"Just ask ChatGPT"
is how cases get lost.
Vanilla LLMs are pattern matchers with a confident voice. That's a great combination for a brainstorm and a terrible one for an investigation. The fix isn't a smarter model — it's training. Here's exactly where an untrained model breaks, and how a Skill file trains each break out.
It hallucinates the things you can't make up
Court citations. Statute numbers. Case names. Document IDs. Vanilla LLMs invent them in a tone of complete confidence, because their training rewards plausibility, not provenance.
Confident answer, fabricated case
Cites People v. Reyes (2018), 32 Cal.App.5th 122. That case does not exist. Quotes a passage with a specific page number. The page number is also fabricated. A junior attorney drops the citation into a brief.
Grounded answer, real authorities, hedged
Skills jones-and-gps-tracking-standards + reasonable-expectation-of-privacy-framework cite United States v. Jones (565 U.S. 400, 2012) and link to oyez.org and the slip opinion. They name California's Electronic Communications Privacy Act (Cal. Penal Code §§ 1546–1546.4) and link to the statute. Where state law is unsettled, they say so explicitly.
It has no tradecraft discipline
Real investigators don't just answer the question — they document the path. Source provenance, chain of custody, deconfliction, OPSEC, redaction. Vanilla LLMs skip all of it because nothing in their training prompts them to care.
Synthesized prose with no audit trail
Returns a polished paragraph that mixes verified facts, AI guesses, and old information into a single tone. No timestamp on any finding. No URL. No 'this was current as of'. No note about which jurisdictions were excluded.
Methodical report with provenance baked in
Skills investigative-report-structure + evidence-citation-format + chain-of-custody-documentation + source-vetting-and-reliability-grading produce a report with: an executive summary, a methodology section, a numbered finding list with per-finding source URL + timestamp + reliability grade (A1–F6), an excluded-sources appendix, and an OPSEC log of what was probed and how.
It has no platform-specific tradecraft
LinkedIn's 'people you may know' graph is an inference attack. Telegram channel discovery isn't search — it's pivoting through forwarded messages. Snapchat's ephemeral content has a 24-hour preservation window. Vanilla LLMs give you a generic 'check their profile' for all of them.
Surface-level profile read
Summarizes the profile's About section, job history, and skills. Maybe mentions 'consider checking their connections'. Doesn't know about the pivoting techniques, the metadata leakage, or the platform-specific search syntax.
Platform-aware deep read
Skill linkedin-investigation walks the agent through: PYMK graph inference, public activity feed scraping, recommendation network analysis, profile-photo reverse search, employment-record cross-checks against corporate-filings-research, and the LinkedIn-specific Boolean syntax (NEAR/3 etc.) — all stopping at the TOS and CFAA line.
It treats jurisdictions as interchangeable
DMV records in California are restricted by the DPPA. In Texas, same DPPA — different state-level access rules. A wiretap in NY needs Title III. In Florida, two-party consent applies to recording. Vanilla LLMs collapse all of this into one generic answer.
Generic 'check with your local DMV' answer
Says DMV records are sometimes available. Suggests 'consider contacting the DMV directly'. Doesn't mention DPPA's permissible-purpose list. Doesn't distinguish state-level access regimes. Doesn't surface FCRA/GLBA implications for vendor-pulled data.
State-aware, statute-grounded, vendor-aware
Skill dmv-and-motor-vehicle-records-by-state + restricted-records-driver-privacy-protection-act + fcra-glba-compliance-for-investigators walk the agent through: DPPA permissible purpose check, state-specific access rules (CA, TX, FL, NY, etc.), licensed-investigator vs ordinary citizen distinctions, vendor sourcing (Accurint, TLO, IRBsearch) with FCRA flags, and how to document the legal basis for each pull.
See it work. The Starter bundle is free.
4 hand-picked skills across OSINT, forensics, legal, and tradecraft. Install in two minutes. Decide if the methodology is worth a paid volume.