leadterraWhat is a Claude Code skill?
A skill packages reusable instructions in a SKILL.md file. Claude Code can select a relevant skill or you can invoke it with its slash command. Anthropic documents the loading and invocation behavior.
For outbound, a skill might turn supplied company notes into a sourced account brief, audit a CSV, or review a draft email against approved claims. Those are distinct jobs with different evidence requirements.
Choose a task with an observable result
“Find good leads” is difficult to evaluate because it leaves the definition of good unstated. A stronger task is “compare these accounts with this buyer brief, cite the evidence, and mark unresolved criteria as unknown.”
- Specify the offer and audience in the input.
- Define the output fields before running the task.
- Explain which decisions require evidence and which may remain unknown.
- Keep research, drafting, and sending as separate tasks.
That last distinction lets you review the work before it changes a campaign. A company brief can suggest an angle without pretending the account has already expressed interest.
Give the skill a narrow structure
Use a folder named for the task, with SKILL.md as its entrypoint. The Agent Skills format includes name and description in YAML frontmatter, followed by the instructions. Read the format specification.
company-research/
└── SKILL.mdThe body should state the task, the evidence rules, and the deliverable. Supporting files are useful when they contain a real schema or example the review needs; a short skill can remain one file. Our company-research template uses a criteria table, unknowns, sources, and a proposed conversation angle.
Install one of the starter templates
Download the company-research skill, then save it as .claude/skills/company-research/SKILL.md in your project. Start Claude Code there and invoke /company-research with your offer, criteria, and source material.
The skill needs evidence to work from. Supply website excerpts or research notes, or use browsing tools available in your environment. It can report the limitations of supplied material; it does not make unavailable data appear.
Test the decisions with contrasting examples
Use a small set of inputs where you can check the answer yourself. Include a clear match, a clear mismatch, an account with incomplete information, and a source that does not establish the criterion you care about.
| Check | What to look for |
|---|---|
| Evidence | Each material fact points to the inspected source. |
| Uncertainty | Missing budget, revenue, or buying intent remains unknown. |
| Fit | The assessment follows your criteria rather than broad enthusiasm. |
| Output | The brief is usable without reconstructing the sources. |
If a polished answer fails those checks, fix the instruction responsible for the mistake. Re-run the same examples so the revision can be compared. These are suggested evaluation steps, not a claim that every model run will behave identically.
Three useful starting points
- Company research: separate sourced facts, account fit, and a proposed conversation angle.
- Lead-list review: preserve source rows while reporting missing fields and duplicate emails.
- Cold-email review: check relevance, proof, offer terms, placeholders, and the next step.
For the surrounding project setup, read CLAUDE.md vs skills vs MCP. For a managed campaign, explore Leadterra’s cold email service.