ai deep research agent workflow

The problem

LLMs often hallucinate certainty, mix vendor claims with verified facts, and lose track of source provenance during complex research tasks. Developers need a systematic way to ensure AI-generated research is traceable, objective, and evidence-controlled rather than just a plausible-sounding summary.

What it does

• Enforces a strict evidence-control chain: CLAIM to SOURCE to EVIDENCE to INTERPRETATION.
• Classifies all information into specific buckets: FACT, OBSERVATION, INFERENCE, ASSUMPTION, or UNKNOWN.
• Resolves source contradictions by comparing publication dates, methodology, and scope instead of averaging them.
• Identifies research gaps and missing primary evidence rather than filling blanks with generic text.
• Generates structured comparison matrices and evidence tables with inline citation anchors.

Why this beats prompting it yourself

Generic prompts fail to maintain rigorous evidence classification over long contexts. This skill implements a systematic verification workflow that forces the model to audit its own sources, distinguish between documented capabilities and real-world performance, and preserve uncertainty when evidence is thin.

Use cases

• Technical architecture and vendor comparisons based on primary documentation.
• Verifying marketing claims against independent benchmarks and standards.
• Feasibility assessments for new product features or regulatory compliance.
• Detecting contradictions between conflicting API documentations or versions.

Known limitations

Does not provide its own search engine or web access. Accuracy depends entirely on the host environment's ability to provide source documents or real-time browsing capability.