What a Skill is, in plain English.
A Skill is a written procedure with one new property: a robot follows it perfectly every time. Before you build one, get the mental model clear.
Skill versus prompt versus Project
- Prompt. One conversation, one ask. Lives in your head or a Notes file. Dies when the conversation closes.
- Project. A workspace with persistent instructions and a reference library. The instructions apply to everything you do in that Project.
- Skill. A reusable instruction set Claude loads on demand when the task matches. The instructions apply across Projects, across teammates, across months.
Why Skills compound
Most AI investment depreciates. Models change, prompts drift, the team forgets which one worked. Skills compound. Every Skill the firm writes raises the floor on what the next junior hire produces on day one. Every revision gets baked in for everyone, instantly.
A Skill is a written procedure with one new property. A robot follows it perfectly every time. If you have a process you keep re-explaining to new hires, that is a Skill in waiting.
Pick your first three Skills.
Three is the right number. Fewer and you do not feel the leverage. More and you stall in scope. Use the pay-back-in-30-days filter to choose.
The pay-back-in-30-days filter
A good first Skill meets four tests: someone runs the process at least monthly, the firm has opinions on how it should be done, the output has a clear shape, and the senior person is the bottleneck.
Worked examples by role
- Operations. Submittal review, weekly status memo, RFI triage.
- BD. Go / no-go on a bid, prospect research brief, RFP qualification questions.
- Marketing. Case study draft, LinkedIn post in the principal's voice, proposal cover letter.
- Estimating. Bid leveling abstract, scope-gap checklist, qualification letter.
Build your first Skill from scratch.
A Skill is a folder with a SKILL.md file and any helpers. Twenty-five minutes from blank file to working Skill.
SKILL.md anatomy
Common mistakes
- Writing the description like marketing copy. Claude uses it to decide when to load the Skill, so it should describe the trigger, not sell the value.
- No examples. Instructions tell Claude what to do. Examples tell Claude how good looks.
- Trying to capture every edge case. Ship the Skill at 80 percent. The next ten edits are easier than the first one.
The first useful answer in 30 seconds test
Run the Skill cold against a real input from this week. If the first useful answer arrives inside 30 seconds, the Skill is ready to deploy. If it does not, the description, the examples, or the template is wrong. Fix the one closest to the failure.
Get the team actually using it.
A Skill that nobody uses is worse than no Skill. Adoption is the hard part. Fifteen minutes on the deployment patterns that actually stick.
The kickoff conversation
Three AEC use cases
The patterns where a Skill pays back fastest in AEC firms.
Manage the library over time.
A Skills library is a living artifact. Versioning, ownership, deprecation. The honest trade-off between building yourself and bringing in POLR.
Ownership and versioning
- One named owner per Skill. The person whose process it encodes.
- A short changelog at the bottom of every SKILL.md. Date, change, reason.
- A quarterly review. Skills that nobody ran in 90 days get archived, not patched.
Build it yourself or bring in POLR
Build it yourself when the process lives in one person's head and that person has 90 minutes to write it down with you. Bring in POLR when the process spans three teams, the firm has been re-explaining it for years, or the Skills library is going to be 8 to 12 Skills inside 90 days.
POLR Coaching Retainers exist for the second case. We build the Skills with you, train your team to use them, and maintain the library as the firm changes.
What to read next
- Getting Started with Claude Connectors and Plugins: extend your Skills with reach into Gmail, Calendar, Drive, HubSpot, QuickBooks.
- Getting Started with Claude Cowork: if you have not run Cowork against a real folder yet, start there.