The mistake I made for months
I assumed a good plan file was a handoff. Write the mission, the task list and the status log, and any new session could pick it up and carry on.
It half worked. A new session would follow the letter of the plan and drift anyway, because the file carried the conclusions without the argument behind them. Tasks came back over-engineered, or with a scope that had already been agreed quietly widened, or as a file of several thousand lines nobody asked for. Nothing in the plan forbade any of it. The reason not to had only ever existed in a conversation that was not in the new session's context.
I built a skill to generate those files automatically. It helped at small scale and stopped helping as the work grew. I do not use it now.
What works better
Stay in the conversation that did the thinking. Work the problem until there is a conclusion, ask for the plan back in its own words so I can check the readback against what I meant, then work through it a step at a time.
The reasoning stays in the context that way, rather than being compressed into a file and then partially reconstructed from it. Even a well-written plan is a summary, and summaries drop the parts that seemed obvious at the time.
The other benefit is mine rather than the tool's. Having been there while the decisions were made, I notice sooner when something starts going wrong.
Handoffs are for when there is no choice
Sessions end and work changes hands. When it does, I want something short about the present. What is done, what is half done, what the next step depends on, and what is still open.
Open questions have to stay visibly open. If one drops off the list, somebody will make a perfectly sensible choice without knowing it needed approval, and that choice will be found much later by someone wondering why the data looks like that.
The opposite failure is carrying too much. A long conversation accumulates abandoned ideas and instructions that were overruled two hours ago, and all of it still counts as input. A handoff describes now. The product notes are there for anyone who needs to know why.
Further reading
- Chroma: Context Rot, on how reliability degrades as input length grows.