Six steps from idea to ✓✓ stamp.
Every step has a gate. No step is skipped. Skills that fail any gate do not receive the ✓✓ stamp.
Operator-led selection: we look at where AI saves the most billable hours per week. Industries with clear, repeatable workflows and measurable time cost are prioritized. We track practitioner forums, job cost data, and regulatory burden density before choosing a domain.
Federal regulations, peer-reviewed studies, industry standards bodies, and authoritative government sources. No invented citations. No hallucinated authority. If a claim requires a source and no credible source exists, the claim is cut. Every power quote in a Skilski is traceable to a real document.
Skills are written in our SKILL.md format: structured frontmatter, body content covering the workflow, inline citations, and a buyer-facing power quote grounded in the cited material. The format enforces scope discipline — one skill, one job.
Four automated phases before operator review: schema validation and markdown integrity, static security grep (injection patterns, unsafe dependencies, misuse vectors), isolated sandbox execution against normal and adversarial inputs, and field-readiness testing covering idempotency, humanized errors, latency, and resource bounds.
The founder reviews every Skilski before the ✓✓ stamp is granted. This is not a rubber stamp — it is a final read for scope drift, overclaiming, missing human-review markers, and research-citation gaps. Skills that surface problems here go back to step 2 or 3.
Passed skills are pushed to Git, deployed to Vercel, indexed in the MCP registry, and surfaced in /changelog. The release is logged with a timestamp and verification record. The catalog is live. The moat compounds daily.
What gets rejected.
The ✓✓ stamp is not granted by default. Skills fail the gate for specific, documented reasons. The most common disqualifiers are listed here. Failed skills are not listed in the catalog.
Skills with invented citations or hallucinated authority — if we cannot trace a claim to a real document, it does not ship.
Skills that produce final filings or autonomous decisions in regulated domains — those are tagged human_pass_required, not auto-execute, and are designed only for informed + reviewed use.
Skills that duplicate existing Operator-Verified flagships without meaningful scope differentiation — depth beats breadth every time.
What we do not claim.
Honest scope is part of the methodology. We build things we can actually deliver. We do not market ourselves into roles we do not fill or obligations we cannot fulfill.
We are not your lawyer, doctor, or CPA. Skills in legal, medical, and financial domains are drafting aids. A qualified human reviews before the output acts on the world.
We are not a SaaS app. Skilskis are a registry your AI agent loads from — discrete, composable capabilities delivered over MCP. No portals, no dashboards, no workflow builder lock-in.
We do not train models on your invocations. Your prompts and agent outputs are not used for model training. We do not license or sell invocation data to third parties.
Every Skilski in the catalog passed this process. Trust and data handling →