Choosing the shape

Decide the smallest arrangement that solves the problem. Guide the user, do not interrogate.

Simple vs rich

• Simple problem → a single skill. One workflow or one body of knowledge. It is already a minimal plugin under the hood.
• Richer problem → a plugin. Reach for a plugin when the problem has several distinct parts (each its own skill), related skills belong together, or you need more than instructions (a rule, hook, or MCP server).

Do not escalate to a plugin for its own sake. If a single skill solves it, that is the answer.

Which primitive for which intent

Generic hook vs native hook

Choose a cross-agent hooks entry when the behavior is agent-independent and one of the generic lifecycle events can express it: PreToolUse, PostToolUse, UserPromptSubmit, SessionStart, or Stop. This is the paved road for linters, validators, and notifications that should work across supported agents.

Choose nativeHooks only when the requirement depends on an agent-specific event, payload field, or behavior that the generic contract cannot express. Name the target agent and the missing generic capability in the plan. Do not use the native tier merely because the user happens to run one agent today.

Skill vs rule: do not default to skills

The most common mistake is making everything a skill. A skill fires when the agent judges a task relevant; a rule is always in context. If the user says the agent "keeps getting X wrong", "every time", or "never follows our convention", that is an always-on convention, which is a rule (sometimes a rule plus a skill), not a skill on its own. Weigh this explicitly for any conventions-style request before defaulting to a skill.

Heuristics

• A skill is instructions the agent reaches for; a rule is a convention it always obeys; a hook is deterministic lifecycle behavior; an MCP server is a live capability.
• One responsibility per plugin; two unrelated jobs are two plugins.
• Do not add primitives to look complete.

Some behaviours are not skills or rules at all

Not every "the agent should X" belongs in skill or rule prose. If a behavior is better enforced deterministically at an agent lifecycle event, use a hook. If it is a broad observable invariant better expressed as a verifier (tessl change verify, an LLM judge at glob level), point the user at the change-verify skill. Context is the right tool for workflows and conventions the agent reasons about; hooks and verifiers are the right tools for deterministic guardrails.

Eval is downstream

Evaluation is not part of choosing the shape, and not part of this plugin. Once the composition is built, skill-optimizer handles eval as a separate step. The only thing composition owes eval is a well-formed skill or plugin for it to run on.