What is a Claude skill exactly?

A skill is a packaged unit of context (prompt, instructions, examples, even scripts) that Claude loads on demand when the user asks for something the skill knows how to handle. Think of it as a function the model calls into, except the function is markdown your team can read.

How long does a custom skill take to build?

Two to four weeks for a focused skill suite (5 to 10 skills covering one workflow area). Discovery and skill design takes one week, build and evals run two to three weeks. Faster if your conventions are already documented.

Anywhere a Claude Agent SDK app loads them: Claude Code, Claude Desktop with the Agent SDK, custom internal tools, or Claude API calls in your own services. We package skills so they install identically across all of them.

Can a skill call our internal tools?

Yes. Skills can use the same MCP servers, Bash commands, and tools the parent agent uses. We typically pair a custom skill suite with a custom MCP server when your team needs to act on internal data, not just reason about it.

How do you keep skills from going stale?

Eval suite in CI catches regressions. A monthly review process flags skills that have not been invoked in 30 days. We hand off the maintenance runbook so your team owns updates after launch.

Skill work is estimated against scope rather than sold at a list price. Writing one process properly and building a library across a team's recurring work are very different projects. The section above sets out what moves the number.

Service AI Engineering

Custom Claude skills your team can actually use.

Skills encode your conventions, runbooks, and tools so Claude operates as a domain expert. Your team gets a teammate that knows your codebase, your process, and your deploy script.

Start a project See our work

Projects are scope-dependent. Free discovery call.

Why this matters

Generic Claude is smart. Generic Claude does not know your conventions.

Out of the box, Claude can write a Rails controller. It cannot know that your team prefers Pundit over CanCanCan, that your background jobs use Sidekiq with specific retry semantics, or that your deploy runs through your custom CLI. Skills are how you teach Claude that. Done right, your team gets a senior engineer that already knows your stack.

A skill loaded

Natural ask, the right skill fires.

Skills are progressive - Claude loads just the context the task needs. Below is a real skill invocation from a SaaS engineering team's setup. One ask, three references pulled, one ready answer. Scroll up and back down to replay.

You

Review this PR for input-validation issues. Trust nothing from query params.

skill_load name:"security-review", scope:"php+wp" running

skill security-review v2.1

categories 12 (xss, sqli, csrf, …)

refs_pulled 3 files (242 lines)

context_used 8.4k tokens

security-review skill

Progressive context skill loads only the references the task needs · no token waste
Eval-backed every skill ships with a Vitest eval suite · regressions block merge
Distributable .zip install · works in Claude Code, Claude Desktop, and Claude API

Skills your team will actually invoke

Skill names and descriptions designed so Claude routes to them when your team types natural language. No "what skill do I run for this?" lookup. Trigger phrases match how your team actually talks.

Skills that trigger when they should, checked against real requests.

Progressive disclosure built in

Skills load lightweight context first, then deeper detail only when needed. Your team gets the right information without burning the entire context window on one task.

Context loaded on demand, so one task cannot eat the whole window.

Versioned, reviewable, owned

Every skill ships in a git repo. Pull requests review skill changes the same way your team reviews code. No more "who changed the prompt?"

Skill changes follow your existing change management.

Eval suite per skill

Real cases, expected outputs, automated runs in CI. A skill change that breaks a real workflow blocks merge before it ships.

Skill regression rate falls to near zero.

Distribution that works

Plugin packaging for Claude Code, marketplace listings, install one-liners, runbooks. Your team installs skills in seconds, not by hand-copying files.

Written so the team picks them up without a training session.

Documented for non-experts

Each skill ships with a one-page reference your team can read without knowing how the skill works internally. Your engineers can extend the skill without consulting us.

Your team can ship new skills after week six.

Tested

triggering, against the phrasings people actually use rather than the ideal one

A skill that never fires is worse than no skill at all. We tune the description against real requests until the right one loads.

