Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Extract atomic insights from raw source material. Takes articles, transcripts, meeting notes, or any unstructured input and produces structured claims with provenance metadata. The foundational extraction step of the 6R processing pipeline. Triggers on: "extract", "reduce", "distill", "summarize source"
.claude/skills/miosa-osa-reduce/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 772% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 29% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -47% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 950% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 422% | 0% |
> Extract insights from source material into atomic, reusable claims.
Transform raw input (articles, transcripts, notes, documents) into structured atomic claims — each with provenance, confidence, and topic tags. This is the core extraction step: it turns noise into signal. Every claim stands alone and can be recombined later.
bash# Reduce a file /reduce path/to/transcript.md # Reduce inline text /reduce --text "Ed called about pricing. He wants $2K per seat for enterprise..." # Reduce with explicit source metadata /reduce path/to/article.md --source "HBR" --date 2026-03-15 # Reduce with depth control /reduce path/to/paper.pdf --depth deep # Reduce multiple files /reduce path/to/*.md --batch
| Flag | Type | Default | Description | |------|------|---------|-------------| | <input> | positional | required | File path, glob pattern, or --text for inline | | --text | string | — | Inline text to reduce (alternative to file input) | | --source | string | auto-detect | Source attribution (author, publication, URL) | | --date | date | today | Date of the source material | | --depth | enum | standard | quick (key points only), standard (claims + context), deep (claims + evidence + counterpoints) | | --format | enum | atomic | atomic (one claim per block), outline (hierarchical), table (comparison grid) | | --max-claims | int | 50 | Maximum number of claims to extract | | --tags | string] | auto | Topic tags to apply (auto-detected if omitted) | | --batch | flag | false | Process multiple files, one output per input | | --output | path | stdout | Write to file instead of stdout |
fact, opinion, decision, action-item, question, insight. Assign confidence (0.0–1.0).Each reduced file produces structured output:
yaml--- type: reduction source: "path/to/original.md" source_title: "Q1 Strategy Call with Ed" source_date: 2026-03-15 reduced_at: 2026-03-20T14:30:00Z claim_count: 12 topics: [pricing, enterprise, ai-masters] --- ## Claims ### 1. Enterprise pricing target is $2K/seat - **type:** decision - **confidence:** 0.9 - **speaker:** Ed Honour - **context:** Discussed during pricing review segment - **tags:** [pricing, enterprise, ai-masters] ### 2. Current conversion rate from free tier is 3.2% - **type:** fact - **confidence:** 0.7 - **speaker:** Roberto - **context:** Referenced but not sourced — verify against analytics - **tags:** [metrics, conversion, funnel]
/seed — Often receives input from seed (but can run standalone)/reflect — Output feeds into reflect for connection discovery.pdf inputs| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 5,254 | 8,570 | +63% | 1 | 1 | 0% | 302 | 2,634 | +772% | 0 | 0 | — |
case-02 | fail→pass | 7,333 | 7,003 | -5% | 1 | 1 | 0% | 1,670 | 2,154 | +29% | 0 | 0 | — |
case-03 | fail→pass | 38,389 | 16,598 | -57% | 1 | 1 | 0% | 7,028 | 3,708 | -47% | 0 | 0 | — |
case-04 | fail→pass | 3,719 | 17,569 | +372% | 1 | 1 | 0% | 354 | 3,718 | +950% | 0 | 0 | — |
case-05 | fail→pass | 3,824 | 9,582 | +151% | 1 | 1 | 0% | 492 | 2,567 | +422% | 0 | 0 | — |
case-06 | pass→pass | 16,769 | 15,049 | -10% | 1 | 1 | 0% | 2,137 | 3,196 | +50% | 0 | 0 | — |
case-07 | fail→pass | 38,274 | 2,636 | -93% | 1 | 1 | 0% | 1,800 | 1,366 | -24% | 0 | 0 | — |
case-08 | fail→pass | 11,462 | 2,301 | -80% | 1 | 1 | 0% | 2,023 | 1,287 | -36% | 0 | 0 | — |
case-09 | fail→pass | 17,186 | 2,068 | -88% | 1 | 1 | 0% | 2,250 | 1,279 | -43% | 0 | 0 | — |
case-10 | fail→pass | 8,418 | 5,828 | -31% | 1 | 1 | 0% | 1,524 | 2,005 | +32% | 0 | 0 | — |
case-11 | fail→pass | 6,546 | 5,700 | -13% | 1 | 1 | 0% | 1,018 | 2,068 | +103% | 0 | 0 | — |
case-12 | fail→pass | 31,644 | 2,260 | -93% | 1 | 1 | 0% | 4,235 | 1,277 | -70% | 0 | 0 | — |
case-13 | fail→pass | 10,523 | 2,442 | -77% | 1 | 1 | 0% | 1,807 | 1,373 | -24% | 0 | 0 | — |
case-14 | pass→pass | 10,009 | 3,584 | -64% | 1 | 1 | 0% | 1,602 | 1,407 | -12% | 0 | 0 | — |
case-15 | pass→pass | 17,824 | 2,691 | -85% | 1 | 1 | 0% | 1,721 | 1,363 | -21% | 0 | 0 | — |
case-16 | fail→pass | 8,881 | 3,910 | -56% | 1 | 1 | 0% | 1,062 | 1,769 | +67% | 0 | 0 | — |
case-17 | fail→pass | 7,340 | 7,360 | +0% | 1 | 1 | 0% | 1,222 | 2,101 | +72% | 0 | 0 | — |
case-18 | pass→pass | 8,452 | 9,879 | +17% | 1 | 1 | 0% | 1,463 | 2,500 | +71% | 0 | 0 | — |
case-19 | fail→pass | 11,072 | 7,864 | -29% | 1 | 1 | 0% | 2,310 | 2,292 | -1% | 0 | 0 | — |
case-20 | pass→pass | 6,525 | 2,227 | -66% | 1 | 1 | 0% | 868 | 1,254 | +44% | 0 | 0 | — |
case-21 | pass→pass | 19,953 | 3,457 | -83% | 1 | 1 | 0% | 2,010 | 1,366 | -32% | 0 | 0 | — |
case-22 | fail→pass | 16,836 | 2,627 | -84% | 1 | 1 | 0% | 2,604 | 1,332 | -49% | 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 21 counted toward the lift figure. The other 1 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 +73 percentage points is the difference between those two pass rates over the 21 comparable cases.
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.