The slop is in the handoffs.
AI has made drafting cheap and calibration scarce. Work that looks finished arrives subtly wrong, and someone downstream pays to repair it. Handoff contracts make quality expectations explicit and portable — between people, and between people and AI agents.
The format
Every contributor declares five things.
Agreed up front, while the collaboration is still being negotiated, and spanning every handoff from scheduling to final delivery — not sprung at the moment work changes hands. What I contribute. What I require. What I return. What “done” looks like. And what the work is for — because an exemplar that shone in one context can mislead in another, every contract names its intended use, and what it is not for. All of it anchored by an exemplar (“like the NewCo report”) plus a short delta (“shorter; board audience”). Senior practitioners already talk this way; nobody has written it down.
- I contribute
- Positioning analysis and two recommendations.
- I require
- The competitor feature matrix, confirmed before I begin.
- I return
- An annotated outline — twelve to fifteen slides’ worth of argument.
- Done looks like
- The designer can build slides without asking what any bullet means.
- This is for
- Choosing which positioning move to fund — not product-team planning.
handoff_contract: "0.2" contributor: id: agent-claude contributes: "Competitor feature matrix" requires: - input: "Confirmed competitor list" blocking: true # fails closed done_looks_like: "Every cell sourced; no unverified claims" exemplar: ref: "exidx://team/atlas-matrix" was_for: "Diligence sprint; 5-day funding decision" delta: scope: "Add pricing tier column"
The asymmetry is deliberate: an agent will not proceed without its declared inputs; a human gets an advisory flag when a deliverable diverges from the contract.
Why now
The joints are where the value lives — and where the damage lands.
Slop crosses boundaries.
“Workslop” transfers repair costs to whoever receives it; across organizations it erodes trust in the process itself.
Davenport, 2026 →Chains, not tasks.
AI’s value accrues to chains of activities — and the cost of handing off intermediate outputs shapes where that value lands.
NBER w34859 →Structure is solved. Quality isn’t.
Flash-teams research answered who does what, when. How good, in what register, at what depth — no format specifies it. This one does.
Retelny et al., UIST →Status
Built to be tested, not admired.
A working draft with pre-specified conditions under which we would conclude it is wrong. Some pieces already exist; others are dated experiments: