Monetize Claude Code Skills by Acceptance
Monetize a Claude Code Skill by defining what the buyer can accept before you choose what to charge. License the package when the buyer wants to operate it, charge for hosted deliverables when acceptance can be checked reliably, or sell a reviewed service when the work still needs judgment.
Test the deliverable before setting the price
Pick one output the buyer can inspect and write down what counts as complete. For a catalog cleanup Skill, that might mean preserving every product identifier, normalizing units, and listing any row that needs a decision. Run the same inputs in Claude Code using direct Skill invocation. Then try ordinary requests from a fresh session to see whether buyers can get the workflow to run using their own wording.
Use three illustrative inputs: a clean 20-row catalog, a catalog with one ambiguous unit, and a file missing product identifiers. Expect a completed file, a flagged row, and a request for corrected input, respectively. If direct invocation produces the expected output but ordinary wording misses the workflow, tighten the Skill description and retest before making conversational discovery part of the offer.
Put one accepted and one rejected example in the buyer’s handoff. The examples should show where the contract draws the boundary between an included revision and a new delivery.
Extend Claude with skills - Claude Code DocsPrice each offer around who runs the Skill
A file license fits buyers who want to install, inspect, and operate the workflow themselves. Anthropic's Skill repository shows the package as a folder of instructions, scripts, and resources. Decide which files the buyer receives, how many installations the agreement covers, and which updates or support you include. Make acceptance a clean installation followed by a successful sample run in the buyer's environment.
A hosted offer fits buyers who want to submit inputs and collect outputs while you retain the package. Define a billable delivery and decide how you handle retries or corrections. A managed service fits work that still requires you to reconcile exceptions or approve the output. Quote that review explicitly and agree on an exception limit or a separate fee for additional work. Track review time during the pilot before offering a standing price.
Time the review work on failures as well as successes. Review labor can dominate the unit economics even when model usage is inexpensive.
GitHub - anthropics/skills: Public repository for Agent Skills · GitHubCalculate margin on accepted work
For an illustrative hosted pilot, assume 100 execution attempts produce 80 accepted reports sold at $4 each. Assume a blended execution cost of $0.25 per attempt and 45 seconds of review per accepted report at $30 per hour. Revenue is $320, execution costs $25, and one hour of review costs $30. That leaves $265 before hosting, payment fees, and support, or about $3.31 per accepted report before those costs.
The Claude API pricing table separates token charges by model and includes prompt caching categories. For your own forecast, confirm input, output, and applicable cache rates for the exact model and platform. Include any applicable tool and deployment charges and reconcile a pilot against the bill. Divide the resulting cost by accepted outputs so the calculation captures failed attempts and retries.
Pricing - Claude Platform DocsPin the version that passed acceptance
For a hosted offer using the Claude Messages API, specify the Skill identifier, type, and tested version in the container configuration along with code execution. Store the release identifier with each job and replay the acceptance samples before promoting a new version. Use one billing record per customer job so that retries can contribute to your costs without automatically creating duplicate charges.
Suppose a new release changes how the catalog Skill handles ambiguous units. If it completes every run but silently guesses a unit, hold the rollout because the report fails the buyer's acceptance rule. Keep serving the previously validated version while you investigate. If human review is still needed after revisions, include that review in the service scope or limit the self-serve offer to inputs the workflow can handle consistently.
Using Agent Skills with the API - Claude Platform Docs