Meeting transcripts are long, repetitive, full of half-formed thoughts, and rarely organized around the decisions that actually matter. A fluent narrative summary that captures the “feel” of the conversation is easy to generate and usually insufficient for follow-through. What most small teams need is a short decision log: what was decided, what remains open, who owns the next visible step, and what information is still missing. The workflow below is the version I now recommend. It is narrow on purpose. It produces less polished prose and more accountability. That is the trade-off worth making.
Why Structured Decision Logs Beat Narrative Summaries
Narrative Summaries Hide Ownership Gaps
A well-written paragraph can describe a lively and productive discussion without ever naming who is supposed to do what by when. Decision logs force ownership and next-step fields into the open. If an owner is missing, the gap is immediately visible and can be corrected before the artifact is shared.
Actionability Requires Explicit Capture of Open Questions
Many meetings end with unresolved items that everyone notices in the room and then forgets within a day. Capturing those items as a prioritized open-question list turns ambient uncertainty into a concrete agenda for the next working session.
Human Confirmation Protects Against Invented Consensus
Models can infer agreement where the transcript only shows exploration or polite discussion. A mandatory human pass against the memory of the room or the recording remains the cheapest and most reliable insurance against false consensus appearing in the official record.
The Core Workflow Steps in Order
Step 1: Prepare a Minimally Clean Transcript
Remove only the noise that actively interferes with comprehension—repeated speaker labels if they dominate the text, obvious transcription artifacts. Do not heavily edit content or rephrase statements. The model should see essentially the same messiness the participants produced. Over-cleaning can remove the very signals that indicate whether a decision was actually reached.
Step 2: Issue a Strictly Scoped Instruction
Ask for a decision log with four required fields for every item: the decision itself, the owner, the deadline or next visible step, and the supporting evidence from the transcript. Separately request a list of open questions that were raised but not resolved. Explicitly forbid narrative summary paragraphs or commentary. The prohibition against narrative is restated because models often drift toward storytelling when the transcript is long.
Step 3: Perform Lightweight Structural Validation
Before any content review, check that every decision line has an owner field filled. Any line lacking an owner is marked incomplete and returned. This single structural rule catches a large fraction of low-value outputs and prevents reviewers from spending time on material that cannot yet be acted upon.
Step 4: Conduct Human Confirmation Against Room Memory
The person who attended the meeting, or a designated reviewer with access to the recording, confirms that each listed decision actually occurred and that ownership is correctly assigned. Open questions are reviewed for completeness and accurate reflection of unresolved tension rather than for perfect wording. This step is non-negotiable.
Instruction Patterns That Improve Reliability in Practice
Force Ownership or Explicit Absence
The prompt requires either a named owner or the exact phrase “owner not assigned in transcript.” Silence on ownership is not permitted. This small constraint dramatically increases the percentage of outputs that are usable after light editing.
Separate Decisions From Discussion
Instruct the model to treat statements of intent, exploration, or polite interest as non-decisions unless clear commitment language appears in the transcript. The distinction reduces the frequency of invented consensus and keeps the decision count honest.
Impose Soft Length Limits by Design
Request a maximum number of decision lines and open questions. Length limits discourage padding with low-value items and keep the final artifact short enough that busy participants will actually read and act on it.
Common Failure Modes and the Immediate Defenses
Invented or Inferred Owners
The model occasionally assigns an owner based on who spoke most frequently or who seemed most enthusiastic rather than who explicitly accepted the action. Defense: the human confirmation step remains mandatory for every ownership field; no exceptions for “obvious” cases.
Softened or Rephrased Open Questions
Unresolved tensions are sometimes rewritten as mild suggestions or next-step ideas, losing the original uncertainty. Defense: instruct the model to preserve uncertainty language where it appears, then have the reviewer restore stronger wording if the model has softened it.
Deadline or Timeline Hallucination
When no timeline was stated in the meeting, the model may supply a plausible one that makes the log look more complete. Defense: require every deadline to be either directly supported by the transcript or marked “not specified.” Invented dates are treated as review failures.
The following table lists the minimum required fields and the review rule applied to each:
Field | Required Content | Review Rule |
|---|---|---|
Decision | Clear statement of what was decided | Confirm it actually occurred in the meeting |
Owner | Named person or “not assigned” | Confirm acceptance or correct the assignment |
Next Step / Deadline | Transcript-supported or “not specified” | Reject any invented timeline |
Open Question | Unresolved item raised but not settled | Check for completeness and preserved uncertainty |
How to Roll the Workflow Out Without Creating Resistance
Begin With Recurring Internal Meetings
Start where the cost of missing owners is high and the external political risk is low. Success on internal decision logs builds the habit and the trust required before the same process is applied to client-facing or cross-functional meetings.
Make the Decision Log the Default Post-Meeting Artifact
Replace long narrative notes with the structured log as the official record. Culture follows the artifact that people are actually expected to open and use. If the narrative summary remains the default, the decision log will be treated as optional extra work.
Review the Review Load After Several Complete Cycles
If human confirmation consistently takes longer than the meeting itself, either the prompt needs tightening or the meeting culture needs clearer decision discipline. Net time saved across the full cycle is the only productivity metric that matters.
Closing the Decision-Log Workflow
Turning transcripts into actionable decisions is less about generating elegant prose and more about forcing ownership, next steps, and open questions into visible, checkable fields. Opus 5.2 can accelerate the drafting of those fields when the instructions are narrow and the prohibitions are explicit. It cannot replace the human who was in the room and who must still stand behind the record.
Useful remains a higher bar than impressive. A meeting workflow earns its place when the resulting log is short enough to read, complete enough to act on, and accurate enough that participants trust it the next morning. The process above is built to clear that bar more reliably than any generic “summarize this meeting” request I have tested.
