You spend three evenings refining a billing pipeline with an AI, only to watch it propose Prisma on day four. You rejected Prisma on day one due to cold-start latency, but the standard session summary failed to capture that binding constraint. The summary was technically accurate but strategically useless. It recorded what happened, not what is settled. For developers working across multiple AI sessions, this drift is not a bug; it is a feature of how compression favors narrative over state.

The Failure of Narrative Compression

When you ask a model to summarize a long thread, it compresses the conversation. This process favors the story arc: what we discussed, what we built, and where we ended. It systematically drops the 'noise' that looks irrelevant but is actually load-bearing. Three critical data types vanish: rejected options, soft constraints stated once in message nine, and open questions that get mislabeled as resolved. A new session reads the summary as the full truth, treating absent decisions as undecided. This forces you to re-litigate settled arguments, wasting tokens and time.

Context Rot and the 'Lost in the Middle' Problem

Long threads degrade because of architectural limitations in how LLMs process context. Research titled 'Lost in the Middle' shows models recall information best at the start or end of the input, failing in the middle. In an eighty-turn chat, your early constraints sit in that dead zone. Anthropic’s engineering team describes this as 'context rot,' where recall decreases as token count grows. Their recommended compaction strategy keeps architectural decisions and unresolved bugs while discarding redundant tool output. A good handoff document is not a recap; it is a state injection that explicitly preserves commitments and boundaries.

Implementing Decision Locks

The solution is a structured handoff document with six fixed sections, where sections four and five do the heavy lifting. Section four lists 'Confirmed Decisions & Constraints,' including approaches tested and rejected. Section five lists 'Unresolved Questions & Open Blockers.' The key innovation here is the 'decision lock.' Instead of a bare 'no Prisma,' the document specifies 'Rejected Prisma due to cold-start latency in serverless workers.' This gives the next session the logic to respect the choice. It prevents the model from reopening the case unless a new requirement genuinely changes the underlying logic. Each line must carry a decision and its reason.

The Bootstrap Protocol and Practical Pitfalls

A handoff document is only half the transfer. You need a bootstrap prompt that tells the next session how to read it. The recommended protocol asks the model to confirm understanding in three concise bullet points before executing the first priority step. This serves as a cheap alignment check. If the model restates the state incorrectly, you catch it before it writes a single line of code. However, beware of serializing errors. If the previous thread contained a wrong assumption that was never corrected, the model will record that assumption as fact. Always read the handoff once before copying it to ensure the premise is clean.

Key Takeaways

  • Summaries record narrative; handoffs record state. Always use a structured handoff for multi-session projects.
  • Include 'decision locks' that pair rejected options with their specific technical reasons to prevent re-debate.
  • Use a bootstrap prompt that requires the AI to confirm understanding via three bullet points before executing.
  • Skip handoffs for short, stateless tasks like quick translations; they only add overhead without benefit.

The Bottom Line

If you cannot find the decision locks in your handoff, you are just pasting a history book into a compiler. Stop summarizing and start locking down the state.