Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Turn meeting notes into cited decisions, owners, tickets.
.claude/skills/nousresearch-meeting-action-items/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 83% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 132% | 0% |
| case-17 | ✗→✓ | ▲ Improved | 346% | 0% |
| case-22 | ✗→✓ | ▲ Improved | 169% | 0% |
| case-12 | ✓→✗ | ▼ Worse | -40% | 0% |
Convert an existing transcript or notes set into accountable follow-through. teams-meeting-pipeline can retrieve Teams artifacts; this skill begins once notes/transcript content is available, from any source.
Don't use for: retrieving meeting recordings or transcripts (use teams-meeting-pipeline or the relevant connector first).
Use read_file on the provided notes/transcript files. Identify meeting title/date, participants, source files, transcript completeness, and whether speaker/time references exist. Done when missing portions and low-confidence transcription are stated.
Extract into distinct lists:
Do not turn brainstorming into decisions. Done when each candidate item has a supporting quote, timestamp, page, or note reference when available.
For every commitment record:
| Field | Rule | |---|---| | outcome | Concrete result, not a vague topic | | owner | Explicit named owner; otherwise unresolved | | due date | Explicit date or unresolved; never invent one | | dependency | What must happen first | | acceptance | Observable completion condition | | source | Transcript/note reference |
Done when every action has supported fields or visible unresolved values.
Load the user's tracker connector (notion, github-issues, or whichever system owns the work). Search for matching open items before creating anything — recurring meetings breed duplicate tickets. Preserve conflicts in owner/date/status for confirmation rather than silently overwriting. Done when proposed creates vs updates are distinguished.
Draft concise minutes with decisions, action table, unresolved questions, and next checkpoint. Prepare proposed tickets/tasks and a follow-up email/chat message, but do not publish yet — drafting is not sending. Done when the user can approve each external effect individually.
Create/update only approved records, attaching meeting provenance. Read back assignees, dates, status, and links from the provider. For ambiguous timeouts, search for the provenance marker before retrying — a blind retry duplicates records. Done when each approved item has a verified destination result.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-21 | fail→fail | 12,640 | 4,434 | -65% | 1 | 1 | 0% | 229 | 934 | +308% | 0 | 0 | — |
case-01 | fail→fail | 5,259 | 7,059 | +34% | 1 | 1 | 0% | 327 | 1,268 | +288% | 0 | 0 | — |
case-02 | fail→fail | 5,222 | 6,563 | +26% | 1 | 1 | 0% | 875 | 1,159 | +32% | 0 | 0 | — |
case-03 | fail→fail | 5,710 | 5,382 | -6% | 1 | 1 | 0% | 363 | 1,072 | +195% | 0 | 0 | — |
case-04 | fail→pass | 4,788 | 4,298 | -10% | 1 | 1 | 0% | 741 | 1,353 | +83% | 0 | 0 | — |
case-05 | pass→pass | 3,664 | 4,114 | +12% | 1 | 1 | 0% | 512 | 1,328 | +159% | 0 | 0 | — |
case-06 | pass→pass | 3,439 | 3,882 | +13% | 1 | 1 | 0% | 466 | 1,305 | +180% | 0 | 0 | — |
case-07 | fail→fail | 8,389 | 6,298 | -25% | 1 | 1 | 0% | 839 | 1,136 | +35% | 0 | 0 | — |
case-08 | fail→fail | 2,976 | 6,434 | +116% | 1 | 1 | 0% | 473 | 1,008 | +113% | 0 | 0 | — |
case-09 | fail→fail | 10,111 | 6,931 | -31% | 1 | 1 | 0% | 37 | 1,221 | +3200% | 0 | 0 | — |
case-10 | pass→pass | 4,218 | 7,465 | +77% | 1 | 1 | 0% | 584 | 1,980 | +239% | 0 | 0 | — |
case-11 | fail→fail | 25,974 | 6,033 | -77% | 1 | 1 | 0% | 3,983 | 1,134 | -72% | 0 | 0 | — |
case-12 | pass→fail | 14,849 | 6,842 | -54% | 1 | 1 | 0% | 1,836 | 1,108 | -40% | 0 | 0 | — |
case-13 | fail→fail | 5,609 | 5,784 | +3% | 1 | 1 | 0% | 344 | 1,121 | +226% | 0 | 0 | — |
case-14 | fail→fail | 5,243 | 6,722 | +28% | 1 | 1 | 0% | 373 | 1,164 | +212% | 0 | 0 | — |
case-15 | fail→pass | 16,640 | 30,598 | +84% | 1 | 1 | 0% | 2,553 | 5,923 | +132% | 0 | 0 | — |
case-16 | fail→fail | 3,843 | 4,788 | +25% | 1 | 1 | 0% | 557 | 911 | +64% | 0 | 0 | — |
case-17 | fail→pass | 4,777 | 12,710 | +166% | 1 | 1 | 0% | 645 | 2,877 | +346% | 0 | 0 | — |
case-18 | fail→fail | 35,002 | 8,362 | -76% | 1 | 1 | 0% | 236 | 1,318 | +458% | 0 | 0 | — |
case-19 | fail→fail | 29,516 | 4,452 | -85% | 1 | 1 | 0% | 4,075 | 954 | -77% | 0 | 0 | — |
case-20 | fail→fail | 4,137 | 9,272 | +124% | 1 | 1 | 0% | 681 | 1,458 | +114% | 0 | 0 | — |
case-22 | fail→pass | 4,690 | 8,147 | +74% | 1 | 1 | 0% | 795 | 2,142 | +169% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 7 counted toward the lift figure. The other 15 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +14 percentage points is the difference between those two pass rates over the 7 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.