autonomous-investigation skill

Autonomous Investigation Protocol

Purpose

Provide the canonical contract for investigation skills — research the AI performs in the world (web search, published data, public filings) while you review the evidence instead of feeding it context. Where workshop-facilitation governs skills that ask you questions one at a time, this protocol governs skills that proceed without you: they budget their questions, show their plan, label every claim, and produce output stable enough to diff against last quarter's run. That last property is the payoff — an investigation honoring this contract can run as an agent task, in a loop, or on a schedule.

Input

Nothing required — this skill defines the protocol other investigation skills follow. Also useful when invoked standalone: the target of the investigation and, above all, the decision the research should support. Research without a decision is a hobby; every investigation skill asks for the decision because it determines what "just enough" means.

Anything supplied with the invocation itself — text after the skill name, a pasted context dump, or an appended ARGUMENTS: line — counts as answers already given. Use it, credit it against the question budget, and don't re-ask.

Arriving empty-handed? That works too. The protocol's whole design is to proceed on best-available evidence with labeled assumptions when nobody answers questions. When another skill references this protocol, that skill's Input section governs what to provide.

Example invocation: Run an autonomous investigation on [TARGET]'s move into workflow automation — this supports our Q3 roadmap bet on the same space.

Key Concepts

Two protocols, two jobs

The contract

Every investigation skill honors all seven clauses. They are not a menu.

nobody answers, proceed with labeled assumptions. This is what makes investigations schedulable: an unattended run degrades gracefully instead of stalling.

1. Question budget — a hard cap (usually 3) on clarifying questions. When the budget is spent or

types, how you'll separate fact from inference. Continue unless the user revises it. Why it teaches: reviewing a plan takes 10 seconds; reviewing a wrong report takes 10 minutes. The gate is the cheapest correction point in the whole workflow.

1. Search-plan gate — before researching, show a 3-bullet plan: what you'll search, which source

1. Evidence labels — every key claim carries exactly one label:

Keep labels short. Things you couldn't find are not a fourth label — they go in an explicit gaps list. Why it teaches: most competitive "facts" in strategy decks are unlabeled inference. Three-level honesty is the habit that separates intelligence from confident storytelling.

• Fact — source-supported; a checkable URL sits next to it
• Inference — evidence-based interpretation; the evidence is cited, the leap is yours to judge
• Assumption — working guess made to keep moving; listed for validation

(competitors, pricing, market share, patent contents, customer wins...) and forbids inventing them. Real, checkable URLs only; a claim without a source and date is an opinion wearing a badge. Why it teaches: the list tells the human exactly what to verify first.

1. Do-not-invent list — each investigation skill names its domain's specific fabrication risks

decision. Verbose Mode exists only on request. Research value is decision support, not page count.

1. Just Enough Mode — default output is the strongest findings in short bullets, sized to the

run N+1 are diffable. Delta monitoring, scheduled refreshes, and "what changed since last quarter" all depend on this clause.

1. Stable output schema — section order and structure never drift between runs, so run N and

to run, assumptions to validate). Accept 1, 1 and 3, Verbose Mode, or a custom path.

1. Final Step block — end with exactly 4 numbered next options (artifacts to build, deeper passes

The Confidence Stacking Rule

Labels grade individual claims; stacking grades the story. When signals arrive from independent collection channels (see intelligence-collection-disciplines):

~~~ 1 channel flags it → Watch item. Log it, do nothing. 2 channels agree → Working hypothesis. Assign someone to probe. 3+ channels agree → Actionable intelligence. Brief leadership, adjust plans. Channels conflict → The most interesting case. Someone is bluffing. Dig. ~~~

One corollary that generalizes everywhere: treat announcements as intent until funding, procurement, hiring, or contracts corroborate them. Ambition shows up in press releases; commitment shows up in filings, job posts, and purchase orders.

Guardrails

All collection under this protocol is legal, ethical, open-source work:

specifically to extract a former employer's secrets, scraping in violation of terms you accepted.

• Yes: anything published, filed, posted, or observable in public.
• No: pretexting (lying about who you are), soliciting NDA-protected information, hiring someone

The rule of thumb, borrowed from the competitive-intelligence profession (SCIP Code of Ethics): if you'd be uncomfortable explaining your method on stage at the target's user conference, don't use the method.

Application

For skills implementing this protocol

1. Declare this skill in References as the governing protocol.
2. State the skill's question budget (default 3) and the questions themselves.
3. Define the domain's do-not-invent list — name the specific things AI fabricates in this territory.
4. Define the stable output schema with numbered sections; mark it "do not reorder."
5. End the schema with a Final Step block of exactly 4 options.

For agents running an investigation

Assumption.

1. Read inline invocation context first; credit it against the question budget.
2. Ask only unanswered budget questions. If silence, proceed — label every gap-filling guess

not just the signals.

1. Show the 3-bullet search plan. Continue unless revised.
2. Research in Just Enough Mode: mixed source types, real URLs captured with dates.
3. Label every key claim Fact / Inference / Assumption. Put what you couldn't find in a gaps list.
4. Apply confidence stacking when multiple channels speak to the same move; report the stack level,

scheduled run), file the output and stop.

1. Emit the skill's schema exactly — same sections, same order — so this run diffs against the last.
2. Close with the Final Step block. If the user picks a number, execute; if they answer nothing (a

A copy/paste investigation brief — the contract's seven clauses as fill-in decisions, for briefing an agent or designing a new investigation skill — lives in template.md.

Examples

Opening of a protocol-honoring run (user gave target + decision inline, so no questions spent):

Search plan (say "revise" to change it):

- Search [TARGET]'s pricing pages, release notes, and last two earnings transcripts

- Source mix: company site, filings, credible press, review sites

- Facts get URLs; interpretations get labeled Inference; gaps become Assumptions to validate

(research happens)

Key finding: [TARGET] removed its mid-tier plan in May — Fact

(pricing page diff, May 12). Packaging is consolidating toward

enterprise — Inference (tier removal + two enterprise-only features shipped since April).

They will raise the entry price within two quarters — Assumption (pattern-based; validate

against their next pricing-page change).

Final Step — reply 1, 2, 3, 4, a combination, or "Verbose Mode":