You switch assistants during a project. The second assistant reviews the current state and proposes an option you already considered and rejected two days ago. It is not ignoring you; the decision lived only in a conversation it never saw.
Preserve the decision as a small record. Do not make the next assistant rediscover your stopping point from a transcript.
Distinguish a decision from a recommendation
An assistant may recommend option A. A decision exists only when an authorized human chooses A for the work. Record that distinction directly:
- Recommendation: what an assistant proposed and why.
- Decision: what the human chose.
- Owner: the person accountable for the choice.
- Checkpoint: when or at which project state it was made.
This prevents “Claude preferred A” from turning into “we decided A.” It also prevents your own decision from being demoted to a model suggestion after the assistant changes.
If no decision was made, say “open.” A handoff should not manufacture closure to make the brief look tidy.
Record the paths you closed
For each plausible option the next assistant might raise, include one line explaining why it is closed. Focus on alternatives a reasonable reviewer would expect, not every idea generated during the session.
A useful record might say:
- Chosen: keep the existing endpoint and extend its response.
- Rejected: add a second endpoint, because clients would need two discovery paths.
- Rejected: replace the response shape, because compatibility is required.
The reason must be honest. If it depends on an assumption, label the assumption. Recording the option you rejected is what keeps silence from looking like an oversight.
Do not copy the whole debate. The next assistant needs the result, the strongest reason, and enough evidence to challenge a false premise.
Include consequences and a reopen condition
A decision changes what happens next. Record the consequences the receiving assistant must honor:
- Files, interfaces, or documents that now need work.
- Risks accepted by the owner.
- Follow-up checks or deadlines.
- Behavior deliberately left unsupported.
- People or systems affected.
Then write the condition that should reopen the decision. Good conditions are observable:
- Reopen if the compatibility test cannot pass.
- Reopen if the provider removes the required API.
- Reopen if measured latency exceeds the agreed threshold.
- Reopen if the named legal or security owner rejects the constraint.
“Revisit later” is not a condition. It tells the next assistant neither when nor why.
Carry the record to your own inbox
Draft the decision record to your own email on the RelayLink account both assistants use. Review the owner, rejected options, consequences, and reopen condition, then explicitly confirm it. Because the recipient is your own account, there is no second account or contact request and no email or magic link. The record stays unread until any assistant pulls it; the account portal is the fallback.
RelayLink carries the approved structured decision text, not hidden context, chat history, files, or project state. A path or reference grants no access; the second assistant needs separate authorized access to the working artifacts. Continuing a project in a different assistant uses this record as part of a broader checkpoint.
Keep secrets, customer information, confidential source, and restricted material out of the package unless the provider and receiving environment are authorized. References and paths are usually enough to locate the decision’s evidence.
Let evidence reopen work, not amnesia
The receiving assistant should treat the decision as a constraint until a reopen condition occurs or supporting evidence fails. That does not make the record infallible. It makes challenges specific.
If the second assistant finds that a rejected option was closed on a false assumption, bring the choice back to the owner. Record the replacement decision and its consequences. Do not quietly rewrite the earlier record as though the project had always chosen the later path.
Preserving decisions saves more than repeated discussion. It keeps the human choice, the model recommendation, and the evidence that could change the choice in their proper places.