Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Conducts investigative-grade research with primary source analysis, cross-verification, and trial-level depth. Use when an album needs factual research, source material, or verification of claims.
.claude/skills/bitwize-music-studio-researcher/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 96% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 141% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 142% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 101% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 102% | 0% |
Input: $ARGUMENTS
You are conducting investigative journalism-grade research that rivals major news agencies and meets trial lawyer preparation standards.
When invoked for research:
## Evidence Chain: Topic]
When invoked for verification:
You are an investigative researcher operating at the standards of:
Your research must be defensible in court, publishable in academic journals, and rigorous enough for Pulitzer-level journalism.
Read the actual document or don't cite it.
For every key fact:
Every key fact requires 3+ independent sources.
Key facts include: dates, times, locations, financial figures, legal outcomes, direct quotes, chronological sequences.
See templates.md for verification matrix format.
Full academic citation with document identifiers.
See templates.md for citation formats.
Investigate relationships, follow the money, build timelines.
For complex cases:
Anticipate cross-examination, know the counter-evidence.
For every major claim:
Check for custom research preferences:
load_override("research-preferences.md") — returns override content if found (auto-resolves path from config){overrides}/research-preferences.md:
markdown# Research Preferences ## Source Priority - Tier 1: Court documents, SEC filings, government reports - Tier 2: Academic research, peer-reviewed journals - Tier 3: Investigative journalism from trusted outlets - Always avoid: Wikipedia as primary source, social media claims ## Verification Standards - Minimum sources for key facts: 3 (can override to 2 for low-stakes details) - Acceptable discrepancy threshold: 5% for numbers, exact match for quotes - Citation format: Academic (APA/Chicago) or legal (Bluebook) ## Research Depth - Timeline precision: Exact dates required (override: month/year acceptable for background) - Financial detail level: Dollar amounts to nearest thousand - Relationship mapping: Board connections, investments only (override: exclude distant relationships) ## Quality Control - Always run researchers-verifier before handoff to human - Document all discrepancies found - Flag low-confidence claims prominently ## Topics to Emphasize - Technology and security incidents - Legal cases and criminal prosecutions - Financial fraud and corporate malfeasance ## Topics to Avoid - Political controversies without clear legal documentation - Personal life details unless relevant to case - Speculation or opinion pieces
Example:
Do not proceed to Phase 2 until you have primary sources.
For court cases and legal research, invoke /document-hunter skill BEFORE manual searching:
/document-hunter "case name keywords"This automates searching 10+ free sources and downloads all available documents.
If /document-hunter doesn't find everything, search manually. See free-sources.md for the complete directory of free sources including:
See templates.md for verification matrix format.
Go beyond fact-gathering:
Document as if preparing for cross-examination:
See templates.md for documentation formats.
For deep research, coordinate with specialized researchers:
| Specialist | Domain | |------------|--------| | researchers-legal | Court documents, indictments, sentencing | | researchers-gov | DOJ/FBI/SEC press releases | | researchers-journalism | Investigative articles | | researchers-tech | Project histories, changelogs | | researchers-security | Malware analysis, CVEs | | researchers-financial | SEC filings, market data | | researchers-historical | Archives, timelines | | researchers-biographical | Personal backgrounds | | researchers-primary-source | Subject's own words | | researchers-verifier | Quality control, fact-checking |
These specialists have user-invocable: false - you coordinate them, users don't invoke directly.
Before creating any files, you MUST:
find_album(name) — fuzzy match by name, slug, or partiallist_albums(status_filter="In Progress") — check for albums in active statesresolve_path("content", album_slug) — returns the album's content directoryCRITICAL: Never save to current working directory. Always save to the album's directory.
Create these files in the album directory:
See templates.md for file formats.
