When to Split an AI Task into Two Passes

Prompt Craft

Combining extraction, judgement and polished writing in one request can make errors harder to spot. Split the work where a person needs to inspect an intermediate result.

Prompt evaluation worksheet with normal, edge and failure cases
Original editorial diagram by Prompt Task Journal. Illustrative worksheet, not an application screenshot.

Find the inspection point

A useful boundary lies between extracting facts and drafting a narrative. For example, first build a table of documented product changes; then write an announcement from the approved table. Splitting a task into arbitrary tiny steps can add overhead, so choose the boundary that exposes a meaningful failure.

Make the intermediate output explicit

Give the first pass a schema and a rule for missing information. Include references to the supplied source. The reviewer should be able to approve or correct that output without reading a polished story around it. Do not let an unverified first pass become authoritative simply because it appears in a table.

Carry only approved context forward

Feed the second pass the corrected intermediate result and the writing brief. Do not ask it to silently restore details that were removed during review. Keep unresolved items labelled. If the second pass needs a new fact, return to the source instead of relying on the model’s memory of the earlier conversation.

Compare total effort

A two-pass process is worthwhile if it makes review clearer or reduces costly rework, not merely because it produces more output. Try it on a recurring task and record the corrections required. Keep the simpler one-pass approach for low-risk drafts where intermediate review adds little value.

An example to adapt

PASS 1: Extract factual changes into a table: change | affected audience | source reference | uncertainty. Do not write the announcement.
HUMAN REVIEW: Correct and approve the table.
PASS 2: Draft the announcement from the approved table. Preserve uncertainties and do not add factual claims. Flag any information still needed.

Illustrative starting point. Adapt it to your inputs, permissions and review process.

Before you use it

  • The split creates a meaningful review point.
  • The intermediate output is verified.
  • Only approved facts reach the second pass.
  • The extra step improves the workflow.