跪拜 Guibai
← All articles
Artificial Intelligence

Turning a 67-Minute Meeting Recording into Executable Minutes with a Three-Pass AI Workflow

By 小月土星 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most AI meeting-minutes tools produce a fluent summary that reads well but buries uncertainty. This workflow treats the first draft as a draft to be interrogated, not a product to be shipped, and adds a dedicated fact-check pass that surfaces exactly where the model is overconfident. The result is minutes a team can act on without a second round of Slack threads asking who owns what.

Summary

Processing a 67-minute, 11-person meeting recording used to take 60–90 minutes of manual transcription and cross-referencing. A three-pass workflow using TRAE Work now produces a first draft in 8 minutes. The key is separating generation, fact-checking, and final formatting into distinct stages, each with strict boundary instructions that forbid the model from guessing, filling gaps, or upgrading tentative suggestions into decisions. A dedicated fact-check pass flags every instance where the summary lacks a source timestamp or where uncertain language was hardened into a conclusion. The final output follows a fixed four-section structure—one-page summary, topic-by-topic record, action-item table with a traceability column, and a pending-confirmation list—so the team can jump directly to what they need. Measured across the same material, executable action items rose from 2 to 5, rework dropped by half, and the total processing time shrank by roughly 80%.

Takeaways
First-draft time dropped from 45 minutes to 8 minutes by splitting the task into generation, fact-checking, and final formatting passes.
Executable action items increased from 2 to 5 because the workflow forces missing owners, deadlines, and deliverables to surface as “待确认” (to be confirmed) instead of being silently omitted.
Rework cycles fell from 2 to 1, and secondary verification time dropped from 20 minutes to 5.
The fact-check pass is the linchpin: it outputs a table of “Original Sentence | Risk | Suggested Rewrite | Person Who Needs to Confirm” without rewriting the full document.
Every action item in the final table includes a “Basis” column that points back to the original speech timestamp, making handoff and verification straightforward.
A reusable prompt template is provided that encodes the three-pass logic and the prohibition against guessing, supplementing facts, or making decisions on behalf of attendees.
Conclusions

The workflow’s value is not in generating text faster but in eliminating the hidden cost of re-listening to scattered segments to verify who said what and whether a suggestion became a decision.

Separating generation from fact-checking into distinct rounds prevents the model from smoothing over uncertainty for the sake of fluency—a failure mode common in single-pass summarization.

Explicitly instructing the model to mark missing fields as “待确认” rather than inferring them turns gaps from silent errors into visible, actionable items.

The insistence on a traceability column in the action-item table shifts the deliverable from a narrative summary to an auditable artifact, which is what makes it reusable across teams and meetings.

Concepts & terms
Three-pass AI workflow
A meeting-minutes pipeline that splits the task into three distinct stages: a first pass for structured generation, a second pass dedicated solely to fact-checking and surfacing uncertainty, and a third pass for final formatting. Each pass has its own prompt and constraints.
Fact-check pass
A dedicated AI round that compares the generated summary against the source material, flagging unsupported conclusions, overconfident language, and missing action-item fields. It outputs a risk table rather than a rewritten document.
Traceability column (依据列)
A column in the action-item table that links each task back to the original speech timestamp or document source, enabling anyone who inherits the minutes to verify decisions without re-listening to the full recording.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