Five Practical Agent Solutions

These are the five problems from the practical-solutions lecture, with the file in this pack that answers each one. Use the part that answers your actual difficulty and keep the rest for a later job.

1. Choosing which work to give an agent

Describe the result you want, then use the agent-fit interview to find a useful split. The interview asks up to 5 questions, then up to 5 more only when an important uncertainty remains. You can also work through the same questions directly in the conversation.

• Installable Skill: skills/agent-fit/ (install steps in AGENT-FIT-INSTALL.md)
• Conversation version: AGENT-FIT-PROMPT.md

2. An agent misunderstands the job or returns something hard to use

Begin with the result you expected and why you need it. Then compare the return with the original assignment, the inputs and the available actions. In the course we described the links that failed to open, and we added Write when the critic needed to save its own report.

• Repair prompt: AGENT-REPAIR-PROMPT.md
• Feedback prompt: AGENT-FEEDBACK-AND-PROGRESS.md (first section)

3. Claude chooses a different agent

Name the intended agent in the request and inspect which agent runs. If needed, check its loaded definition and its description, which Claude uses to select it. A newly saved agent may need a fresh conversation before Claude can find it.

Use the [exact agent name] agent to [specific job]. Give it [named inputs] and have it return [useful result]. Tell me which agent actually ran.

• More detail: AGENT-FEEDBACK-AND-PROGRESS.md (Name the intended agent)

4. Adding agents without a useful division of work

Give each agent a distinct question, the inputs it needs and its own report to return. Reviews that examine different material can run together. A step that needs an earlier result waits until that result is available. Start with the roles the result needs, choose models suited to those roles and check whether each return adds something useful.

Before starting agents, list each proposed agent with its question, its inputs and the report it returns. Mark which agents can start now and which steps must wait for another result. Remove any agent that would reread the same files to answer the same question.

• Worked example: skills/content-review/ with demonstration-inputs/content-review/
• Model and turn-limit settings: AGENT-CONFIGURATION.md

5. Not knowing what finished

Give each agent a completion instruction: say what you completed, where the result is and what remains open. Have the main agent check those returns, update the shared progress checklist and tell you when the overall job is finished. You can open the Markdown checklist and refresh it as stages finish.

• Completion prompt: AGENT-COMPLETION-PROMPT.md
• Progress checklist: AGENT-FEEDBACK-AND-PROGRESS.md (A simple progress checklist) and demonstration-inputs/content-review/progress.md