Running Multi-Stage Requests: Handoffs, Review Points and Refinement

1.What each stage needs to receive

In a staged request, each stage only knows what it is given. If a stage continues in the same conversation, the earlier work is still there. Long jobs often move to a new conversation, though, and a new conversation starts with nothing from the last one. The AI Fluency course for educators handles this by having learners either continue the earlier conversation or output a summary of that conversation to use as input for a new one. Its lesson takeaway is that established context makes each new workflow better than starting fresh.

Take a bid team that agreed three criteria for judging partner organisations in an earlier stage. They now start the scoring stage in a new conversation. The scoring request must state those three criteria again, together with the material to be scored. Otherwise Claude has to guess the standard. Anthropic's orchestrator-workers cookbook describes the same rule from the engineering side: each worker gets the original task as well as its own specific instructions. A stage told only "score these" is missing the reason for the scoring and the standard to score against.

Sources12

2.The handoff risk: summaries drop detail

Every handoff between stages compresses something. That is the point of a handoff, and also its main risk. Anthropic's multi-agent guidance describes teams that split work into separate planning, execution, review and iteration steps. They found they had lost context at each handoff and spent more tokens coordinating than doing the work. The same thing happens in ordinary business work. If a team summarises twenty customer interviews and then decides which product problems to fix using only the summaries, then any nuance, minority complaint or exact customer wording the summaries left out never reaches the decision.

The AI Fluency course for builders states the general principle: a failure at any link cascades downstream. A weak summary in stage two quietly limits everything in stages three, four and five. Two habits reduce the risk. First, tell the summarising stage what the later stages will need, so it keeps the right details. Second, keep the original material available, so a later stage or a reviewer can go back to it when a summary seems thin.

Sources34

3.Where to put the human review point

Being able to check the work between stages is a major reason to decompose a task at all. Anthropic's prompting guidance says explicit chaining is still useful when you need to inspect intermediate outputs or enforce a specific pipeline structure. The engineering guidance adds that you can put checks on any intermediate step to ensure the process is still on track. In Anthropic's document example, the check sits on the outline, before the document is written.

If you only have time for one review, the cascade principle tells you where it should go: the stage whose output every later stage depends on. Problems found there are cheap to fix. Problems found after the final draft may mean redoing several stages. For a policy rewrite in several stages, that usually means reviewing the agreed outline or criteria before the full drafting starts, not just reading the finished text.

Sources567

4.The draft, review and refine chain

Anthropic's prompting guidance names the most common chain: generate a draft, have Claude review it against criteria, then have Claude refine it based on the review. Each step is a separate prompt, so you can read the feedback, accept or reject it, or change direction before the next step runs. The criteria matter. A review prompt that names what to check against gives focused, usable feedback. A vague "make it better" gives the refine step nothing specific to act on.

Anthropic's prompt engineering blog shows the pattern as three prompts in a row, using a research summary as the example.

First prompt: "Summarize this medical paper covering methodology, findings, and clinical implications."

Second prompt: "Review the summary above for accuracy, clarity, and completeness. Provide graded feedback."

Third prompt: "Improve the summary based on this feedback: [feedback from step 2]"

In the example, the review step names three criteria (accuracy, clarity and completeness) and asks for graded feedback. The refine step then gets the feedback pasted in word for word, not a vague instruction to improve. This is the earlier advice applied to a small chain: decide the standard, pass forward exactly what the next stage needs, and keep a point where you can look before moving on.

Sources57

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.When a later stage starts in a new conversation, Claude will still know the criteria agreed in the earlier stage.Why is that wrong?

   A new conversation starts without the earlier work. Carry it forward by passing in a summary of the earlier conversation, or by restating the criteria.

   Covered in What each stage needs to receive

2. 2.Once each source has been summarised, later stages can rely on the summaries alone without losing anything important.Why is that wrong?

   Each handoff can lose context. Any detail dropped at an early link is missing from every stage that follows.

   Covered in The handoff risk: summaries drop detail