Using skills
Skill loads are useful evidence, not completed work
Interpret creator analytics without confusing a tool retrieval with a business result.
Quick answer
A skill load records successful retrieval of a workflow’s instructions. It does not prove that the AI followed every step, completed an outside action, or produced a good result. Use load counts to understand interest and repeat use; use reviewed outputs and explicit feedback to evaluate quality.
Know the event being counted
The Skillset usage ledger records successful first loads. Continuation chunks for the same load do not each become another use. This avoids inflating a long workflow’s popularity simply because its instructions require several responses.
Read each metric literally
A unique user count describes distinct accounts represented in the recorded activity. A returning-user count identifies users who loaded skills more than once. Neither number is automatically a retention rate: a retention measure needs a defined starting group, return window, and qualifying behavior.
| Metric | What it helps answer | What it does not prove |
|---|---|---|
| Loads | Was the workflow retrieved? | Was the task completed? |
| Unique users | How many accounts retrieved it? | How many people adopted it? |
| Returning users | Did accounts retrieve again? | Did the output improve? |
| Net sales | What revenue is recorded? | Did the buyer achieve a result? |
Look for questions, not flattering totals
Many first loads with few returns may suggest a mismatch between the listing and the method, or a task that is naturally infrequent. Repeated loads might indicate a useful recurring job, but they can also come from troubleshooting. Review feedback and actual examples before choosing an explanation.
Build an honest improvement loop
Keep a small log of recurring questions, missing inputs, confusing steps, and reviewed output examples. Change one substantial issue in a new version, then observe whether the same problem recurs. Do not invent success percentages to fill an empty dashboard.
When sharing results publicly, state the period and the measurement. “Thirty instruction loads this month” and “thirty campaigns completed” are different claims. Only make the second claim when there is evidence for those completed campaigns.
Common questions
Are loads the same as downloads?
No. A load is retrieval through the connected workflow. Historical sales or file downloads should not be relabeled as MCP loads.
Can a load count prove success?
No. Success needs evidence appropriate to the task, such as an approved draft or a verified external result.
Why does an older skill start with little usage data?
Hosted metering begins when the ledger is deployed. It should not invent activity for periods that were never measured.
Put a workflow to work.
Connect your library to your AI, or turn your method into a skill.