Session Handoff — Context Compaction Summary

Compresses a long working session into a handoff document a fresh agent can continue from — decisions with their reasons, failed approaches preserved, exact next commands.

Fill in the blanks

Prompt variables

2 variables detected — fill them and Copy inserts your values (0/2 filled).

Summarize this working session into a handoff document. The reader is a fresh agent (or you, tomorrow) with ZERO memory of this session. The test of quality: the reader continues the work without re-discovering anything — especially without re-trying things that already failed. Structure — all sections mandatory: ## Goal The overall objective in 1–2 sentences, plus the acceptance criteria as currently understood (they may have evolved during the session — state the CURRENT version, and note if it drifted from the original ask). ## State: done / in progress / not started - DONE: each completed piece, one line, with how it was verified (test run, manual check) — unverified work is IN PROGRESS, not done. - IN PROGRESS: exactly where work stopped, mid-what. Which file was being edited, what the half-finished idea was. - NOT STARTED: the remaining known work. ## Decisions made, with reasons Every non-obvious choice: "chose X over Y because Z". The reason matters more than the decision — without it, the next agent re-litigates or, worse, silently reverses it. ## Failed approaches — DO NOT RETRY The highest-value section. Each dead end: what was tried, the exact error or reason it failed, and what that rules out. A handoff that omits failures condemns the next agent to repeat them. ## Live gotchas Environment quirks discovered (the flaky test, the service that needs a restart after config changes, the misleading error message and what it actually means), plus any assumptions the current work rests on that were never verified. ## Next actions Numbered, concrete, in order. First entry must be executable immediately with zero investigation: the exact command to run or file+line to open. Include the verification command(s) for checking work as it proceeds. ## Open questions Things needing a human decision or missing information — phrased as answerable questions, each with its blocking effect noted. Rules: facts from the session only — no optimistic smoothing; exact identifiers (paths, commands, error strings) verbatim, never paraphrased; total length under {{max_length}}, spent proportionally to usefulness (failed approaches and next actions earn their space, narrative history does not). Session log: {{session_log}}
Field notes

Comments (0)

LLM Atlas

A free map of prompts that actually work — tested on real models, starred by real people.

© 2026 LLM AtlasPrompts default to CC0 unless stated.