A skill is how Startup OS does a kind of work well: review pricing, specify a feature, design an experiment. Skills are shared across every product, so what one product teaches the platform is available to the next.
Skills, procedures, and tool actions
The platform separates three things that are easy to blur.
| Concept | What it is | Example |
|---|---|---|
| Skill | A reusable method for a bounded task that needs judgment. It never says when it runs. | pricing-review: does the pricing capture the value delivered? |
| Procedure | A multi-step workflow with a cadence. It is the only concept that says when. | weekly-business-review: what changed this week, and do the bets still hold? |
| Tool action | One atomic action with one effect and no judgment. It is where reversibility and verification are declared. | Open a pull request. |
Each skill and procedure also declares the lowest autonomy level it needs. A procedure that only reads needs L0. One that stages a change for review needs L2.
The registry
The platform ships forty-four skills and procedures in four domains. Every one is currently marked experimental. A skill is promoted to stable only after three real uses and one documented failure mode, and none has met that bar yet.
Business
quarterly-planning sets falsifiable quarterly bets. weekly-business-review asks what changed and whether the bets hold. customer-signal-synthesis turns qualitative feedback into weighted themes. pricing-review checks whether pricing captures value.
Engineering
feature-specification scopes a feature so it can be built and verified. architecture-review tests the structure against the next few quarters. decision-record captures a decision with the alternatives it rejected. incident-review turns a failure into a better control. product-development-turn delivers one technical step as one pull request, and development-review-turn reviews one.
Growth
Research and strategy: positioning-review, competitor-analysis, channel-selection, funnel-analysis, acquisition-surface-audit, and experiment-design.
Content: content-topics-turn, content-brief, content-draft-turn, newsletter-draft-turn, launch-marketing-turn, brand-asset-from-template, founder-voice-profile, and write-like-me, which writes from the retained voice evidence.
Maintenance and outreach: changelog-content-maintenance prepares bounded internal content updates; first-customers prepares outreach for founder sending.
Distribution: venue-discovery-turn, social-publishing-turn, community-engagement-turn, and honest-community-reply. Drafting turns never publish. Whether anything goes out is decided by the product’s permissions.
Platform
The operating loop itself: loop-observe, loop-understand, loop-research, loop-review, loop-decide, loop-plan, loop-act, loop-verify, and loop-learn, plus product-operating-run for one full cycle. Alongside them: product-onboarding, evidence-brief for recommendations that meet the evidence standard, product-changelog-audit for retained release-history inspection, and skill-authoring for writing new skills.
How a skill is chosen
When a step needs judgment, the agent working it chooses a skill in four moves:
- Match. Find skills whose intent and signals fit the situation.
- Filter. Drop any whose preconditions fail, whose autonomy level is higher than the product grants, that are deprecated, or that the product has turned off.
- Rank. Prefer the best fit, and the one whose evidence requirements the product can actually meet.
- Select or decline. Use the best, or decline and say why. Declining is a valid outcome.
Turning skills on and off
The manifest’s skills section lists which skills a product may use:
skills:
enabled: all
disabled: []
Today this is guidance the agent follows when it selects a skill, not a check the platform enforces. What actually runs is decided by the product’s scheduled turns in config/schedule.json, each of which names one procedure. You can switch a turn on or off from the dashboard’s configuration page. Skills themselves cannot be edited there.
Go-to-market work
Existing onboarding, Experiment Design, Weekly Business Review and Quarterly Planning connect strategy, tests, evidence and governed decisions. This capability creates no separate GTM scheduler or execution permission. See Go-to-market workflow.