handoff
Skill1 starsChanged 3 months agoMITView on GitHub
Compact the current conversation into a handoff document for another agent to pick up.
191 other projects use a version of this, 88 of them edited.
Adopted by
Most recently taken up by
- xiaoxiaosrm/dsh-mattpocock-skillsedited, first seen Nov 8, 2026
- ethosagent/ethostheir own version, first seen Nov 8, 2026
- dsifry/metaswarmtheir own version, first seen Nov 8, 2026
- glebis/claude-skillstheir own version, first seen Nov 8, 2026
- domengabrovsek/agent-configedited, first seen Nov 8, 2026
- domengabrovsek/claudeedited, first seen Nov 8, 2026
Used with
Projects with it also use decision-mapping, setup-ts-deep-modules and 9 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
FabienDehopreSkilldecision-mapping
Turn a loose idea into a sequenced map of investigation tickets, then drive them to resolution one at a time.
Also in 8 projects that use this
liufeng1976Skillsetup-ts-deep-modules
Wire dependency-cruiser into a TypeScript repo so each package is a deep module — implementation hidden in subfolders, reachable only through its entry-point files. User-invoked.
Also in 42 projects that use this
AWSSkillto-questionnaire
Turn a decision you can't fully answer into a questionnaire for someone else to fill in.
Also in 93 projects that use this
GoogleSkillask-matt
Ask which skill or flow fits your situation. A router over the skills in this repo.
Also in 87 projects that use this
RoentekSkillloop-me
Grill me about specs for the workflows I want to build, within this workspace.
Also in 51 projects that use this
LesaniSkillwriting-beats
Shape an article as a journey of beats, choose-your-own-adventure style. The user picks a starting beat from the raw material, you write only that beat, then offer options for where to pivot next, beat by beat, until the article reaches a natural end. Use when the user has raw material and wants to assemble it as a narrative rather than an argument.
Also in 54 projects that use this
Other versions191 projects use a version of this.
Word for word · 0
None found.
Edited · 88
- mattpocock/skills★ 309k51% same
- QuivrHQ/quivr★ 40k53% same
- The-Vibe-Company/quivr★ 40k53% same
- GreyDGL/PentestGPT★ 16k58% same
- ArcReel/ArcReel★ 6.3k51% same
- withcoral/coral★ 4.9k51% same
- bestofjs/bestofjs★ 3.1k54% same
- compozy/compozy★ 2.9k53% same
- bholmesdev/hubble.md★ 1.5k58% same
- CelestoAI/SmolVM★ 1.2k51% same
78 more not listed
Wrote their own version · 103
- ykdojo/claude-code-tips★ 10ksame name
- vinvcn/mattpocock-skills-zh-CN★ 5.5ksame name
- epicenter-md/epicenter★ 4.8ksame name
- EpicenterHQ/epicenter★ 4.8ksame name
- davidondrej/skills★ 4.3ksame name
- feiskyer/claude-code-settings★ 1.7ksame name
- skforecast/skforecast★ 1.5ksame name
- openclaw/agent-skills★ 1.1ksame name
- suboss87/FDEOps★ 940same name
- JordanCoin/codemap★ 710same name
93 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository39 other files it gives its agents.
- arch-mapSkill
- ask-meSkill
- autodevSkill
- codebase-designSkill
- code-reviewSkill
- context-barSkill
- contractSkill
- copilot-reviewSkill
- crapSkill
- deep-research-consolidateSkill
- deep-research-promptSkill
- domain-modelingSkill
- drop-reviewSkill
- fleet-reviewSkill
- framing-decisionsSkill
- grillingSkill
- grill-meSkill
- grill-with-docsSkill
- herdrSkill
- implementSkill
- mergedSkill
- meta-analysisSkill
- mutation-testSkill
- obsidian-vaultSkill
- paper-outlineSkill
- probeSkill
- prSkill
- setup-matt-pocock-skillsSkill
- show-meSkill
- slack-askSkill
- slack-notifySkill
- sparSkill
- sync-matt-pocock-skillsSkill
- tddSkill
- to-specSkill
- triageSkill
- writing-for-agentsSkill
- writing-git-commitsSkill
- writing-great-skillsSkill
The file
We split this skill into the rules it gives an agent
10 rules. 0 are common, 7 are distinctive, and 3 are very specific to this project.