SKILL.md md

    
      
          1
          # .claude/skills/code-review-rails/SKILL.md
        
          2
          ---
        
          3
          name: code-review-rails
        
          4
          description: Use when reviewing Ruby on Rails pull requests. Checks ActiveRecord
        
          5
            N+1 queries, strong parameter coverage, missing scopes, security
        
          6
            patterns from the Rails guide, and team-specific conventions in
        
          7
            rails-conventions.md. Trigger on "review this Rails PR", "code review",
        
          8
            "PR check", "audit this controller".
        
          9
          ---
        
          10
           
        
          11
          # Code Review: Rails
        
          12
           
        
          13
          Read `rails-conventions.md` for our team rules.
        
          14
           
        
          15
          ## Phase 1: Quick Scan
        
          16
          Read every changed file. List every public method, every callback, every
        
          17
          ActiveRecord query. Note anything that looks suspicious for the next phase.
        
          18
           
        
          19
          ## Phase 2: Deep Checks
        
          20
          For each query, confirm includes/preload covers all loaded associations.
        
          21
          For each controller action, confirm strong params match the model attributes.
        
          22
          For each new background job, confirm idempotency and retry behavior.

Discovery

One week. We shadow your team, identify the workflows worth encoding as skills, design the skill catalog, and lock the trigger phrases. You approve the catalog before any skill is written.

Fixed scope, fixed price.

Build

Two to three weeks. Skills written, evals wired, packaging built. Your team installs and uses them in week two for real-world feedback.

Skills used in real work by week two.

Handoff

One week. Documentation, maintenance runbook, training session for your team to extend skills themselves. We step back when your team can ship the next skill.

Your team owns the suite at the end.

What it costs

Skill pricing follows how many processes and how repeatable they are.

One documented process and a library covering a team's recurring work are the same sentence and very different projects. Here is what moves the number.

What moves the number

How many processes Each skill is a process written down properly. The writing is the work, and it is worth doing whether or not a model reads it.
How well documented already A team with written process is a fast start. Extracting it from the people who hold it is discovery before it is authoring.
How much they do A skill that explains is simpler than one that runs tools and changes things.
Who has to adopt them Written so the team picks them up without training, or they sit unused. That is a copy problem as much as a technical one.
How they stay current Process changes. A skill nobody updates becomes wrong, and wrong is worse than absent.

Small, well-defined work is usually better handled as an hourly block than scoped as a project, and we will say so rather than inflate it into one.

Rough idea to delivery

You send the details What you want built, roughly, plus budget range and timing. The form asks for all of it
Within 4 business hours We read it and reply. Nothing is scheduled before we know what it is about
Then the call Free discovery, booked once there is enough on the table to make it worth your hour
Within 48 hours of the call A written fixed-price quote you approve before anything starts

Get an estimate

Quick enquiry

Same work explained over and over?

Tell us which process eats the most repeated explanation. That is almost always the right first skill.

Prefer the full form? Start a project

What is a Claude skill exactly?

A skill is a packaged unit of context (prompt, instructions, examples, even scripts) that Claude loads on demand when the user asks for something the skill knows how to handle. Think of it as a function the model calls into, except the function is markdown your team can read.
How long does a custom skill take to build?

Two to four weeks for a focused skill suite (5 to 10 skills covering one workflow area). Discovery and skill design takes one week, build and evals run two to three weeks. Faster if your conventions are already documented.
Where do skills run?

Anywhere a Claude Agent SDK app loads them: Claude Code, Claude Desktop with the Agent SDK, custom internal tools, or Claude API calls in your own services. We package skills so they install identically across all of them.
Can a skill call our internal tools?

Yes. Skills can use the same MCP servers, Bash commands, and tools the parent agent uses. We typically pair a custom skill suite with a custom MCP server when your team needs to act on internal data, not just reason about it.
How do you keep skills from going stale?

Eval suite in CI catches regressions. A monthly review process flags skills that have not been invoked in 30 days. We hand off the maintenance runbook so your team owns updates after launch.
What does it cost?

Skill work is estimated against scope rather than sold at a list price. Writing one process properly and building a library across a team's recurring work are very different projects. The section above sets out what moves the number.