Research Reports Your Agent Writes Need Sources Before Sentences

When you request a research report from a coding agent, you often receive well-written prose first and evidence later, if at all. A single forum post might be presented as "research shows," and a 2023 number is taken as the current state without question. The author's beliefs are obscured by the presentation, not clarified by the writing.

The issue lies in the process, not the style: the writing began without a defined question and unvetted sources. To address this, I documented the workflow as a skill, starting with the premise that the quality of research is determined by its sources, not its prose. The workflow's first step mandates defining the report's question and including up to three sub-questions before any search occurs.

This ensures the research stays focused on the question, preventing the agent from collecting random information and shaping the report based on search order rather than the question. The second step involves tiering the sources before accepting them as evidence. Primary sources are official documents, original papers, filings, and measured data.

Secondary sources include reputable press, industry reports, and expert interviews. Tertiary sources, such as social-media posts, forum threads, and aggregator sites, are listed last and only serve to lead to other sources, never as evidence. A crucial rule is that no tertiary source can support a core conclusion. If no second independent source exists, the conclusion is downgraded to a single-source claim and marked accordingly, rather than being subtly presented as more substantial than it is.

The skill enforces that every number in the report must have a source, and numbers without sources are deleted. Each key fact is dated, with undated information treated as possibly stale by default. The writing structure begins with a conclusion, followed by uncertainty, ensuring honesty about what isn't known. This approach prioritizes stating what is unknown before asserting what is known, avoiding the common pitfall of presenting confident statements without evidence.

The skill also requires uncertainty statements at the end, detailing what wasn't found, what is speculative, and conditions under which conclusions may change. This approach is part of a suite of MIT-licensed skills, including others for commit messages, code reviews, meeting minutes, and technical proofreading. Each skill is a single SKILL.md file, easily copied into an agent's skills directory to apply the workflow without further instruction.

The research skill is located at https://github.com/alapha888/agent-skills-en under skills/deep-research-framework/, complete with the full workflow and an anti-pattern list to help maintain integrity in research reports. If your organization has specific source preferences, the skill can be forked and adapted to include those preferences, ensuring the structure remains stable while sources can be customized.