International Journal of Human-Computer Studies, people using a highly but imperfectly reliable decision aid missed the events it failed to flag and followed its wrong prompts even when reliable contradictory indicators were on screen. The task had nothing to do with code. The same reliance forms wherever a fast and usually-right aid answers before you do.
What to do
- Treat the first odd signal as information rather than noise, and stop treating momentum as evidence.
- Make one confirming check from the table: the recorded decision, the exact source, the path, the diff, or the test.
- Match the move to the diagnosis. Restart drift, trim dilution, verify hallucination, queue scope creep, and check automation bias yourself.
Where it breaks
The five names are diagnoses, not measurements. A missing file can be a typo, a forgotten constraint can come from an unclear request, and a large diff can be exactly what the job required. Do not label a session from one ambiguous symptom. Confirm once, then act.
They also travel together. Dilution can precede drift, a hallucinated path can tempt scope creep, and automation bias is how any of them gets accepted. Work from the earliest one you can prove.