builder skill

Agent Instruction AI Skill

Turn a normal task into clear instructions an AI agent can execute, verify, and hand back for review.

What This Skill Helps You Do

This skill converts a loose request into agent-ready instructions with scope, inputs, allowed actions, verification steps, and a definition of done.

Best For

• turning repeat tasks into agent procedures
• preparing Codex or assistant tasks
• documenting operational workflows

Use This When

• You want an AI agent to take action, not just draft text.
• The task touches files, tools, apps, or multi-step decisions.
• You need clear boundaries and verification.

Inputs Needed

• task goal
• available inputs
• allowed actions
• constraints
• definition of done

Copy This Skill

You are helping me turn a task into agent-ready instructions.

Rewrite my request as a clear procedure an AI agent can follow. Include scope, inputs, allowed actions, constraints, verification steps, and what the agent should report back.

Requirements:
- Keep the task bounded.
- Make assumptions visible.
- Include checkpoints before risky actions.
- Define how success should be verified.
- End with a concise agent prompt I can reuse.

Input Template

Task goal: [what should be done]
Context: [background, files, systems, audience, source material]
Available inputs: [documents, URLs, notes, data, repo files]
Allowed actions: [read, edit, run tests, browse, send draft, create asset]
Not allowed: [destructive actions, external sends, unsupported claims]
Verification: [tests, checklist, review criteria]
Definition of done: [what final state should look like]

Output Format

1. Clarified task
2. Inputs and assumptions
3. Agent procedure
4. Verification plan
5. Human review points
6. Reusable agent prompt

Example

Example input:
Task: Update a content site with new guide pages and verify the build.

Example output:
An agent procedure that says which files to inspect, which content types to create, what commands to run, and how to summarize the result.

Quality Checklist

Fix Bad Outputs

• If the procedure is too broad, narrow it to one task and one output.
• If verification is vague, add concrete checks or commands.
• If the agent may take risky action, add approval checkpoints.
• If inputs are missing, ask for a preflight checklist.

Variations

• ChatGPT: use this to prepare a general agent task.
• Claude: use longer context and ask for operational clarity.
• Gemini: include source links when the agent needs current information.
• Codex: use when the task involves repo edits, tests, or deployment checks.