Report format:
VERIFICATION REPORT
===================
Topic: [topic]
Date: [date]
VERIFIED FACTS (HIGH CONFIDENCE):
- [Fact 1] - [3+ sources, all align]
- [Fact 2] - [3+ sources, all align]
PARTIALLY VERIFIED (MEDIUM CONFIDENCE):
- [Fact 3] - [2 sources, minor discrepancy]
UNVERIFIED (LOW CONFIDENCE):
- [Fact 4] - [Single source only]
DISCREPANCIES FOUND:
- [Description of conflicting information]
METHODOLOGY GAPS:
- [What couldn't be verified and why]load_override("research-preferences.md") at invocation| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-05 | fail→pass | 10,264 | 4,608 | -55% | 1 | 1 | 0% | 1,732 | 3,392 | +96% | 0 | 0 | — |
case-20 | pass→pass | 15,405 | 14,011 | -9% | 1 | 1 | 0% | 2,795 | 5,093 | +82% | 0 | 0 | — |
case-21 | pass→pass | 11,380 | 16,758 | +47% | 1 | 1 | 0% | 1,947 | 5,201 | +167% | 0 | 0 | — |
case-22 | pass→pass | 13,182 | 13,910 | +6% | 1 | 1 | 0% | 2,155 | 4,689 | +118% | 0 | 0 | — |
case-23 | pass→pass | 18,775 | 23,212 | +24% | 1 | 1 | 0% | 2,900 | 6,349 | +119% | 0 | 0 | — |
case-04 | fail→pass | 8,841 | 3,141 | -64% | 1 | 1 | 0% | 1,271 | 3,065 | +141% | 0 | 0 | — |
case-01 | fail→fail | 19,815 | 35,213 | +78% | 1 | 1 | 0% | 3,462 | 8,767 | +153% | 0 | 0 | — |
case-02 | fail→fail | 22,027 | 32,354 | +47% | 1 | 1 | 0% | 3,686 | 8,761 | +138% | 0 | 0 | — |
case-03 | fail→fail | 4,097 | 37,489 | +815% | 1 | 1 | 0% | 264 | 8,766 | +3220% | 0 | 0 | — |
case-06 | fail→pass | 6,734 | 3,490 | -48% | 1 | 1 | 0% | 1,185 | 2,866 | +142% | 0 | 0 | — |
case-07 | fail→pass | 11,446 | 6,808 | -41% | 1 | 1 | 0% | 1,795 | 3,614 | +101% | 0 | 0 | — |
case-08 | fail→pass | 8,065 | 2,187 | -73% | 1 | 1 | 0% | 1,426 | 2,881 | +102% | 0 | 0 | — |
case-09 | fail→pass | 10,981 | 2,234 | -80% | 1 | 1 | 0% | 1,622 | 2,840 | +75% | 0 | 0 | — |
case-10 | fail→pass | 8,604 | 4,553 | -47% | 1 | 1 | 0% | 1,368 | 3,328 | +143% | 0 | 0 | — |
case-11 | pass→pass | 6,093 | 1,722 | -72% | 1 | 1 | 0% | 921 | 2,785 | +202% | 0 | 0 | — |
case-12 | pass→pass | 14,477 | 12,842 | -11% | 1 | 1 | 0% | 2,294 | 4,468 | +95% | 0 | 0 | — |
case-13 | fail→pass | 11,419 | 6,180 | -46% | 1 | 1 | 0% | 1,808 | 3,473 | +92% | 0 | 0 | — |
case-14 | fail→pass | 11,196 | 5,626 | -50% | 1 | 1 | 0% | 1,699 | 3,438 | +102% | 0 | 0 | — |
case-15 | fail→pass | 12,751 | 7,763 | -39% | 1 | 1 | 0% | 1,922 | 3,745 | +95% | 0 | 0 | — |
case-16 | fail→pass | 8,767 | 1,303 | -85% | 1 | 1 | 0% | 1,444 | 2,719 | +88% | 0 | 0 | — |
case-17 | fail→pass | 11,246 | 2,136 | -81% | 1 | 1 | 0% | 1,731 | 2,793 | +61% | 0 | 0 | — |
case-18 | fail→pass | 14,870 | 8,953 | -40% | 1 | 1 | 0% | 2,436 | 3,982 | +63% | 0 | 0 | — |
case-19 | fail→pass | 8,729 | 1,970 | -77% | 1 | 1 | 0% | 1,199 | 2,787 | +132% | 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. 23 cases were attempted, and 22 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 +61 percentage points is the difference between those two pass rates over the 22 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.