Anatomy of Claude Code deep-research skill in 349 lines

The exploration began with a surprise when running deep-research, a built-in skill in Claude Code, which returned an unexpected comprehensive report from a single run. It processed 27 sources, extracted 123 claims, verified 25 of them, confirming 18 and refuting 7, alongside an executive summary, caveats, and open questions. Impressed that all of this originated from a single skill execution, the developer set out to examine how it was constructed.

Finding the source code required looking inside the Claude Code binary where it is compiled as a bundled workflow, originating from the bughunter architecture. Spanning just 349 lines of JavaScript, reading through the implementation served as the core research process.

Inside the code, there are three distinct prompts, all sharing a uniform structure: a role in the title, context combining the original question and specific input, a numbered checklist for tasks, explicit decision criteria, and an output format. The verifier stands out by enforcing deep skepticism, requiring at least two out of three refutations to kill a claim, and defaulting to refuted when uncertainty arises.

A skill proves to be much more than just a long prompt. The architecture breaks down into six key components: trigger metadata, tuning constants at the top, a dedicated JSON schema per agent, prompts instantiated as functions, explicit orchestration with deliberate synchronization barriers, and defensive design to handle early exits gracefully.