THE HANDBACK CARD · the handback · the decision first, every claim with its evidence
WAKES WHEN an agent stops, gets blocked, runs out of budget or finishes and a person takes over, or a handoff note or escalation summary is too vague to act on
THE METHOD
template --out HANDBACK.md — the shape, every part a [FILL: …] marker.
write run-state.json --out HANDBACK.md — a draft from the run, inventing nothing.
check HANDBACK.md — the score, the verdict, every gap with its fix, in order.
A person fills the markers and closes the gaps; check again until READY.
REFUSES
Writing evidence the run does not hold
Calling a hope evidence
Inventing the decision's options, recommendation or deadline
Calling a note READY with markers open
IT MAKES template — the handback shape, decision first, every part a [FILL] marker; check — score a handback note (.md or .json) and name every gap; write — draft a handback from an agent's run state (JSON), inventing nothing.
THE LAW The score comes from the note on file, part by part, through the machine. A hope is not evidence. A claim of work done with nothing to check scores nothing. Unfinished work asks for one decision, at the top. A status that is not one of the five is unfinished work. Every change outside the files is a risk to name, with whether it can be undone. Nothing is invented: what the run does not say is a [FILL: …] marker, and a note with an open marker is never READY. It reads shape and words, not truth: a person still reads the note before acting on it.
IN THE BOX scripts/handback.py · the shape · the method by hand · the sources note · ten tests.
PROOF selftest 22/22; a vague note scored 4, a draft from the run state, and the filled note at 100, with the trigger test, on the record (WITNESS.md) · witnessed 2026-10-13 · EDITION 1.
THE SEAL 96 of 100 · LEGENDARY · WITNESSED · Trigger 5 · Machinery 5 · Law 5 · Portability 5 · Proof 4
IT CAN BE WRONG an AI reading a text file: it can be wrong and can invent a number, a rule, a name or a date; check what matters against your own records and the source; the risk is yours. LICENSE-NOTE.md rides in the packet and is the licence.
THE PUNCH · SHA-256 OF SKILL.MD · 5BA434971E75CB216938D2B8
CITE THIS CARDTHE HANDBACK CARD, Edition 1. Lund Studio LLC, Westminster, Colorado. https://lundr.tech/?view=handback-card. Punch 5BA434971E75CB216938D2B8. Retrieved 2026-10-13.
What it is.
The note an AI agent leaves when work goes back to a person: when it finishes, stops, gets blocked, runs out of budget, or finds it unsafe to go on. A good handback lets the person act without asking anything. The card holds it to one shape:
Use it whenever an agent stops, gets blocked, runs out of budget or finishes and a person takes over, or a handoff note or escalation summary is too vague to act on. Not for a ledger a fresh agent resumes from (THE CHECKPOINT), agent-to-agent hand-offs (THE ENVELOPE CARD), an audit receipt of a run (THE RUN RECEIPT).
It carries a machine that shows its working: template — the handback shape, decision first, every part a [FILL] marker; check — score a handback note (.md or .json) and name every gap; write — draft a handback from an agent's run state (JSON), inventing nothing — 22 checks in its selftest.
It answers requests such as “check this handoff note before I send it to the on-call lead”, “the agent stopped halfway: write the handback from its run log”, “is this escalation summary good enough to act on?”, “what should an agent's end-of-run report to a person contain?”.
Six laws, written into the card.
They are in the card itself, so the AI follows them.
THE SCORE COMES FROM THE NOTE ON FILEThe score comes from the note on file, part by part, through the machine.
A HOPE IS NOT EVIDENCEA hope is not evidence. A claim of work done with nothing to check scores nothing.
UNFINISHED WORK ASKS FOR ONE DECISIONUnfinished work asks for one decision, at the top. A status that is not one of the five is unfinished work.
EVERY CHANGE OUTSIDE THE FILES IS A RISK TO NAMEEvery change outside the files is a risk to name, with whether it can be undone.
NOTHING IS INVENTEDNothing is invented: what the run does not say is a [FILL: …] marker, and a note with an open marker is never READY.
IT READS SHAPE AND WORDSIt reads shape and words, not truth: a person still reads the note before acting on it.
The method.
template --out HANDBACK.md — the shape, every part a [FILL: …] marker.
write run-state.json --out HANDBACK.md — a draft from the run, inventing nothing.
check HANDBACK.md — the score, the verdict, every gap with its fix, in order.
A person fills the markers and closes the gaps; check again until READY.
What rides in the card.
In the card’s own voice, with a machine that shows its working.
THE REFERENCESreferences/SOURCES.md · THE-METHOD.md · THE-SHAPE.md
THE MACHINEscripts/handback.py · three commandstemplate — the handback shape, decision first, every part a [FILL] marker; check — score a handback note (.md or .json) and name every gap; write — draft a handback from an agent's run state (JSON), inventing nothing. 22 checks in its selftest. Standard library only; nothing opens a network connection.
