Use and write Skills

A Skill is a reusable method: instructions, and optionally scripts and references, that the Agent loads when a task needs them. In this lesson you point the Agent at a bundled Skill, then turn the analysis from Analyze a dataset into a Skill of your own and run it on the same file.

• Time: about 25 minutes, most of it in three runs of a few minutes each.
• You need: the Session from Analyze a dataset, with growth-report.md among its Artifacts, and plant-growth.csv.
• Skills: the bundled computation-reviewer and skill-creator, then the group-comparison Skill you create.
• You end with: computation-check.md, a published group-comparison Skill, and a report it produced in a new Session.

All 17 bundled Skills, with what each one is for, are listed in Built-in research capabilities.

1. Use a bundled Skill

Go back to the Session from the dataset lesson. Type / in the message box, choose computation-reviewer so it appears as a chip, and send:

Check growth-report.md against the code executions recorded in this Session. For each number in its tables and test results, say which execution produced it, and list any number you cannot trace to an execution. Do not change the report. Write the check to computation-check.md and declare it as an artifact.

What you should see: the timeline shows the Agent loading computation-reviewer, reading the report, looking up earlier executions and their logs, and usually re-running small checks. computation-check.md appears among the Artifacts as a table that pairs each number in the report with the execution behind it, plus anything it could not match. growth-report.md itself is unchanged.

A Skill being available does not prove the model used it. The timeline entry that loads the Skill is the evidence.

2. Create a Skill with skill-creator

In the same Session, start the message with /skill-creator. The composer shows it as a Skill authoring command. Send:

/skill-creator Create a Skill named group-comparison from the analysis in this Session. It should compare one numeric column between the groups of another column in an uploaded CSV: report n, mean and standard deviation per group, run a one-way ANOVA and Tukey HSD with SciPy in the managed Python environment, draw a box plot with the individual points overlaid, and write a Markdown report that is declared as an artifact together with the figure. The user names the CSV file and the two columns.

What you should see: the Agent loads skill-creator, writes a package with a SKILL.md, a script and often a report template, and tests the script on plant-growth.csv in the sandbox. It finishes by creating a pending draft. Nothing is installed yet.

3. Review and publish the draft

Open System configuration → Skills. The Skill manager shows a pending review for group-comparison; the conversation also offers a shortcut to it.

1. Open the draft and read Review changes: the SKILL.md instructions, the script, and any other file. Edit anything you disagree with; this is the point of the review.
2. Leave Publish to at Project Skills and select Publish Skill.

Publishing creates an immutable version in the Project Skills library. Discard draft removes the proposal instead.

4. Make the library available to a Project

Published library Skills reach a run only where their library is turned on. In the left sidebar, open the menu of the Project from the dataset lesson, choose Settings, turn on Project Skills under Skill libraries, and save the Project settings.

5. Run your Skill

Create a new Session in that Project, upload plant-growth.csv again as in Analyze a dataset, and send:

Use group-comparison on plant-growth.csv to compare light_hours between the treatment groups.

Approve the first code execution as before. What you should see: the timeline shows the Agent reading the group-comparison Skill and running its bundled script on the file, with treatment as the group column and light_hours as the value. A report, a box plot and usually a results file appear among the Artifacts, named after the column the Skill compared.

6. Check your result

• The timeline shows computation-reviewer being loaded in step 1, and computation-check.md is an Artifact.
• The draft is no longer pending, and the Project's Skill libraries setting shows Project Skills with a published version.
• In the new Session, the timeline shows the Skill being read and its script running, and the report compares light_hours, not dry_mass_g.

If the new Session does not find group-comparison, check that Project Skills is on in that Project's or Session's settings. To import a Skill you already have instead of writing one, see Import and manage research Skills.

Next

Continue with Connect a tool.