PLATFORMJuly 8, 2026 · 7 min read

We Taught Claude to Investigate
Political Trading Like We Do

Giving an AI access to 18 investigative tools is the easy part — our MCP connector does that in two minutes. The hard part is the craft: which tool answers which question, what the scores honestly mean, and — above everything — what you may and may not claim from a disclosure record. So we wrote the craft down as a Claude skill, ran it through three cold-chat evaluations against live production data, and are releasing it free. The evaluations passed. They also found two real defects in our own data, which is exactly the point.

A skill is a standing operating procedure

A Claude skill is a markdown instruction file Claude loads automatically when a task matches its description — the closest thing the platform has to handing a new analyst your desk manual. Ours opens with the rules, not the tools, because the rules are the product: Signal Congress reports sequence and co-occurrence in public records. It never asserts intent, causation, coordination, or non-public knowledge — and the skill forbids Claude from doing so on our data, no matter how suggestive a result looks. "Traded 17 days before the contract award," never "traded on advance knowledge." Amount ranges stay ranges. Clusters show co-occurrence, not coordination.

Below the trust boundary sit six playbooks — the tool-chaining sequences an experienced analyst would run:

P1
Investigate a member
Resolve the member, pull their baseline, check recent scored trades, measure acceleration against their own history — never against an imagined norm.
P2
Investigate a ticker
Congressional activity, then the corroboration picture (lobbying, contracts, FEC), then cross-branch overlap — presented as separate layers, never blended into one insinuation.
P3
Find the story
Journalist mode: leads where members traded inside their own committee’s jurisdiction, each with a plain-language why, delivered as a dated sequence.
P4
Compliance sweep
Late STOCK Act filers with the airtight phrasing built in: the statute’s 45-day ceiling vs. the unobservable awareness date, and edge cases excluded by rule.
P5
Market posture
What Congress is buying and selling by sector, with distinct-member counts headlined over raw row counts.
P6
Standing watch
Alert subscriptions on tickers, members, clusters, and compliance events.

Honest score interpretation, enforced

Scores are where AI-assisted analysis usually goes wrong, so the skill encodes how to read ours without folklore. A null anomaly score is benign, not suspicious — it means thin history or a non-equity, and the response says so. The high-conviction gate is 57 and the alert band is 70 because that's where the backtests put them, and the skill says to quote the calibrated bands, not remembered values. Convergence responses headline distinct officials and distinct tickers, never join fan-out row counts. And when broad-basket holders are excluded from a convergence answer, the skill requires saying so — surfacing the filter is what makes the remaining signal credible.

The citation rule is absolute: every factual claim must trace to a field in a tool response Claude actually received in that conversation. If a search comes back empty, the answer is "the search came back empty" — never a fill-in from remembered news coverage.

Three cold-chat evals — and what they caught

We validated the skill the way we validate detectors: fresh conversations, live production data, no coaching. All three runs held the playbook order, kept the claim rules under pressure (including a genuinely juicy cross-branch overlap that the run argued against overreading), and presented ranges as ranges. More interesting is what the runs found:

Eval 1 caught a typo in a real filing record
A cold-chat run over live data surfaced a filer title stored as “Untied States Attorney.” One-row fix the same evening — and proof the skill’s cite-what-you-received rule makes errors visible instead of smoothing them over.
Eval 2 caught a duplicate-ingestion defect
The skill’s discipline of separating distinct records from raw rows flagged same-content sibling rows in the trades corpus. That finding became a characterized pipeline issue with a standing data-quality guard — the number is now frozen and monitored every run.
Eval 3 taught us legal nuance back
Asked for a compliance sweep, the cold run independently derived the STOCK Act’s two-clock structure — 30 days from awareness, 45-day hard ceiling — and excluded edge-of-window rows unprompted. Both refinements were folded back into the skill.

A skill audit that finds bugs in the platform it documents is a feature of the method. The claim rules force Claude to look at what the data actually says — and data that gets looked at that closely gets cleaner.

Install it

The skill is free. Download the skill package, then in Claude go to Customize → Skills, hit the + button, and upload the ZIP. Make sure code execution and file creation are enabled in Claude's capability settings — skills require it. The skill drives the Signal Congress tools, so pair it with the MCP connector (two minutes, one URL). Tool depth follows your plan tier.

The skill's first rule bears repeating outside the skill: a signal means "worth a look," never "wrongdoing found." Legal trading with disclosure is the norm. What the skill guarantees is that when Claude presents something striking from our data, it presents the dated, sourced sequence — and lets it speak.
Signal Congress is research infrastructure, not investment advice. Everything the skill works with is derived from public government records — STOCK Act PTRs, OGE Form 278 filings, and federal corroboration sources — and every signal links back to its primary source.
← All postsThe connector how-to →Pricing →