Create A Skill

Create an OpenCode-compatible Agent Skill

Create, revise, or diagnose one portable Agent Skill. Repository instructions, the installed OpenCode v1 contract, and the target repository's chosen skill root govern the artifact.

Choose the mode

• Create or revise — inspect, design, edit, reconnect callers, and validate.
• Diagnose or review — inspect and validate read-only, then report evidence and recommendations. Edit only if the user also requested implementation.

When a skill or reference already owns the requested job, extend it within the user's scope instead of creating a competing owner. Adapt the work to the change; clarify only when it would materially change the requested deliverable.

1. Inspect the target

Inspect applicable AGENTS.md, OpenCode config, existing skills, neighboring commands, and repository validation. Determine which supported root the repository already uses:

• .agents/skills/<name>/SKILL.md for a portable project skill;
• .opencode/skills/<name>/SKILL.md for an OpenCode-specific project skill;
• the corresponding supported user-level root only when the user requested a personal skill.

Do not introduce a parallel skill tree without a material reason and user agreement. State the skill's one job, its boundary, current callers, and whether it needs a separate slash command.

2. Design discovery and invocation

Preserve existing invocation policy unless the task changes it. Default a new task skill to relevance-based discovery, with a direct command when useful. Reserve deliberate-only access for an explicit entry-point contract, not merely because a later operation needs authorization. Reuse authorization already granted for that action; a method or phase change does not require another ask.

Use portable skill frontmatter:

• name — required, kebab-case, and equal to the directory name;
• description — required, concise, and specific enough for model discovery;
• license, compatibility, and metadata — optional only when they convey a real portable contract.

Do not copy client-specific invocation keys into an OpenCode skill. OpenCode loads skills through the skill tool. Control model access with ordered permission.skill patterns in agent or project configuration. When a direct slash entry point matters, add a native command with the same stable ID under the selected OpenCode commands/ root; the wrapper must load the skill and pass $ARGUMENTS as user input.

For access rules, put "*" before narrower patterns because the last matching rule wins. Use ask for a source skill that was intended to require deliberate human invocation; this is an approval approximation, not provenance-aware manual-only behavior.

3. Design the information hierarchy

Apply the authoring principles. Keep instructions needed every time in SKILL.md; put conditional reference, scripts, and assets beside it and link them with an explicit loading condition. Consult only the needed heading in the glossary.

4. Implement and reconnect

Write the outcome, decision criteria and essential constraints. Use fixed sequences or strict gates where order, correctness or safety requires them; leave room for judgment elsewhere. Add only useful bundled resources. Update direct callers, command wrappers, adapter policy, and provenance affected by a name or invocation change. If a repository generates its OpenCode target, change the canonical source or target overlay and regenerate; never hand-edit the generated output.

Complete the edit when names are stable, links resolve, target roots do not compete, permissions preserve the intended boundary, and the diff contains no unrelated client assumptions.

5. Validate and forward-test

Immediately before validation, read the OpenCode validation checklist. Run its safe structural and runtime checks using a disposable target/config copy when the client may create dependency artifacts. Test explicit command invocation, positive skill discovery, a close negative prompt, and every distinct branch that can be exercised without unauthorized external writes.

Check that the skill selects the right task and performs a useful next action without unnecessary questions or scope expansion. A listing or forced skill load does not prove autonomous selection. Report unexercised behavior as unverified. In diagnose mode, a failing target is valid evidence. In create/revise mode, iterate against observed behavior and keep the diff within the intended contract.