Compositions

Instructions + Skills

Instructions name the mandatory workflow; a skill contains the workflow detail. The instruction stays short: “Use the release-readiness skill before proposing publication.” The skill owns steps, references, and output shape.

This avoids loading an entire procedure on every turn while keeping the routing rule durable.

Instructions + Memory

Instructions hold reviewed team rules. Memory holds revisable recall. Link from the instruction contract to memory only when the agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose should consult it, and require freshness metadata for state that changes.

Never let machine-authored memory silently override checked-in instructions.

Instructions + Hooks

Instructions explain why a control exists and what remediation looks like. Hooks run the fast deterministic check at the relevant lifecycle point. CI or release policy repeats the load-bearing check outside the agent process.

This creates three useful layers: model guidance, immediate feedback, and independent authorization.

Instructions + Product Adapters

Keep portable rules in one source and limit CLAUDE.md, AGENTS.md, and Copilot-specific files to routing and native syntax. A change to a shared rule should produce one review that updates every adapter.

Instructions + Evidence

Define completion as proof, not confidence. For example: type-check output, production build result, changed-file scope, and the commit SHA that was actually reviewed. The agent may collect evidence; a deterministic layer should decide whether it is admissible.