What it does once your AI has it.
template The handback shape, decision first, every part a [FILL] marker
check Score a handback note (.md or .json) and name every gap
write Draft a handback from an agent's run state (JSON), inventing nothing
On the record An Example Co clean-up agent's handback taken from 4 to 100: its own note, a draft from the run state with two markers, and the filled note; 44 of 48 in the trigger test
What it refuses, and where it hands off.
It refuses writing evidence the run does not hold; calling a hope evidence; inventing the decision's options, recommendation or deadline; calling a note READY with markers open.
a ledger a fresh agent resumes from after a context reset → THE CHECKPOINT
work handed from one agent to another → THE ENVELOPE CARD
an audit receipt of what a run read, changed and sent → THE RUN RECEIPT
whether the agent may take an action at all → THE AUTONOMY LADDER
Reads the note's shape and words, not whether its evidence is true; a person still reads it before acting. It is an AI reading a text file: it can be wrong, so check what matters against your own records and the source.
What’s in the buy.
CARD.md — the whole card: its boot, law, method, machine, seals, refusals and hand-offs, ledger, score and what it carries · SKILL.md — the door your AI reads first · START-HERE.md — the first five minutes, for you
SOURCES.md + THE-METHOD.md + THE-SHAPE.md — in references/, in the card's own voice
scripts/handback.py — the machine: template · check · write; standard library only, no network; selftest 22/22
WITNESS.md — the run on a realistic case as printed, and the trigger test on Haiku, Sonnet and Opus · SELFTEST.md — the machine's own test, as run at the cut · AUDIT.md, the twelve checks · MANIFEST.md, every file hashed
tests/evals.json — the cases it must take, hand off and refuse
MY-LEDGER.md — the card's memory, on your own machine · TIN.md, the nine lines on the tin · LICENSE-NOTE.md · HOSTS.md, where it runs
system/short-prompt.md and system/system-prompt.md — the doors for any AI without skills
PROOF · SELFTEST 22/22 · WITNESSED ON A RUN · TRIGGER TEST 44 OF 48 ON ITS OWN PROMPTS · HAIKU 12/16 · SONNET 16/16 · OPUS 16/16 · 0 WRONG PICKS · EDITION 1 · PUNCH 5BA434971E75CB216938D2B8
ON EVERY DOOR · A CARD FOR YOUR AI, A FILE FOR YOUR COMPUTER
In Claude: turn on code execution under Settings › Capabilities, then Customize › Skills › + › Create skill, and upload the zip as it is — START-HERE.md in the box walks you through it, and tells you how to keep the ledger. In ChatGPT, Gemini and the rest: a Project, a GPT or a Gem with the card’s files attached and the short door in the box pasted as its instructions. The words run anywhere an AI reads Markdown; the machines run wherever the host runs Python. Witnessed in Claude only. The host marks →
Take it.
Nº 973 — LAUNCHING SOON
Sold as is under the term printed. A card instructs an AI you operate; results depend on your model, your files and your judgement. Not medical, legal, tax or financial advice, never the person’s boss, and not a licence to practise. The terms →
One payment. The file is yours to keep and keeps working at its edition; new editions issued in the year after you buy reach you by email as replacement files. Tax is added at checkout where it applies. Nothing auto-renews.
A TERM, NOT A SUBSCRIPTION: NOTHING RENEWS ITSELF AND NOBODY IS CHARGED AGAIN. WHEN A TERM ENDS THE FILES STAY ON YOUR COMPUTER AND STAY READABLE — FREEZE, NOT LOSS.
THE MACHINE · IN THE OPEN · FROM THE CARD’S OWN WITNESS.MD AND SELFTEST.MD
handback.py, every command showing its working.
A small Python program. Your AI runs it; you can too. The blocks below are pasted from the card’s own record, as printed.
template
the handback shape, decision first, every part a [FILL] marker
check
score a handback note (.md or .json) and name every gap
write
draft a handback from an agent's run state (JSON), inventing nothing
selftest
22 checks that prove the machine; it must print 22/22
HANDBACK · note-from-the-agent.md · status 'mostly done', not one of the five
SCORE · 4/100 · NOT READY
PARTS · head 4/10 · done 0/25 · not done 0/10 · tried 0/10 · decision 0/25 · risks 0/10 · resume 0/10
GAPS · what to fix, in order
head status is 'mostly done': use finished, blocked, stopped, out of budget or unsafe to continue
head no reader: write `for:` and the person or role who takes over
done no Done section: list what was done, each with its evidence
not done no Not done section: list what was not done and why, or write none
tried no Tried section: say what was tried that did not work, and what happened, or write none
decision the work is not finished but no decision is asked: write the one question, two options or more, a recommendation and by when
risks no Risks section: name what could go wrong, what was changed outside the files, or write none known
resume no Resume section: say the next step, where everything is, and the command or state to continue
RULE · score = head 10 + done-with-evidence 25 (in proportion) + not done 10 + tried 10 + the decision 25 (question 6, two options 6, recommendation 5, by when 5, placed first 3) + risks 10 + resume 10; READY at 85 or more, THIN at 60 or more
SOURCE · the card's handback shape (references/THE-SHAPE.md): six sections and a head, the card's own standard
LIMIT · it reads the note's shape and words. It cannot tell whether the evidence is true, only whether it is there and checkable; a person still reads the note before acting on it.
