The guided flow
When you ask the companion to create or improve a skill, it loads the Skill Creator guide, a platform skill that scripts the whole authoring process, before asking you anything. The flow is the same whether you start from a conversation or from scratch.1
It infers what it can
From the conversation and your request, the companion works out the trigger (when should a
future run reach for this?), the inputs that vary per run, the tools and integrations
involved, what “done” looks like, and any guardrails. Starting from a conversation, most of
this is already in the transcript, and it does not re-interview you about it.
2
It asks only the open questions
At most three or four questions, in one round, each with a suggested default. It always
confirms the trigger and the output, because those decide whether the skill ever gets used
correctly.
3
It drafts with the standard structure
Goal, inputs, steps, output, guardrails. See structure
below.
4
It checks its own draft
Would a fresh run with no memory of this conversation succeed? Is every step a concrete tool
call or decision? Does it reference only tools this team can use? Did today’s dates or a
specific contact leak in where a placeholder belongs?
5
It shows you a preview and waits
You see the name, description, and full body, with Yes, create it and Cancel
suggestions. Nothing is saved until you confirm.
6
It saves and offers a test
The skill appears in the Skills tab, marked as generated. The companion offers to run it on a
real example right away, and suggests a routine if the workflow is recurring.
The structure of a good skill
A skill is instructions for a future run that has none of today’s context. Write it so that run succeeds first time. Name says what the skill does, not how. “Monday pipeline review”, not “HubSpot query plus summary v2”. No dates, no instance data. Description is the one line the companion matches on. Say when to use it and what it produces: “Weekly pipeline summary from HubSpot deals, posted as a table. Use for ‘pipeline review’ or Monday check-ins.” Content follows this shape:Example
A skill a manager might save after doing this once by hand:Checklist before you save
- The trigger is in the description, in the words a user would actually say.
- Every step names a real capability the team has connected. A skill that says “check Salesloft” on a team without Salesloft will fail on step one.
- Inputs that vary per run are inputs, not hardcoded values.
- The output section says exactly what to deliver and where.
- Anything the skill must never do is in guardrails, so an unattended routine run honors it.