Self-Skill Forge

The problem

Agents can get stuck on complex tasks when they lack a reusable capability, such as a specific data-format routine, tool integration, or multi-step workflow. Asking a human to find, audit, install, and adjust skills every time breaks the loop. This skill turns “fill the missing capability” into an executable agent process.

How it works

It organizes the workflow around search, audit, install, generate, and iterate:
- Gap detection: decides whether a missing ability should be packaged as a skill rather than treated as a one-off operation.
- Dual search: checks the recommended marketplace first and treats external community results as candidates that still require security review.
- Risk grading: handles P0, P1, and P2 outcomes, covering destructive commands, code injection, outbound requests, and hardcoded credentials.
- Install and landing: installs marketplace skills through the host-provided capability and avoids fake skillId values; when no suitable package exists, it generates a skill directory with a SKILL.md file.
- Iteration: when errors, weak feedback, better tools, or marketplace updates appear, it reads the current implementation, edits the skill, bumps the version, appends CHANGELOG.md, and reruns the audit.

Boundaries

It is aimed at task-driven agents that need evolving capabilities, not one-shot questions. Marketplace installs do not auto-enable, locally modified marketplace skills need a marker to prevent later updates from overwriting changes, and every generated or iterated skill must pass a security self-check. Sensitive-information decisions remain with the user.