In partnership with
Hey builders,
“Done” is not a deliverable. If an AI agent changed your repo, the useful handoff is three things you can inspect: what changed, what it actually checked, and one artifact that proves the result.
Start here: save the copy block below as AGENT-HANDOFF.md before your next low-risk task. It works with any coding assistant you already use—and it still works as a pull-request checklist if you use no agent at all.
📦 From Our Partners
285 Free Prompts to Save Time
Your inbox isn't the problem. Neither are your meetings. It's the hours you spend on work AI could do in minutes.
285 copy-paste prompts: built for email, meetings, admin, and delegation. Each one has a clear role, a specific task, and a defined output.
No setup. No prompt engineering. Just copy, paste, and get your time back. Free when you subscribe to Mindstream.
1,000+ Claude Prompts Top Professionals Actually Use at Work
Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.
These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.
Sign up for Superhuman AI and get:
1,000+ ready-to-use Claude prompts to get real work done in minutes — researched, tested, and used by professionals at Google, Microsoft, and NASA
Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career — the prompts are just the beginning
🔨 BUILD OF THE WEEK
An agent can write code fast and leave a human with the worst status update: “finished.” Independent developer Michael Lynch describes agents declaring work complete too early and presenting plans too messy to review. That is a workflow problem, not a reason to chase another model.
The community signal is visual, too. Big Arrow on the Screen lets supported agents point at a control instead of merely printing “click Allow.” Its Show HN post had 412 points and 188 comments when checked. The Mac tool is niche; visible evidence is not.
My take: Give agents a receipt format, not more room for confident prose.
Hype Scale: 🔨🔨 (2/5) — this is a boring habit, which is why it is more likely to survive the next agent launch.
⚡ THE 5-MINUTE MOVE
Turn “done” into a three-proof handoff
Choose a reversible task. Try a doc typo, a test-only change or a tiny UI label—not a deploy, payment, permission, deletion or production edit.
Save the handoff. Copy the block below into
AGENT-HANDOFF.md(or paste it into the task) and fill the six bracketed fields.Name the one check first. Tell the agent the exact command or manual check it must run. If it cannot run it, it must say
NOT RUN.Require three proofs. Do not accept
COMPLETEwithout a changed-file list, a real validation result and a focused diff, screenshot or expected-vs-actual example.Review the limits. Read
LIMITS & HANDOFFbefore merging. An honestPARTIALis cheaper than a fictional green light.
The rule: a sentence is a claim. A file list, command result and artifact are evidence.
📋 THE THREE-PROOF HANDOFF
Copy this into AGENT-HANDOFF.md, then hand the completed file to an agent. It needs no API key, paid model or particular provider. It asks for evidence; it does not enforce permissions or prove the agent never touched anything else.
# Task contract
Outcome: [THE USER-VISIBLE OR TESTABLE RESULT]
In scope: [FILES / COMPONENT / ENVIRONMENT THE AGENT MAY TOUCH]
Out of scope: [SYSTEMS OR ACTIONS IT MUST NOT TOUCH]
Done means: [ONE OBSERVABLE ACCEPTANCE CONDITION]
Validation command or method: [EXACT COMMAND OR MANUAL CHECK]
Evidence artifact: [FOCUSED DIFF / SCREENSHOT / EXAMPLE]
Before changing anything, restate the outcome and name the validation method.
If you are missing information, need a risky action, or cannot validate, stop
and label the task BLOCKED. Do not silently choose a production, money, delete,
publish, permission, password, 2FA or external-send action.
Before you say complete, return exactly:
STATUS: COMPLETE | PARTIAL | BLOCKED
SCOPE: what changed; files added/edited/deleted; intentionally untouched areas
VALIDATION: exact command/method; PASS | FAIL | NOT RUN; key output/result
EVIDENCE: artifact supplied; what it demonstrates
LIMITS & HANDOFF: what was not tested; human decision/next step, if any
🧭 SUNDAY SUPPLY SHELF
See the side effects a Git diff misses. agent-trail is a MIT, local/offline Claude Code hook that records metadata and salted digests—not file contents. Grab it →
Find role drift in saved agent sessions. clauditor is a MIT, dependency-free Python 3.12+ scanner for Claude Code, Cursor and Codex transcripts. Grab it →
Add an audit trail and policy hooks across agents. Gryph is an open-source local tool that can log and policy-check supported coding-agent actions. Grab it →
🔎 THE TWO-MINUTE REVIEW
Scope: Are every claimed file and side effect inside the task’s allowed area? If not, stop and inspect the diff.
Validation: Did the agent name the command and its result? “Tests pass” is not enough;
NOT RUNis valid and reviewable.Evidence: Can you open the diff, screenshot or example and see the promised outcome? A screenshot proves appearance, not backend behavior.
Limits: Does the receipt say what was skipped—mobile check, auth state, integration test, real data? If it cannot, the job is
PARTIAL.
📐 ONE TERM A DAY
Today’s term: provenance
Plain English: Provenance is the label on a parcel that says where it came from and how it reached you. For an agent result, it is the trail from claim to command to artifact.
Why you should care: It lets you review a result without treating the agent’s summary as a witness under oath.
🎁 FREE SUPPLY DROP
1,000 Powerful Prompts for Claude AI — a prompt pack covering writing, coding, research and business work, so you stop starting from a blank box.
Tools Down
Swati, ByteBuilders