Engineering / Field guide
Turn a team procedure into a reusable AI agent skill
A good first skill produces something your team already recognizes. A proposal checklist or an onboarding brief is easier to test than an instruction to be helpful.
2XD Studio · · 2 min read
Collect the finished work first
Choose one accepted example and one that was sent back for revision. Ask the reviewer what made the difference. Convert the answer into required inputs, decision rules, and a completion checklist. If an account number is mandatory, the skill should ask for it rather than manufacture a plausible value.
Separate references from instructions
Keep a small entry point that explains when the skill applies. Put long templates and sample outputs in referenced files. Assign an owner to each source that changes independently. This lets a team update the pricing appendix without rewriting the procedure for assembling a proposal.
Test the boundary of activation
Try a normal request, an incomplete request, and a similar task that belongs to another team. Check when the skill activates as well as what it writes. A technically perfect proposal template still wastes time if it opens every time a rep asks for a one-sentence meeting summary.
Compare two versions
When updating the onboarding template, run last month's test inputs through both skill versions. Compare required sections, missing-field behavior, and references. Ignore harmless punctuation changes and concentrate on differences a reviewer would act on. If the new version introduces an optional field, make sure its absence does not block a previously valid task.
Reader questions
When does a prompt become a skill?
Packaging becomes worthwhile when instructions, examples, and supporting files need to be reused and tested together.
How should changes be released?
Version the skill, rerun representative tasks, and have the process owner review differences in the outputs before rollout.
Further reading Agent Skills: how skills work