An AI conversation has produced a draft. Within that conversation, the reason for the changes and the next steps seem clear. Hand the work to another person or agent, though, and someone has to explain again what needs checking.
Is the draft still being prepared? Has its content been reviewed? Which version is on the public site? A reply saying “done” does not distinguish these states.
In Fragment Practice's own development and operations, I have been considering what needs to remain outside the conversation when work passes to an AI. The design question is what lets someone who was absent from that conversation continue the task.
What would an article handoff need?
Consider an agent whose work stops after preparing a revision to a public article. The full conversation contains rejected proposals and intermediate decisions. Its latest explanation may not describe the current state.
Four things need to be kept distinct in the handoff.
| What to keep distinct | What an article revision requires checking |
|---|---|
| Task record | What is the purpose, current state and remaining work? What would make the task complete? |
| Material with an identifiable version | Which draft and rules are being reviewed? Which decision applies to which version? |
| Actual state of the public site | What is currently displayed? Did the prepared change take effect? |
| Execution actor and authority | What may the next actor do? Is there a decision authorizing publication of this version? |
Reviewing the content does not decide its publication. Access to a publishing operation does not itself authorize publishing that draft. The state of the work, the decisions about it and what happened outside the conversation need to be checkable separately.
What Git records, and what the public site establishes
In this practice, Git holds the change history of drafts, code and operating rules. It records what changed and makes it possible to identify the version under discussion. GitHub's project and issue records hold the purpose and progress of current work: work cards for passing on remaining tasks and the next decision.
A draft in the change history is still different from a draft on the public site. The publication result has to be checked at the destination. Work involving another service likewise requires looking at that service's present state.
The useful distinction is the question each record should answer. Do I need the draft version, the remaining work, or the result of an operation? In another organization, the starting point could be whether its existing document and work-management systems can answer those questions.
Decisions to revisit on resumption
Finding the draft does not always make publication the next action.
- If only the draft was prepared, continue the remaining editing and review.
- If publication may already have happened, check the destination before repeating the operation.
- If the draft changed after review or approval, check whether that decision still applies to the new version.
A conversation summary can help locate this information. Resuming the work requires comparing the relevant draft, decision record and publication state. AI Delegation Reveals How Work Was Designed considers the related question of what should count as completion.
Does more recording make a handoff easier?
More detailed records also create work to keep them current. Recording the same state in several places raises further questions: which one should be updated, and which should be trusted when they disagree? A mechanism built to support resumption can acquire maintenance work of its own.
I have not yet measured how this separation changes resumption time or human workload. To evaluate it, I would want to see what the next person had to find again, where they waited for a decision, and what rework remained.
Can someone who was absent from the conversation tell what has finished, what must be checked, and what they may do next? For now, I want to understand what makes those records usable for the person continuing the task.