The first assistant builds a plan around an unspoken premise: the client can accept a breaking response change. You switch assistants, and the second assistant receives only the plan. It sees the premise as part of the world rather than as a guess that should be checked.
An assumption survives a handoff safely only when its origin and failure condition travel with it.
Name the premise before the conclusion
An assumption is something you are treating as true so work can proceed. It may be reasonable and still be unverified.
Write the premise as a direct sentence:
- The client can update before deployment.
- The source data contains one record per account.
- The reviewer cares more about reversibility than speed.
Then preserve the two origin labels RelayLink carries:
- Human-stated: you explicitly supplied the premise.
- AI-inferred: the assistant derived it from wording, behavior, or missing information.
Record supporting basis separately: a named source and its check status, or a working convention the team chose. Those are notes about support, not extra provenance labels. A source-based premise is still human-stated or AI-inferred according to who supplied it.
Human-stated does not mean true. AI-inferred does not mean useless. The labels answer origin, which is what the receiving assistant needs in order to weigh and challenge the claim. Separating facts from a guess keeps fluent grammar from erasing that boundary.
Record why the assumption matters
Do not send a long inventory of harmless guesses. Preserve the assumptions that change the answer.
For each important premise, add:
- Impact: what recommendation, design, or draft depends on it.
- If false: what breaks or must be reconsidered.
- Evidence for: the observation or source supporting it.
- Disconfirmer: the observation that would show it is false.
- Recheck: who should verify it, how, and before which action.
“Assume the client can update” becomes useful when paired with “If false, keep the old response shape; check the supported client versions before changing the interface.”
The disconfirmer matters because assistants are good at extending a premise once it is present. A named way to defeat it tells the second assistant that the assumption is a question, not an instruction to defend.
Keep assumptions attached to decisions
If a decision depends on a premise, link them in the briefing. State which choice must be reopened if the premise fails. Otherwise the second assistant may correct the assumption while leaving the decision built on it untouched.
Preserving decisions across assistants adds the owner, rejected options, consequences, and reopen condition. The assumption record supplies one common trigger for that condition.
Do not retroactively relabel an inference as a confirmed fact because a human approved the package. Approval authorizes the text to travel. Verification requires evidence.
Send assumptions in a self-handoff
Draft the assumption register to your own email on the RelayLink account both assistants use. Review each premise and its stated-versus-inferred label, then explicitly confirm it. Same-account delivery needs neither a second account nor a contact request and creates no email or magic link. The register stays unread until any assistant pulls it; the account portal is the fallback inbox.
What travels is the approved structured assumption text. RelayLink does not transfer hidden memory, the complete chat, files, or browser state, and it does not verify a premise. A reference grants no access. The stated-versus-inferred labels you confirm remain part of the composed package.
Do not include secrets, customer data, confidential source, or restricted evidence unless the provider and receiving environment are authorized. Prefer a safe reference or path to the system where the check should occur.
Recheck before the assumption becomes expensive
The receiving assistant should begin by identifying assumptions tied to irreversible or costly steps. Verify those before implementation, publication, or commitment. Lower-impact premises can remain open if their risk is stated.
If evidence disproves one, update the record, revisit every dependent decision, and preserve the result. If evidence supports it, record what was checked rather than silently deleting the assumption.
The point is not to eliminate uncertainty. It is to stop uncertainty from changing its name when the assistant changes.