autonomous-investigation

autonomous-investigation (AI Autonomous Investigation Protocol)—Making AI Research Trustworthy, Rerunnable, and Comparable

Skill Overview

autonomous-investigation is a protocol that constrains AI autonomous research behavior. Through a bounded question budget, a search-plan gate, three levels of evidence labels—Fact / Inference / Assumption—and a no-fabrication checklist, it enables AI to complete credible research tasks without supervision, while producing structurally consistent outputs that can be compared across periods. It is commonly used for competitive intelligence and market intelligence research.

Applicable Scenarios

Automating quarterly competitor scans: Configure market or competitor research as agent tasks, loops, or scheduled tasks that can be rerun without supervision. Because the output structure remains consistent across runs, the current results can be directly diffed against the previous quarter, allowing the team to focus only on “what changed.”

Research requiring a distinction between facts and speculation: When AI research results support decision-making, every key conclusion carries a Fact / Inference / Assumption label. Facts include verifiable source URLs and dates, so you can immediately see which points should be verified first.

Cross-validating signals across multiple channels: When the same action—such as a competitor price increase, layoffs, or transformation—appears simultaneously across multiple intelligence channels, the confidence-stacking rules determine whether it is an observation, a working hypothesis, or actionable intelligence. When channels conflict, the protocol prompts: “Someone may be bluffing; this is worth investigating further.”

Core Features

Seven contractual provisions: Question budget (three by default; once the questions are answered or receive no response, continue with clearly labeled assumptions), search-plan gate (present three research plans before beginning; changing the plan costs far less than reworking the report), evidence labels, no-fabrication checklist, Just Enough Mode (optimize for “enough to support a decision” rather than page count), stable output schema, and a Final Step closing block (with four fixed next-step options). The seven provisions function as a whole and must not be selectively adopted.

Confidence-stacking rules and no-fabrication checklist: One channel signal = observation; two consistent channels = working hypothesis; three or more consistent channels = actionable intelligence. The protocol also emphasizes that “an announcement is merely intent; funding, hiring, procurement, and contracts count as commitments.” The no-fabrication checklist identifies content AI is most likely to invent in each field—such as competitors, prices, market share, patent details, and customer case studies—and requires that only real, accessible, dated sources be cited.

Stable, diffable output schema: The chapter order and structure must never drift across runs. This is the prerequisite for scheduled refreshes, incremental monitoring, and determining “what changed compared with the previous period.” The protocol also requires all intelligence gathering to rely solely on legal and ethical research of public sources, rejecting methods such as pretexting—using a false identity to obtain information.

Frequently Asked Questions

How does the protocol prevent AI from fabricating facts and citations during research?

The protocol includes a “no-fabrication checklist” provision: every research skill must identify the specific content AI is most likely to fabricate in that field and explicitly prohibit fabrication. All key conclusions must be supported by real, accessible, dated source URLs. Claims without source support may only be labeled Assumption and placed on the verification list. Information that cannot be found is added to an explicit gaps list rather than disguised as fact. This makes the entire output credible and verifiable.

What do Fact, Inference, and Assumption mean in AI research results?

These are the protocol’s three levels of evidence labels: Fact means the claim is supported by a source with a verifiable URL provided alongside it; Inference means an interpretation based on cited evidence—the evidence is referenced, while you decide whether the inference itself is valid; Assumption is a working hypothesis introduced to keep the research moving and is explicitly listed for subsequent verification. This classification helps separate “facts found by AI” from “AI’s speculation,” preventing inferences from being treated as facts in decision-making.

Can this protocol enable scheduled, unattended AI research tasks?

Yes. That is its design objective and what distinguishes it from interactive, one-question-at-a-time skills. The question budget ensures that, when no one responds, the research continues with clearly labeled assumptions instead of becoming blocked. The search-plan gate is automatically approved in unattended mode. The stable output structure ensures that the current run can be directly diffed against the previous period. During scheduled runs, no one selects the next-step options, so the output is archived and the process stops, waiting for the team to review the incremental changes.

> Scope note: This skill is itself a “protocol/specification” and must be referenced and implemented by specific research skills, such as market-landscape-scan, competitive-research-snapshot, competitive-intel-watch, and battle-card-builder. It does not directly produce reports; instead, it defines how these skills must ask questions, gather evidence, label findings, and structure outputs. All information gathering is limited to public sources and does not involve unauthorized scraping or the acquisition of confidential information.

Author

Category

Install