Anatomy of a skill
This started with a surprise, not with a plan. I needed a survey of the state of the art on a topic I was considering working on, so I ran deep-research , a skill that ships with Claude Code, and it came back with a report I didn't expect from a single run —27 sources, 123 claims extracted, 25 verified, and of those 18 confirmed and 7 refuted— with an executive summary, caveats, and open…
The story begins with a surprising discovery. The author used a deep-research skill, provided by Claude Code, to examine the state of the art on a topic he was considering working on. The skill produced 27 sources, 123 claims extracted, 25 verified, 18 confirmed, 7 refuted, an executive summary, caveats, and open questions. The author was intrigued not by the volume of information but by the fact that all of it came out of a single skill.
The skill, which is not easily found within the .claude/skills/ or ~/.claude directories, is instead compiled into the Claude Code binary as a bundled workflow. The skill consists of 349 lines of JavaScript. The author spent time examining the code, specifically the prompts that make up the skill.
There are three prompts within the skill, each following a consistent structure: a role in the title, the context (original question plus specific input), a task as a numbered checklist, an explicit decision criterion, and an output format. The verifier prompt is particularly clear, guiding the model to question the claim and providing specific steps to determine its validity. The verifier follows a rule-based approach, defaulting to a refuted status if the evidence is insufficient.
Breaking down the skill into its six components reveals that it is more than just a long prompt. There's trigger metadata, tuning constants, one schema per agent, prompts as functions, explicit orchestration, and defensive design. The orchestration moves seamlessly from search to fetch, ensuring that each angle progresses without waiting for the others. The skill also includes a barrier to ensure the pool of claims is complete before ranking can occur.
Finally, the skill incorporates defensive design measures, such as early exits and a cost calculation for agents. The author demonstrates their understanding of the pattern by reusing it in subsequent runs, creating a more efficient version of the original skill with the same outcome. The takeaway is that a good skill is not a lengthy prompt but a concise, well-structured prompt that can be instantiated multiple times through an efficient orchestration, measuring its costs along the way.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
