End a working session like a shift change, not an abandonment: state captured, decisions logged, next step named — ready for the next session to pick up.
Working sessions with AI end badly by default: the context evaporates with the tab, and the next session starts by re-explaining. A session handoff treats the boundary like a shift change — the package captures what was completed, what was decided (and with whom it was agreed), what constraints bind, what remains, and exactly where to resume. This setup loads a refactoring session mid-flight, with the next step explicitly stated: the new session's first act is continuing the extraction, not rediscovering it.
Close with a capture
The last act of a session is building its handoff — state is freshest at the boundary.
Name the resume point
The Recommended Next Step field turns "where were we?" into "do this first".
Resume without warmup
The next session reads the package and continues — the re-explaining tax is gone.
"Fidelity: Standard" marks how much detail the package preserves — a balanced capture that keeps the key decisions, constraints, and open questions without transcribing the whole conversation. It sits on the first line so the receiving session knows the depth of context it's inheriting. If you need more or less granularity, you regenerate the handoff at a different fidelity in your own tool before pasting it.
The HANDOFF INSTRUCTIONS block tells the receiving session to accept everything above as established context, preserve the key decisions, keep tracking the open questions, and continue from the recommended next step — a continuation, not a fresh start. Pasting it as the first message of a new chat is what carries that instruction across; the model has no memory of the prior session on its own.
Read the KEY DECISIONS, CONSTRAINTS, OPEN QUESTIONS, and PENDING TASKS sections against what actually happened — the NOTE flags that items were pulled verbatim and deduplicated, not paraphrased, so they can be stale or over-broad. Fix anything wrong and confirm the RECOMMENDED NEXT STEP is truly where you want the next session to resume. The package is a draft you review, not an authoritative record.
NewPrompt builds the handoff text; you paste it as the first message of a new chat yourself — it isn't run or delivered for you. Because it's plain text, it works in ChatGPT, Claude, Gemini, or any chat model. The HANDOFF INSTRUCTIONS travel with it and tell that model to continue from the recommended next step rather than restart.
When the window fills, don't retype the project from memory. A handoff package carries the decisions, constraints, and pending work into the new session as its first message.
Done work tells stories; open work needs carrying. The package separates completed from pending and resumes from the first open task.
Creative collaborations build invisible state: the voice you settled on, the structures you rejected. Carry it forward so the next session writes like the last one.
Token budget planning for real workloads: how much of the window a transcript actually consumes, what is left for the answer, and how much headroom remains.
The "message too long" error has a structural fix: split at paragraph boundaries into sequenced chunks with wait rules, instead of pasting fragments and hoping.
What a converged conversation contains: a rejected voice, an accepted framing, a hard constraint, a format decision, and a final-version moment — each mapped to its prompt section.
Continue work in a new chat without losing context — decisions, constraints, and tasks extracted into a state package.