The agent-fit Skill

---
name: agent-fit
description: Help a Claude Code learner decide whether a job belongs in the main conversation, a bounded agent delegation or a reusable specialist agent, then optionally create the agreed project agent.
---

# Choose a useful agent job

Understand the learner's work before recommending an agent. Stay in the main conversation for this interview so each answer informs the next question. An agent is a useful assignment, not a collection of impressive role names.

## Interview

Ask one question at a time and wait for its answer. Begin with the result the learner wants, unless their request already explains it. Use each answer to choose the next useful question. Find out what starting information the work needs, what decisions it involves, whether it repeats, what the agent should return and what the learner wants to keep under their own control. Ask for a concrete recent example when an idea is vague. If the learner struggles to name one, offer a few concrete choices grounded in their business work, such as reviewing customer questions, checking a plan or improving a page, within one question. Do not turn these topics into a fixed questionnaire or ask again for information already supplied.

Ask up to five questions in the first round. Stop earlier only when you understand the job well enough to make a useful recommendation or the learner asks to stop. An unknown or vague job is a reason to continue helping the learner describe it, not a reason to end with "there is not enough defined work" or a premature main-conversation verdict.

After five answers, if the job remains unclear, continue with up to five more adaptive questions. Before asking question six, state the specific fact you still need and explain how different answers could change your recommendation. Put that explanation before the next question in the same reply. A general statement that you need more information is insufficient. Stop as soon as enough information is known; ten questions are a ceiling, not a target. After question ten, state unresolved assumptions and give the best useful recommendation. Keep one interview and count questions across rounds; do not restart it unless the learner explicitly introduces a new job. A compound question containing several unrelated requests is several questions.

## Recommend and reflect

Reflect the learner's intended result in plain words, then recommend the simplest useful approach:

- Keep the work in the main conversation when discussion, shared context or a small one-off task is enough.
- Delegate one bounded part when that part has distinct inputs and a useful result the main conversation can use. Keep dependent steps in order. Check whether an available agent can take the assignment before proposing a new definition.
- Create a reusable specialist agent when the same responsibility, perspective and instructions will help on repeated jobs. Check whether an available agent already covers it.
- Strengthen the idea with a clearly useful responsibility when that addition improves the learner's desired result. Explain the benefit and recommend how to do the improved job. Do not add duties merely to justify creating an agent.

Explain the reason using the learner's actual work. Provide a compact brief with purpose, perspective, starting information, allowed actions, expected result and completion condition. Name unresolved assumptions. Ask for confirmation only when an important choice remains or creation has not been requested. A request to decide what fits does not itself authorize creating files or running the proposed work.

For a main-conversation recommendation or a delegation that an existing agent can perform, finish with the useful next prompt. Do not offer a plugin installation or create an agent definition for these outcomes. Do not execute the original business task merely because the interview described it.

## Optional creation

If a new reusable agent fits and the learner wants it created, read [the official creation route](references/official-creation.md). That reference owns the tooling check, installation consent and native-agent target. The interview itself requires no plugin.

Default to this project's `.claude/agents/` folder. Use the agreed brief, preserve existing files, and keep the definition's permissions within the learner's authorization. Creating an agent does not grant permission to send messages, publish, deploy, spend money or perform other external actions.

After writing, read the definition back and compare it with the brief. Invoke the named agent on a small, relevant task using learner-approved inputs. Inspect which agent actually ran and the result it returned. Report the path, result and any unfinished item. If loading or invocation fails, describe the observed failure and give the next useful repair step. A file on disk is evidence of creation, not evidence that the agent ran successfully.

# Optional official agent creation

Read this only after recommending a new reusable agent and establishing that the learner wants it created. This is course-authored routing guidance, checked against Anthropic's documentation on 7 October 2026. Recheck the current official instructions when a version or command differs from the learner's installation.

## Check what is available

Show the learner Anthropic's official [Agent Development Skill](https://github.com/anthropics/claude-plugins-official/tree/main/plugins/plugin-dev/skills/agent-development) and [agent-creator agent](https://github.com/anthropics/claude-plugins-official/blob/main/plugins/plugin-dev/agents/agent-creator.md). They are separate components in `plugin-dev@claude-plugins-official`.

Check whether both components can actually load or be invoked in the current session. A download, cached file or marketplace listing alone does not establish availability. If both are available, use them without another installation request. If the package is installed but not loaded, use the current documented loading step, then check again. Do not reinstall the package merely because a component has not loaded.

## If installation is needed

Explain that the official plugin-dev package includes plugin-development guidance beyond agent creation. Propose project scope and explain that it enables the package in this repository through `.claude/settings.json`; collaborators would still need to download it themselves. Offer local scope if the learner wants the enablement kept to their own use of this repository. Keep the agent definition project-local in either case.

Ask whether the learner wants the missing official package installed at the stated scope and the agreed agent created. Wait for consent before installing, adding a marketplace or changing plugin enablement. If the learner declines, offer to have Claude write the native project agent directly with the agreed brief, using [the built-in creation route](https://code.claude.com/docs/en/sub-agents). Proceed on that route only if creation is authorized. Do not claim the official tooling was used on the direct route.

After approval, follow [current official installation instructions](https://code.claude.com/docs/en/plugins/install). At the check date, the project-scope shell command was:

```bash
claude plugin install plugin-dev@claude-plugins-official --scope project
```

If the official marketplace is missing, the documented prerequisite is `claude plugin marketplace add anthropics/claude-plugins-official`. Only add it within the approved installation step. Load the installation in the active session using the documented route, currently `/reload-plugins`, or a new session. Inspect the actual installation result and confirm that both components load. If either fails, report the failure and offer the direct creation route without repeated installation attempts.

## Create a native project agent

Load the official Agent Development Skill for relevant instruction-design guidance and invoke the official agent-creator agent with the agreed brief. Explicitly tell the creator: "Create a native Claude Code project agent in `.claude/agents/<agreed-name>.md`. This is a project agent, so use the current native-agent requirements. Preserve the brief's scope, allowed actions and completion condition. Return the file path and anything you could not complete."

The official package teaches plugin authoring. Its plugin-root `agents/` path, color field, triggering examples and plugin-validator advice must not become assumed requirements for a native project agent. Check [native-agent requirements](https://code.claude.com/docs/en/sub-agents) for the actual target. At the check date, `name` and `description` were required; tools and model were configuration choices. Use a short description of when the agent should run, the least tool access that supports the brief and a suitable supported model. Keep detailed instructions in the body. Do not create a plugin manifest, change a global agent folder or preload unrelated Skills.

If the official components are unavailable and the learner accepts direct creation, ask Claude to write that same native definition from the brief. Describe exactly which route was used. On either route, return to the main Skill's readback and small named-agent test before reporting success.