Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Summarize a long document into an executive summary, key points, and role-specific implications. Use when the user says "summarize this document", "give me the key points", "do I need to read this?", or drops a long PDF, report, or contract and asks what matters in it.
.claude/skills/sgharlow-summarize-document/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 85% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 98% | 0% |
| case-04 | ✓→✓ | = Same ✓ | 11% | 0% |
| case-13 | ✓→✓ | = Same ✓ | -12% | 0% |
| case-19 | ✓→✓ | = Same ✓ | 85% | 0% |
Read a long document and return a decision-ready digest tuned to the reader's role. The summary must be faithful: every claim in it traces to the document, with section or page references where the format allows.
what they will do with it (present, decide, share). If context is missing, ask these three questions first; a summary without them is generic and half as useful.
(headings, tables, executive sections), then the sections that bear on the reader's stated purpose.
it matters to this reader.
stated purpose.
add it or state why it was omitted. Never present an inference as the document's claim — label reader-specific interpretation as "Implication", not "Finding".
file) — say exactly what was unreadable instead of summarizing the fragment silently.
Full walkthrough, examples, and variations: recipes/Recipe-004-Document-Summarization.md.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 3,843 | 4,112 | +7% | 1 | 1 | 0% | 583 | 1,079 | +85% | 0 | 0 | — |
case-02 | fail→fail | 7,867 | 4,494 | -43% | 1 | 1 | 0% | 1,232 | 1,231 | -0% | 0 | 0 | — |
case-03 | fail→fail | 4,883 | 4,024 | -18% | 1 | 1 | 0% | 755 | 1,095 | +45% | 0 | 0 | — |
case-04 | pass→pass | 9,341 | 8,883 | -5% | 1 | 1 | 0% | 1,604 | 1,780 | +11% | 0 | 0 | — |
case-05 | fail→fail | 2,995 | 4,841 | +62% | 1 | 1 | 0% | 441 | 1,236 | +180% | 0 | 0 | — |
case-06 | fail→fail | 2,798 | 4,655 | +66% | 1 | 1 | 0% | 396 | 1,193 | +201% | 0 | 0 | — |
case-07 | fail→fail | 4,455 | 4,543 | +2% | 1 | 1 | 0% | 720 | 1,186 | +65% | 0 | 0 | — |
case-08 | fail→fail | 5,233 | 4,446 | -15% | 1 | 1 | 0% | 829 | 1,178 | +42% | 0 | 0 | — |
case-09 | fail→fail | 8,661 | 5,186 | -40% | 1 | 1 | 0% | 1,278 | 1,292 | +1% | 0 | 0 | — |
case-10 | fail→fail | 8,578 | 3,951 | -54% | 1 | 1 | 0% | 1,382 | 1,091 | -21% | 0 | 0 | — |
case-11 | fail→fail | 18,956 | 7,971 | -58% | 1 | 1 | 0% | 3,246 | 1,756 | -46% | 0 | 0 | — |
case-12 | fail→fail | 9,440 | 3,559 | -62% | 1 | 1 | 0% | 1,451 | 1,029 | -29% | 0 | 0 | — |
case-13 | pass→pass | 14,406 | 9,675 | -33% | 1 | 1 | 0% | 2,189 | 1,928 | -12% | 0 | 0 | — |
case-14 | fail→fail | 6,019 | 4,693 | -22% | 1 | 1 | 0% | 966 | 1,282 | +33% | 0 | 0 | — |
case-15 | fail→fail | 4,208 | 3,951 | -6% | 1 | 1 | 0% | 654 | 1,075 | +64% | 0 | 0 | — |
case-16 | fail→pass | 3,284 | 3,800 | +16% | 1 | 1 | 0% | 524 | 1,036 | +98% | 0 | 0 | — |
case-17 | fail→fail | 7,144 | 3,737 | -48% | 1 | 1 | 0% | 1,120 | 1,045 | -7% | 0 | 0 | — |
case-18 | fail→fail | 4,222 | 6,373 | +51% | 1 | 1 | 0% | 639 | 1,504 | +135% | 0 | 0 | — |
case-19 | pass→pass | 5,350 | 7,746 | +45% | 1 | 1 | 0% | 975 | 1,804 | +85% | 0 | 0 | — |
case-20 | pass→pass | 32,229 | 26,151 | -19% | 1 | 1 | 0% | 5,494 | 4,696 | -15% | 0 | 0 | — |
case-21 | pass→pass | 3,518 | 3,196 | -9% | 1 | 1 | 0% | 651 | 955 | +47% | 0 | 0 | — |
case-22 | fail→fail | 2,204 | 2,692 | +22% | 1 | 1 | 0% | 324 | 833 | +157% | 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. The headline lift of +9 percentage points is the difference between those two pass rates over the 22 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.