0
Common
Most of the 172 edited versions we read keep these.
7
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 7
Write or load a handoff document so a fresh agent can continue the work after compaction.
- Very specific: line 22
If the user asks /handoff load (Claude Code) or $handoff load (Codex), resolve this session's identity exactly as the writer does: …
- Distinctive: line 26
Do not infer a Codex identity from a scratchpad path.
- Very specific: line 28
If that pointer is missing, do NOT silently fall back to ${AUTODEV_HOME:-$HOME/agents}/handoffs/.latest or to "the newest file in the …
- Distinctive: line 37
Every handoff MUST contain these sections, in this order.
- Distinctive: line 48
Do not duplicate content already captured in other artifacts (PRDs, plans, ADRs, issues, commits, diffs).
- Distinctive: line 50
If the user passed arguments, treat them as a description of what the next session will focus on and tailor the doc's emphasis accordingly …
- Very specific: line 66
Before writing, migrate old handoffs once if needed.
- Distinctive: line 80
When composing Standing items, cross-reference by path:
- Distinctive: line 85
After writing the handoff, update the deterministic reload pointer, then file a compact-request so the auto-handoff watcher (if installed) …
“Common” is measured: kept by at least 6 in 10 of the 172 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Write or load a handoff document so a fresh agent can continue the work after compaction.
Explicit invocation is runtime-specific: Claude Code uses /handoff; Codex uses $handoff.
AUTODEV_HOME defaults to ~/agents; every agents path below honors it.
Save handoffs into ${AUTODEV_HOME:-$HOME/agents}/handoffs/ - not the current workspace and not the OS temp
directory. Create the directory if it does not already exist. Name new files:
$(date +"%Y.%m.%d %H.%M.%S") handoff-<short-kebab-topic>.md.
Files in ${AUTODEV_HOME:-$HOME/agents}/handoffs/ older than one month are auto-pruned by an external monthly
cron job. If the cron still targets $HOME/handoffs/, update it outside this skill.
Load
If the user asks /handoff load (Claude Code) or $handoff load (Codex), resolve this session's
identity exactly as the writer does: ${CLAUDE_CODE_SESSION_ID} for Claude Code, or
${CODEX_SESSION_ID:-${CODEX_THREAD_ID:-}} for Codex. Read
${AUTODEV_HOME:-$HOME/agents}/handoffs/.latest.<resolved-session-id>, then read and present that handoff document.
Do not infer a Codex identity from a scratchpad path.
If that pointer is missing, do NOT silently fall back to ${AUTODEV_HOME:-$HOME/agents}/handoffs/.latest or to
"the newest file in the directory". Both are machine-wide and last-writer-wins: on a box with
several concurrent sessions they usually name somebody else's mission. Instead, list the newest
timestamped handoffs and ask the user which one to load.
Required sections
The first line under the document title is Context at handoff: <n>% (the $pct from the Write Workflow's first step, or unknown).
Every handoff MUST contain these sections, in this order. Do not drop any of them even for a "small" handoff — scale each to the work, but cover all.
- Broad context — orient a reader with ZERO prior context: what this work is, why it matters, where it sits in the larger effort, and the current state (repo, branch, tip commit, what's shipped, what's green, what's pushed). Give the through-line of the whole arc, not just the latest task. Reference artifacts by path/URL for detail rather than duplicating them.
- Standing items — every open thread with its state made EXPLICIT (e.g. DONE / DEFERRED-with-named-closer / BLOCKED-on-X / AWAITING-USER). Include what's waiting on the user specifically (pushes, approvals, rebases, decisions) and each blocker's unblock condition.
- Next phases / steps — the concrete sequence of work the next session should pick up, in order, with any ordering constraints or dependencies between steps spelled out ("do X before Y because …").
- Recommendation — your explicit, opinionated recommendation for how to tackle the next steps: which item to start with and why, the approach you would take, the traps to avoid, and any sequencing/leverage judgment. Make the call you would make — this is your judgment, not a neutral menu of options.
