The Skill Creator Agent — Extending the Framework Deliberately
How Datarim's Skill Creator agent researches best practices, audits what already exists, and produces new skills, agents, and commands only when nothing existing covers the need.
Every time a capability gap appears — a new domain, a workflow the framework does not cover — the reflex is to add something. That reflex is often right, but sometimes the gap is already covered by an existing component that just needs updating, or it is covered by two components that could be merged. The Skill Creator agent does the audit before it does the creation.
It runs at /dr-addskill, and it runs on an Opus-class model. The work involves research, judgment, and architectural decisions about how new components fit into the existing framework — exactly the kind of reasoning that warrants the heavier model.
The workflow
The first step is research: searching for existing implementations and community patterns relevant to the requested domain. The second step is a framework audit — reading the current skills, agents, and commands to determine whether the need is already covered, partially covered, or genuinely new. Update is preferred over create; extending an existing component is preferred over adding a new file.
When a new component is warranted, the agent designs it following the Datarim conventions: YAML frontmatter with name, description capped at 155 characters, and a required model field. The description is written as a triggering condition — not a workflow summary — so the agent that reads it knows exactly when to load the skill, not just what the skill contains.
The model assignment rule
Every skill and agent the Skill Creator produces includes an explicit model field. The rule is straightforward: Opus for critical reasoning, architecture, and security; Sonnet for standard code work, structured tasks, and content; Haiku for simple lookups and structured output. Reference-only skills — rules and patterns the caller applies inline — omit the field to inherit from the calling context. No exceptions.
A concrete example
Say there is a need for a skill that handles database migration reviews. The Skill Creator first checks: is there a database skill already? A migration guide template? A review checklist? If none of those exist, it searches for community patterns on migration review workflows, extracts the structure, and designs a skill that covers the specific gap — not a generic "database operations" skill that would overlap with things already in the framework.
The approval gate
Nothing is written to disk without operator approval. The agent presents the proposed file paths and content preview, explains why this structure and this scope, then waits. After approval it creates the files and confirms what was installed and how to invoke it.
Full details on the Skill Creator reference card, or read what Datarim is for context.