TsukiNi22/ai-utils / skills/ai/export
export
UnverifiedExport everything needed to hand the work over to another machine, session or model into one compact file (PASSATION.md by default) - the goal and context, what is done, what runs, what is left (with the exact next step), the decisions and constraints of the user, the files and branches touched, the commands / tools / scripts used and how to build, test, run, the skills to load and the MCP servers / tools to set up, the state of the environment, the traps met - written for an AI reader in the rtk way (terse facts, no prose, no token wasted), never a secret - plus the ready-to-paste prompt that makes the new session pick the work up, saying whether the file must be sent with it (not when the new session works in the folder holding it). Use whenever the user asks to export / hand over / pass the work on ("export", "passation", "handoff", "fais une passation", "fais un md de passation", "exporte le contexte", "un prompt pour reprendre", "pour une autre session / machine / IA"), or summons /export.
Export: a handover file for another session / machine / model
Stop and resume in the same project, at the next continue: the pause skill, not this one.
Write one file another AI reads to continue the work without this conversation: what is wanted, what is done,
what is left and exactly how to go on. Default PASSATION.md at the root of the project (the cwd); another name or
place only when the user says so. The reader is an AI: apply rtk (terse facts, one per line, no prose, no
decoration, nothing said twice) - complete, not pretty.
1. Gather (read-only)
- The conversation: the requests of the user (their words for the important ones), their answers to questions, their preferences and refusals, what was decided and why.
- The task lists (TaskList / todos) and their states; the background work still running (jobs, shells, builds, tests) and how to check it.
- The repository:
git status/ branch / last commits / unpushed work / worktrees / stashes, the files changed and not committed (git diff --stat), generated or temp files that matter (scripts kept in a temp folder: copy their path, or their content when the other machine will not have them). - The environment: OS, the tools and versions that matter, how to build / test / run / install, the services or processes in use (never stop them), the settings / env vars changed.
- The memory and the project instructions (CLAUDE.md, rules the user gave) that the next one must follow.
- What the next one needs loaded: the skills used and to load first (by name), the MCP servers relied on (
claude mcp list,.mcp.json: name, whether the new session has them, how to add a missing one), the CLI tools and their versions, the permissions / modes that mattered (auto mode, allowed tools), the subagents / background jobs pattern used.
2. Write it (sections in this order, empty ones dropped)
# PASSATION <project> <date> (<from: model / machine>)
## goal the request(s) in one or two lines each; the user's own words when they matter
## rules constraints of the user: git (commit / push allowed?), language, style, what never to do
## state = done N · running N · todo N · blocked N
## done - <what> (<where: file / commit / version>), checked how
## running - <what> · check: <command> · expected: <result>
## todo - [P1] <what> · next step: <exact action> · files: <paths> (ordered: what to do first)
## blocked - <what> · why · what unblocks it
## decisions - <choice> · why · rejected: <alternative>
## files $P=<common dir>; - $P/a.cpp: <what changed, uncommitted?>
## repo branch · last commit · uncommitted / unpushed · worktree · how the release / commit is done
## env os · tools (versions) · build: <cmd> · test: <cmd> · run: <cmd> · install: <cmd>
## tools the scripts / helpers used and for what (path + one line)
## load skills to load first: <names> · MCP needed: <name> (present? add: <cmd>) · CLI: <tool version> · mode: <auto / plan...>
## traps - <what went wrong> → <how it is avoided> (flaky tests, quotas, guards, aliases...)
## refs docs, issues, URLs, paths of logs / screenshots / reports
- Every item actionable: a path, a command, an id, a version, an exact error text - never "see above" or "as
discussed". Relative paths under a
$Pwritten once. todois the core: ordered, each with its next step precise enough to start without asking.- What the next one must not do (destructive commands, processes to keep, settings not to change) goes in
rules. - Never a secret: no token, password, key, cookie, private URL with credentials, personal data beyond what the
project needs; write where it lives instead (
keyring: claude-cluster,env: API_KEY). - Long logs / diffs: never pasted; their path, or the 1-3 lines that matter.
3. The prompt for the new session
A short prompt the user pastes into the new session (a code block, in the language the user writes to it, French by default here), that makes it pick the work up at once:
- The file reachable from there (the new session works in the project folder, or in claude-cluster on that project, or
on this machine with access to the path): no need to send it - the prompt says to read it first, by its absolute path:
lis <path>/PASSATION.md entièrement, charge les skills et vérifie les MCP listés dans "load", puis reprends au premier todo (P1) sans refaire ce qui est "done" ; respecte "rules". - Not reachable (another machine, claude.ai, a model without this folder, another project): say plainly that the file must be sent with the prompt (attached, or pasted after it), and write the prompt so it refers to "the file attached". Nothing reachable at all and the file short: offer the file content inlined in the prompt instead.
- Add the one or two things the next one must do first (the exact next step of the top todo, a check of what still runs)
and what it must never do (from
rules); never repeat the whole file.
4. Check and report
- Read the file as the next AI: can it continue alone (what, where, how, in which order, what not to touch)? Fill the holes; remove what it does not need.
- Measure (
wc -c, ~4 chars per token) and keep it short: a big context compressed, not copied. - Tell the user the path, the size and the counts in one or two lines, then the prompt (its code block) and whether the file must be sent with it; not committed unless they ask (or the project's rules allow it and they asked for a commit).
- With claude-cluster running and the user wanting it: the new session can be started with that prompt directly (a new session in the project folder) - offered, never done without their go.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 464k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 464k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 464k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 192k | Updated | Claude CodeCodex | Verified | Install |