Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Research current topics with multiple sources and produce a structured brief, comparison, recommendation, or fact-check. Use when the user asks for investigation, market/product landscape scans, option evaluation, due diligence, source-backed validation, or a research report. Do not use for summarizing a single provided URL/document, or for GitHub issue/PR operations.
.claude/skills/understudy-ai-researcher/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✓→✗ | ▼ Worse | 4% | 0% |
| case-05 | ✓→✗ | ▼ Worse | 7% | 0% |
| case-06 | ✓→✗ | ▼ Worse | -69% | 0% |
| case-07 | ✓→✗ | ▼ Worse | -23% | 0% |
| case-09 | ✓→✗ | ▼ Worse | -6% | 0% |
Use this skill for bounded, source-backed research.
Default goal: turn an open-ended question into a concise research output with explicit evidence, tradeoffs, and uncertainty.
Use this skill when the user wants any of:
Do not use this skill for:
Prefer current sources over memory. Use a small search budget first, then expand only if the evidence is weak or conflicting.
Unless the user already gave a narrow format, produce:
If the user asks for a persistent artifact, write a Markdown report under research/ with a short kebab-case filename that matches the topic.
Follow these phases in order.
Before searching, extract or infer:
If one missing detail would materially change the answer, ask a short clarifying question. Otherwise proceed with a stated assumption.
Break the work into 3-7 subquestions. Keep them concrete and decision-relevant.
Examples:
Start with a tight first pass:
Avoid aimless searching. Stop when additional searches are no longer changing the answer.
When possible, prioritize:
Use secondary summaries only to discover leads, not as the sole basis for important conclusions.
For each important claim:
Call out any inference you are making from the evidence instead of presenting it as a confirmed fact.
Keep the final answer structured and useful. Include:
Use this shape by default:
Use this shape:
Use this shape:
Prefer web_search to discover candidates, web_fetch to read exact page contents, and pdf when a primary source is a PDF.
Use the browser only when a relevant source requires interactive navigation, login, or a page state that the normal web tools cannot reach.
Do not end with a pile of links. Synthesize.
Do not present stale or weakly supported claims as settled.
Do not hide uncertainty. If the evidence is thin, say so clearly and narrow the recommendation.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 24,541 | 6,700 | -73% | 1 | 1 | 0% | 4,041 | 1,417 | -65% | 0 | 0 | — |
case-02 | fail→fail | 35,006 | 7,282 | -79% | 1 | 1 | 0% | 5,946 | 1,309 | -78% | 0 | 0 | — |
case-03 | fail→fail | 14,080 | 11,812 | -16% | 1 | 1 | 0% | 2,137 | 1,596 | -25% | 0 | 0 | — |
case-04 | pass→fail | 8,304 | 20,100 | +142% | 1 | 1 | 0% | 1,181 | 1,226 | +4% | 0 | 0 | — |
case-05 | pass→fail | 7,333 | 5,032 | -31% | 1 | 1 | 0% | 1,111 | 1,187 | +7% | 0 | 0 | — |
case-06 | pass→fail | 32,542 | 9,406 | -71% | 1 | 1 | 0% | 4,964 | 1,524 | -69% | 0 | 0 | — |
case-07 | pass→fail | 12,095 | 7,884 | -35% | 1 | 1 | 0% | 1,754 | 1,356 | -23% | 0 | 0 | — |
case-08 | fail→fail | 16,871 | 18,557 | +10% | 1 | 1 | 0% | 2,515 | 2,896 | +15% | 0 | 0 | — |
case-09 | pass→fail | 8,061 | 13,980 | +73% | 1 | 1 | 0% | 1,282 | 1,203 | -6% | 0 | 0 | — |
case-10 | pass→fail | 18,603 | 8,275 | -56% | 1 | 1 | 0% | 2,745 | 1,267 | -54% | 0 | 0 | — |
case-11 | fail→fail | 19,673 | 8,612 | -56% | 1 | 1 | 0% | 3,129 | 1,427 | -54% | 0 | 0 | — |
case-12 | pass→fail | 8,260 | 30,864 | +274% | 1 | 1 | 0% | 1,330 | 1,377 | +4% | 0 | 0 | — |
case-13 | pass→fail | 26,923 | 12,334 | -54% | 1 | 1 | 0% | 3,837 | 1,477 | -62% | 0 | 0 | — |
case-14 | pass→fail | 23,860 | 7,040 | -70% | 1 | 1 | 0% | 3,598 | 1,317 | -63% | 0 | 0 | — |
case-15 | fail→fail | 18,955 | 9,300 | -51% | 1 | 1 | 0% | 2,679 | 1,374 | -49% | 0 | 0 | — |
case-16 | pass→fail | 6,578 | 16,184 | +146% | 1 | 1 | 0% | 898 | 1,178 | +31% | 0 | 0 | — |
case-17 | fail→fail | 41,716 | 8,833 | -79% | 1 | 1 | 0% | 6,169 | 1,387 | -78% | 0 | 0 | — |
case-18 | pass→fail | 21,528 | 8,472 | -61% | 1 | 1 | 0% | 3,292 | 1,350 | -59% | 0 | 0 | — |
case-19 | pass→fail | 7,465 | 6,905 | -8% | 1 | 1 | 0% | 1,165 | 1,179 | +1% | 0 | 0 | — |
case-20 | pass→fail | 21,728 | 8,962 | -59% | 1 | 1 | 0% | 3,429 | 1,412 | -59% | 0 | 0 | — |
case-21 | pass→fail | 25,083 | 11,846 | -53% | 1 | 1 | 0% | 3,821 | 1,431 | -63% | 0 | 0 | — |
case-22 | pass→fail | 10,385 | 16,631 | +60% | 1 | 1 | 0% | 1,641 | 1,494 | -9% | 0 | 0 | — |
case-23 | fail→fail | 37,936 | 9,951 | -74% | 1 | 1 | 0% | 6,077 | 1,621 | -73% | 0 | 0 | — |
case-24 | fail→fail | 2,867 | 8,373 | +192% | 1 | 1 | 0% | 418 | 2,205 | +428% | 0 | 0 | — |
case-25 | fail→fail | 3,011 | 6,095 | +102% | 1 | 1 | 0% | 396 | 1,846 | +366% | 0 | 0 | — |
case-26 | fail→fail | 2,598 | 1,790 | -31% | 1 | 1 | 0% | 233 | 1,176 | +405% | 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. 26 cases were attempted, and 4 counted toward the lift figure. The other 22 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 -58 percentage points is the difference between those two pass rates over the 4 comparable cases. 20 cases got worse with the skill loaded, and they are included in that figure.
Other measured skills in the registry, with their headline benchmark lift.