- Insights from this session — the non-obvious things learned that are NOT captured in code or commits: inverted premises, false-positives found, decisions and their why, antipatterns avoided, dead ends not worth re-treading, and any discipline/meta-lessons. These are the most perishable and often the most valuable part of the handoff — record them so the next session does not rediscover them the hard way. Write every one of them down here; the only question is whether a copy also belongs somewhere permanent. Handoffs are pruned after a month, so an insight that would still be true long after this work ends outlives its handoff: if the project has a durable home for it — a PM repo's
INSIGHTS.md, an ADR, the repo's own docs — put it there too and cite that path here. Where no such home exists, this section is the home. - Suggested skills — skills the next agent should invoke (e.g. brainstorming, writing-plans, subagent-driven-development), each with a one-line reason.
Rules
- Do not duplicate content already captured in other artifacts (PRDs, plans, ADRs, issues, commits, diffs). Reference them by path or URL instead — but DO synthesise the through-line and the cross-artifact state that no single document captures.
- Redact any sensitive information, such as API keys, passwords, or personally identifiable information.
- If the user passed arguments, treat them as a description of what the next session will focus on and tailor the doc's emphasis accordingly — but still include ALL required sections above; the arguments shape emphasis, not coverage.
Write Workflow
First, before anything else, print the context % to the user. This is the first visible output of the handoff turn, and it runs for manual and watcher-triggered handoffs alike. Resolve the session id as elsewhere in this file and read the reading the statusline (Claude Code) or Stop hook (Codex) wrote; never estimate it:
sid="${CLAUDE_CODE_SESSION_ID:-${CODEX_SESSION_ID:-${CODEX_THREAD_ID:-}}}"
pct=$(sed -n 's/^pct=\([0-9.]*\).*/\1/p' "${AUTODEV_HOME:-$HOME/agents}/state/$sid.ctx" 2>/dev/null)
[ -n "$sid" ] && [ -n "$pct" ] && echo "Context at ${pct}% — writing handoff." || echo "Context: unknown (no .ctx reading)"Show that line to the user as text, then continue. Keep $pct (or "unknown") for the document header.
Before writing, migrate old handoffs once if needed. Preserve the ~/agents/ timestamp naming
convention by prefixing each migrated file with its original modification time:
mkdir -p "${AUTODEV_HOME:-$HOME/agents}/handoffs"
for f in "$HOME"/handoffs/*.md; do
[ -e "$f" ] || continue
epoch=$(stat -c %Y "$f" 2>/dev/null || stat -f %m "$f")
ts=$(date -d "@$epoch" +"%Y.%m.%d %H.%M.%S" 2>/dev/null || date -r "$epoch" +"%Y.%m.%d %H.%M.%S")
base=$(basename "$f")
mv "$f" "${AUTODEV_HOME:-$HOME/agents}/handoffs/$ts $base"
doneWhen composing Standing items, cross-reference by path:
- latest sparring artifacts under
~/agents/adversarial/{slug}/, including.session-id,sparring-log.md, and the latest round Q/A if present;
After writing the handoff, update the deterministic reload pointer, then file a compact-request so the auto-handoff watcher (if installed) finishes the cycle:
hf="<full handoff path>"
# .latest is a SHARED, machine-wide, last-writer-wins pointer. It is for humans
# ("what handed off most recently on this box"), and it is NOT what your session
# reloads from — with several sessions running concurrently, whoever writes last
# wins and everyone else would resume the wrong mission.
tmp="${AUTODEV_HOME:-$HOME/agents}/handoffs/.latest.tmp"
printf '%s\n' "$hf" > "$tmp" && mv "$tmp" "${AUTODEV_HOME:-$HOME/agents}/handoffs/.latest"
