Relay racing | Master AI Automation in 4 hours Master AI Automation in 4 hours Course About Ayush Modules Sample chapter Toolbox The Microcap Minute Module 07: The AI Relay Team / Chapter 2 Relay racing Watch first, then read.
Same lesson, your pace.
What you will learn - What a handover is between AIs (and humans) - The four classic handover failures - State: what must travel with the baton The baton In a relay, runners don't share a brain, they share a baton : a physical object carrying everything that matters.
AI handovers work identically.
When Qwen's draft goes to Claude for review, Claude knows nothing except what you paste.
No shared memory exists between chats, models or apps (Module 1 Chapter 2, every desk starts empty).
A handover is therefore an act of deliberate information transfer: Everything the next worker needs to succeed without asking questions.
Not "here's my draft, improve it", that's not a baton, it's a shrug.
The four classic failures Study any bad multi-AI workflow and one of these four will be the culprit: Failure Symptom Cure Context collapse Next AI "forgets" constraints ("I didn't know it was for 12-year-olds") Constraints ride inside the baton, never in old chats Verbosity tax Pasting 3,000 words where 200 suffice; attention dilutes Curate: relevant facts only Format rot Each hop mangles structure until output is mush Declare format rules once, carry them verbatim Goal drift By step 4, nobody remembers the original ask Goal statement leads every handover, word for word Notice they're all the same sin wearing different clothes: assuming shared understanding that doesn't exist.
What travels in the baton The minimal viable state for any handover: Goal , what done looks like (one sentence) Constraints , audience, length, tone, hard limits Artifact , the current work itself (clearly marked) Ask , precisely what this receiver should do to it Verboten , what NOT to change/do (as important as the ask) Chapter 3 turns this list into a reusable template with a name, so your fingers stop re-deciding it.
But the concept comes first: if a handover lacks any of the five, expect the matching failure above.
Try it yourself Deliberately commit failure #1: give Kimi a task with constraints ("under 50 words, no jargon"), get output, then paste only the output into ChatGPT saying "improve this", watch it violate both constraints.
Redo the handover with all five baton elements and compare.
Log both versions side by side in learn/handover-anatomy.md .