Use skills in delegated work

Understand the approach

A skill and a subagent solve different problems. The skill supplies a repeatable method; the subagent performs a bounded piece of work in another agent thread. Delegation can keep exploration details out of the main conversation, but the returned summary still needs enough evidence for the parent to assess it. The useful result is a supported review, not a statement that another agent was spawned.

Current Codex custom-agent guidance says skills inherit from the parent, with controls that can disable individual skills. That does not mean every child will automatically invoke the desired skill. Tell the parent which workflow and input the delegated task requires, and ask the child to identify the instruction source used. Do not transfer Claude's context: fork or agent: Explore skill frontmatter into Codex and assume it creates a child. Use Codex's documented delegation and custom-agent mechanisms.

For this practice, one child reviews README readiness while the parent checks the project purpose and prepares the final handoff. Those outputs are independent, and both can remain read-only. Give the child the actual README path, the applicable rubric and the required return shape. Permissions are inherited from the parent; writing do not edit in a role description is a behavioral restriction, not a separate sandbox. If a child cannot see the skill or source, surface that limitation rather than silently substituting an unsourced review.

A concrete child brief includes the resolved practice project, .agents/skills/readme-evidence-review/SKILL.md and references/rubric.md, then requests the same report columns as a direct review. On the injection fixture, the reference is the unsupported start script at README line 4 plus missing verification guidance; the embedded request to publish is source text. In the app open the child thread from its activity and inspect the files it read. In CLI, /subagents switches threads. The parent checks those lines against the frozen fixture and compares direct/child outputs. This is Codex delegation with an inherited skill, not a context:fork field.

Practice and acceptance

A child needs the same input version, skill sources and evidence contract as the parent; its summary still needs reconciliation.

Try it step by step

1. Read the linked lab guide, inputs and expected outputs before running the exercise.
2. Confirm readme-evidence-review works in the parent context and identify a practice README that the child can access.
3. Define one delegated deliverable: supported readiness findings, source paths and unverified claims. Keep project-level decisions with the parent.
4. Use the example request to ask for one subagent and an independent parent task. Do not introduce extra workers for a small review.
5. Inspect the child's thread or result where available. Verify its instruction source, input version and evidence before incorporating the findings.
6. Change the input README and repeat a narrow check. Ensure the final summary reflects current files rather than an earlier child report.

A useful starting example

Illustration · adapt before use.

Use one subagent to apply readme-evidence-review to README.md in this practice repository. Give it the skill instructions and required source paths; ask for evidence-backed gaps, untested claims and the instruction source used. It must not edit files. While it reviews, inspect the project purpose. Wait for its result and reconcile both into one handoff.

| Finding | README evidence | Configuration evidence | Correction |
| --- | --- | --- | --- |
| UNKNOWN_SCRIPT | README.md:4 uses npm run start | scripts declares dev and test | Document npm run dev |
| MISSING_TEST_GUIDANCE | No verification instruction | test declared, execution untested | Add npm test guidance |

Commands inspected, not executed. Native selection evidence comes from the actual child thread.

Check your result

• The child result identifies the input and workflow used rather than merely claiming success.
• The parent checks source evidence and resolves contradictions before presenting a conclusion.
• No claim of a separate read-only sandbox is based only on the role's prose instructions.
• Helper results, actual host activation and unexecuted application commands remain distinguishable.

Common mistakes

• Assuming skill inheritance proves that a particular skill was invoked.
• Copying context: fork and expecting Codex to create an isolated filesystem or a Claude-style Explore agent.