The short version

Design AI skills as reusable workflow assets, not as long one-off prompts.

Editorial analysis

The useful lesson is not a single prompt formula. It is a way to reduce ambiguity: state the outcome, supply only the context that changes the answer, demonstrate the pattern when format matters, and test whether the result remains reliable across several inputs.

A useful way to read this study is as a decision guide: identify the problem it solves, the conditions where it works, the tradeoffs it introduces, and the evidence you would need before relying on it.

Source context

Date

July 1, 2026

Blog post covered

Lessons from building Claude Code: How we use skills

https://claude.com/blog/lessons-from-building-claude-code-how-we-use-skills

Key concept

Design AI skills as reusable workflow assets, not as long one-off prompts.

Why it matters

The post explains that skills have become one of the most used extension points in Claude Code. Anthropic describes skills as folders of instructions, scripts, resources, assets, data, examples, and configuration that agents can discover and use to work more accurately and efficiently.

This matters because repeated AI work should not rely on someone remembering the perfect prompt every time. If a workflow matters, it should be packaged into a reusable skill with trigger guidance, gotchas, templates, and verification.

How it works

A prompt is a one-time instruction. A skill is a reusable operating manual for Claude.

A good skill answers:

  • When should Claude use this?
  • What inputs are needed?
  • What steps should Claude follow?
  • What examples, templates, scripts, or reference files are available?
  • What mistakes should Claude avoid?
  • How should the result be checked?

The important shift is from asking Claude to remember everything in the chat to giving Claude a small reusable workflow folder it can inspect when the task appears.

Where it matters

Finance

Create a variance memo skill with metric definitions, source tables, commentary template, known edge cases, and reviewer checklist.

Create a contract review skill with clause playbook, risk categories, fallback questions, redline template, and escalation rules.

Research

Create a source synthesis skill with search rules, citation standards, contradiction checks, and summary templates.

Productivity

Create a weekly recap skill that reads prior logs, tickets, and docs, then produces delta-only updates.

Implementation examples

Weak example

“Review this contract carefully and check important risks.”

Why it is weak:

  • Too vague
  • No risk taxonomy
  • No examples
  • No review checklist
  • No known gotchas
  • No standard output format

Strong example

“Use the contract review skill. Apply the clause matrix, risk rubric, fallback questions, redline template, and final checklist. State assumptions and escalate missing information.”

Why it is strong:

  • Clear trigger
  • Reusable assets
  • Known failure points
  • Standardized output
  • Built-in review process

Weak-vs-strong example

Weak skill design:

  • One long markdown instruction
  • Repeats obvious advice
  • No examples
  • No templates
  • No verification
  • No memory of past runs

Strong skill design:

  • Clear trigger description for the model
  • Short core instructions
  • Separate reference files for detailed context
  • Gotchas section based on real failures
  • Templates for final deliverables
  • Scripts or checklists for verification
  • Optional log/history file for recurring workflows

Implementation checklist

  • Use skills for recurring workflows.
  • Do not restate what Claude already knows.
  • Add a high-signal gotchas section.
  • Use progressive disclosure with reference files.
  • Include templates and scripts where useful.
  • Add verification when output quality matters.
  • Update the skill when Claude fails or the process changes.

Try it in practice

Pick one task you repeat weekly and draft a mini skill card:

  • Trigger: When should Claude use this skill?
  • Inputs: What files, links, systems, or context are needed?
  • Steps: What process should Claude follow?
  • Gotchas: What mistakes happen often?
  • Output: What should the final deliverable look like?
  • Verifier: How should Claude check the work?