wrote DRAFT.md · draft scores 93/100 THIN · 2 [FILL: …] marker(s) to fill from the person or the run's record
RULE · nothing the run state does not say is written: every gap is a [FILL: …] marker, never a guess
# HANDBACK · merge duplicate customer records in the CRM export
status: blocked
for: [FILL: who takes over, by role]
date: 2026-10-13
## Decision needed
question: For the 34 groups with two live logins, keep both accounts or merge into the newest?
options:
- A: keep both and link them as one household: no login lost, duplicates stay in reports
- B: merge into the newest account and email the other address first: one record, one login removed
recommend: A: nothing is lost and it can be merged later; B cannot be undone once the old login is removed
by: [FILL: a date, or what happens if no answer]
## Done
- found duplicate groups by email and phone — evidence: dedupe-report.csv: 1,214 groups, 2,611 records
- merged groups with identical name, email and phone — evidence: merge-log.jsonl: 1,180 merges, 0 errors
- sent the merge summary to the sales team list — evidence: message id 4471 in the outbox
## Not done
- 34 groups with different emails — why: each holds two live logins; merging drops one
## Tried
- merged 34 groups whose emails differ — result: stopped: two live accounts per group; a merge would drop one login
## Risks
- the merge summary was emailed to the sales team list — done outside the files: undoing the edits does not undo it
## Resume
next: answer the decision; the agent then runs the 34 groups
where: work/dedupe-report.csv, work/merge-log.jsonl, work/blocked-34.csv
command: python3 merge.py --groups work/blocked-34.csv --mode household
THE SELFTEST, AS PRINTED
PASS a complete blocked handback scores 100/100, READY, no gaps
PASS the empty template is NOT READY and its open [FILL] markers are counted
PASS a done item with no evidence is named, and done scores 2 of 3 (17/25)
PASS 'should work' is a hope, not evidence
PASS a blocked note with no decision loses all 25
PASS a status that is not one of the five ('mostly done') is not finished: no decision and no Tried score nothing
PASS one option is not a choice
PASS a decision buried at the bottom loses 3 points
PASS a finished note with nothing left needs no decision and scores 100
PASS a status outside the five is named
PASS a side effect with no risk named is flagged
PASS a resume with no place loses 3
PASS section names in other words are read: What I tried, Next steps, Left to do
PASS a note over 500 words is warned, not failed
PASS check from an empty folder: READY, exit 0, the same bytes twice
PASS the JSON form of a complete note also scores 100
PASS write maps the run state: good steps to Done, failed steps to Tried, the to-do list to Not done
PASS a note with an open [FILL] marker is never READY, whatever its score (98)
PASS write invents nothing: missing evidence, the question and the command are [FILL] markers; side effects become risks
PASS the draft's [FILL] markers count as missing until filled
PASS unreadable input exits 2 with a message
PASS template then check: NOT READY, exit 1
selftest 22/22
THE TRIGGER TEST. The card's line was tried on Haiku, Sonnet and Opus beside 16 neighbouring skills, in a clean environment: 44/48 on its own prompts (Haiku 12/16 · Sonnet 16/16 · Opus 16/16), 0 wrong picks; no skill (answered directly) 4.
THE RUNS ARE PASTED FROM THE CARD’S OWN RECORD AS PRINTED. THE CASES ARE THE STUDIO’S OWN EXAMPLES (EXAMPLE CO IS FICTION); NO PRIVATE PERSON IS NAMED.
QUESTIONS AT THE TABLE
Asked before the wire.
What does it need from me?
Python 3; the note (Markdown or JSON) or the agent's run state (JSON).
Does it send anything anywhere?
No. The machine is a standard-library Python program that runs on your own computer and sends nothing anywhere; the words run wherever an AI reads Markdown. No server, no account, no key.
What is on the record?
The selftest (22/22), a run on a realistic case with every command and verdict pasted into WITNESS.md as printed (An Example Co clean-up agent's handback taken from 4 to 100: its own note, a draft from the run state with two markers, and the filled note; 44 of 48 in the trigger test); and the trigger test, the card's line tried on Haiku, Sonnet and Opus beside 16 neighbouring skills: 44/48 on its own prompts, 0 wrong picks; the twelve-check audit (AUDIT.md). The punch on the face is the SHA-256 of SKILL.md, so a downloaded copy can be checked against this room.
Can it be wrong?
Yes — it is an AI reading a text file. Reads the note's shape and words, not whether its evidence is true; a person still reads it before acting.
What will it never do?
It refuses writing evidence the run does not hold; calling a hope evidence; inventing the decision's options, recommendation or deadline; calling a note READY with markers open.
Who may use it?
One person, any number of their own AIs and computers; a team needs a company licence.
