This content originally appeared on DEV Community and was authored by DaslavCl
TL;DR: As your conversation gets long, answers degrade before the hard limit (30% of context limit is the degradation spot for me). Keep two living docs the model can refresh on demand: README.md (holistic view) and HANDOFF.md (everything a fresh instance needs to continue seamlessly).
*When to trigger it
*
You notice omissions/contradictions, weird latencies, or invented paths/versions. Don’t wait for a hard token error.
*What to maintain
*
README.md: purpose/scope, quick arch note, stack & versions, common commands, recent decisions.
HANDOFF.md: current status, open issues + next steps, paths/artifacts, latest test results, data/IO schemas, exact env (venv/conda/poetry) and package versions.
One-shot prompt to refresh both
“”” Please update two files based on our recent work and decisions.
README.md — keep a holistic, up-to-date view:
purpose/scope, brief architecture, stack with exact versions,
how to run (commands, seeds), recent decisions (changelog-lite).
HANDOFF.md — for a fresh instance (who will take this conversation when we hit the context limit) to continue smoothly: (remember the new instance has not context about our work or previous conversation) Please add (examples)
current status, open challenges, next steps,
paths/artifacts/datasets, recent test results + logs,
schemas/contracts and expected outputs,
exact environment (venv/conda/poetry), package versions, and notes to avoid creating duplicate envs.
Use the versions/configs we’ve used so far. Do not invent tools or upgrade versions unless I ask. “””
*Why this helps
*
Mitigates “context drift” long before you hit limits. Makes instance swaps (or model switches) painless. Turns your chat into project memory rather than ephemeral Q&A.
If you’ve got a tighter checklist or a better trigger than my “degradation symptoms,” please share.
This content originally appeared on DEV Community and was authored by DaslavCl