# This is the one that matters. --handoff records the PER-SESSION reload pointer
# (.latest.<sid>) naming the file you just wrote, which is the only thing the
# watcher and the SessionStart hook will read back after compaction. Pass it
# ALWAYS: without it the helper can only copy the shared .latest, which another
# session may already have clobbered.
#
# It also files the compact-request that makes the watcher run /compact + reload
# even below the 35% threshold. No-op when autodev is not installed; defers on
# its own when the watcher is already mid-cycle.
rh=""
for root in "$HOME/.agents/skills" "$HOME/.claude/skills" "$HOME/.codex/skills"; do
[ -x "$root/autodev/bin/request-handoff.sh" ] && { rh="$root/autodev/bin/request-handoff.sh"; break; }
done
[ -x "$rh" ] && bash "$rh" --compact-only --handoff "$hf" 2>/dev/null || trueALWAYS print the handoff's absolute path to the user — every time, no exceptions. The full
expanded path exactly as $hf holds it: not ${AUTODEV_HOME:-$HOME/agents}/handoffs/…, not a bare filename, not "the
handoff file". The user has to open, move, or hand that file to another session, and a path they
must reconstruct is a path they cannot click. This holds when the write was routine, when a tool
call already showed the filename, and when nothing else in the turn is worth saying. Print the
context % alongside it using the same $pct reading from the first step, never a fresh estimate.
Then VERIFY the trigger exists — do not assume the script filed one. request-handoff.sh
legitimately declines to file when the watcher is already mid-cycle, and it is a silent no-op
when autodev is not installed. Running it is therefore not proof that anything will happen:
sid="${CLAUDE_CODE_SESSION_ID:-${CODEX_SESSION_ID:-${CODEX_THREAD_ID:-}}}"
ls -la "${AUTODEV_HOME:-$HOME/agents}/state/$sid".{compact-request,handoff-request} 2>/dev/null
ls -la "${AUTODEV_HOME:-$HOME/agents}/handoffs/.latest.$sid" 2>/dev/null # the reload pointer
cat "${AUTODEV_HOME:-$HOME/agents}/state/$sid.ctx" 2>/dev/null # authoritative pct — never guess itResolve as follows:
- A marker file exists → the cycle will run. Emit the block below and end the turn.
- No marker, but the script said it was already mid-cycle → fine, the in-flight cycle will
compact. This is the normal outcome when the watcher itself invoked the handoff skill you are
answering: its cycle lock is live for the whole turn. Say so explicitly in your message so the
user can see why nothing was filed, and check
.latest.$sidbelow — the pointer is recorded even on this path, so if it is missing something else is wrong. - No marker and no mid-cycle → retry the script; if it still files nothing, tell the user
plainly that automation is not attached and ask them to run
/compact. Never end the turn silently here. .latest.$sidis missing → the reload after compaction will deliberately fail closed (you will be told to re-orient from your own transcript rather than handed a handoff). Re-run the--handoffcommand above so the pointer exists.
The pct threshold does not fire on its own from an idle session: the Stop-hook watcher evaluates
once per turn end, and an idle session's pct never rises. auto-handoff-sweep.sh (systemd timer,
every few minutes) is the safety net that re-evaluates parked sessions — but it is a backstop, not
a substitute for filing and verifying the marker. Do both.
Then emit this explicit instruction block:
Handoff written: <absolute path — the full expanded $hf, always, no exceptions>
Context at handoff: <pct>% (or "unknown" when no .ctx reading exists)
.latest updated, compact-request filed.
• End the turn only when the marker and this session's registered pane were verified above. The
watcher will run /compact and reload at the next idle Stop; the marker is what makes this fire
below the 35% threshold. It cannot act while a background agent or turn is still running because
input would be queued. "Do nothing" is correct only because the state was verified.
• Claude Code also shows the 🔴 AUTO-HANDOFF badge when attached. Codex uses its native status
line and has no custom badge, so its marker, `.ctx`, `.runtime`, and registered pane files are
the authority. If the required state is absent, run /compact yourself now.
Reload contract: after compaction, the next agent turn is handed THIS session's handoff to read,
via the SessionStart hook reading ${AUTODEV_HOME:-$HOME/agents}/handoffs/.latest.<resolved-session-id> (using
CLAUDE_CODE_SESSION_ID, or CODEX_SESSION_ID with CODEX_THREAD_ID as its fallback). If that pointer
is missing the reload fails closed on purpose — it will never hand you the shared ${AUTODEV_HOME:-$HOME/agents}/handoffs/.latest,
because that file is machine-wide and would resume another session's mission.The agent cannot self-trigger the interactive /compact command — only the user or the
state-verified watcher can. On Codex, the watcher confirms the command only when the pane visibly
contains an explicit compact/summarize confirmation prompt.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

