The summary reads beautifully. Check it anyway.
An AI summary of your workshop will be fluent, confident and about 80% right. The missing 20% is where the disagreements went.
What changed
Synthesis used to be the facilitator’s unglamorous evening. Photograph the wall, type up the notes, find the themes, write the summary the sponsor reads on the train. Now the model does it in forty seconds and the summary is better written than mine ever was.
It’s also wrong in ways you can’t see by reading it. Summaries smooth. They merge two different ideas because they sound similar. They drop the one person who disagreed because one voice doesn’t make a theme. They turn “we might try” into “the team recommends”. And the result reads like a conclusion, so nobody checks.
The skill
Trace the summary back to the source. Every claim in it should point at a sticky note, a quote or a vote. The ones that don’t get flagged or cut. It takes fifteen minutes and we do it with the team in the room, not afterwards.
| The summary said | The wall said | What leaked |
|---|---|---|
| “The team recommends calling customers before the bill lands.” | Four votes for it. One strong objection from the ops lead about call volumes. | Dropped disagreement |
| “Customers want a simpler bill with clearer comparisons.” | Two separate ideas: reorder the bill (9 cards) and add a comparison line (4 cards). Different owners, different cost. | Merged ideas |
| “The letter should be sent within 48 hours of the call.” | One note: “maybe 48h? check with print”. No vote, no owner. | “Maybe” became a conclusion |
| “Three prototypes will be tested with five customers on Thursday.” | Decided, written on the wall, owner named. | Traced. Keep. |
Fig. 7 — Four lines from a real summary. Three leaked. Read without the wall, all four looked fine.
The three leaks we look for every time:
- Omitted disagreement. Who objected, and where did their objection go? If someone in the room argued against a point and the summary doesn’t mention it, the sponsor will never know there was a question.
- Merged ideas. Two cards with different owners, costs or customers, folded into one sentence because they rhyme. Unmerge them. They’ll be built by different people.
- Tentative to definite. “Maybe”, “we could”, “worth trying” in the notes. “Recommends”, “will”, “should” in the summary. The model upgrades confidence by default. Downgrade it back.
What goes wrong
We sent an unchecked summary to a sponsor once. It read well. She approved a direction the team hadn’t actually agreed on, because the one dissenting voice had been smoothed out. Three weeks later the dissenter turned out to be right, and the sponsor felt misled. By a summary. Nobody had lied; the model had just written a tidier story than the room had told.
The other failure is doing the check alone. The facilitator reads the summary against the wall by themselves and misses the merge, because they weren’t in the conversation where those two ideas turned out to be different. Do it with the room. Read each line out. Ask “where’s this from?” The person whose idea got merged will tell you.
Run it Monday
- Generate the summary while the wall is still up and the team is still there.
- Read it line by line. For each claim, someone points at the note it came from.
- Three flags on the wall: dropped disagreement, merged, upgraded. Mark each leak.
- Fix the summary in the room. Add the dissent, split the merge, restore the “maybe”.
- Only then does it go to the sponsor. With the wall photo attached.
A summary that reads like a conclusion isn’t one until you can point at where each line came from.
The synthesis-check prompt, including the version that asks the model to cite the note for every claim, is in the facilitation deck.
