playbooks

home / skills / ctdio / dotfiles / oracle

oracle skill

/agents/skills/oracle

This skill lets you consult the oracle for a second opinion on problems by sending prompts to GPT-5.2 via Codex CLI.

npx playbooks add skill ctdio/dotfiles --skill oracle

Review the files below or copy the command above to add this skill to your agents.

Files (1)
SKILL.md
1.4 KB
---
name: oracle
description: Consult the Oracle (GPT-5.2 via Codex CLI) for a second opinion or alternative perspective on a problem.
color: purple
---

# Oracle

Consult GPT-5.2 via the Codex CLI for a second opinion.

## Usage

When the user asks you to "consult the oracle" or invokes `/oracle`, pass the prompt to the Codex CLI and return the response.

## How to Consult the Oracle

Execute the following command:

```bash
codex --sandbox=read-only --model=gpt-5.2 exec "<prompt>"
```

Where `<prompt>` is either:
1. The prompt provided by the user after `/oracle`
2. A question or problem from the current conversation context

## Workflow

1. Identify what question or problem to ask the oracle
2. Format the prompt clearly for the oracle
3. Run the codex command with the prompt
4. Present the oracle's response to the user
5. Optionally synthesize or compare perspectives if relevant

## Examples

**User provides explicit prompt:**
```
/oracle How would you refactor this function to be more maintainable?
```

**User asks for second opinion on current work:**
```
/oracle
```
In this case, summarize the current context/problem and ask the oracle for their perspective.

## Notes

- The oracle runs in read-only sandbox mode for safety
- Use this for getting alternative perspectives, not for tasks requiring file writes
- The oracle's response is informational - use your judgment on whether to incorporate suggestions

Overview

This skill lets you consult the Oracle (GPT-5.2 via the Codex CLI) for a second opinion or alternative perspective on a problem. It forwards a clearly formatted prompt to the Codex CLI in a read-only sandbox and returns the model's response. Use it when you want an independent take without modifying files or state.

How this skill works

The skill identifies the question or extracts a prompt from the conversation, formats it for clarity, and executes codex --sandbox=read-only --model=gpt-5.2 exec "<prompt>". It captures the Oracle's output and returns that response to the user. Optionally, the skill can synthesize or compare the Oracle's perspective with other input before replying.

When to use it

  • Get an alternative solution or second opinion on a design, algorithm, or decision.
  • Ask clarifying questions about ambiguous requirements or trade-offs.
  • Validate ideas without performing file writes or system changes.
  • Request a concise rewrite, refactor suggestion, or critique of an approach.
  • Obtain a different perspective when stuck or facing competing options.

Best practices

  • Frame a single clear question or problem for the Oracle to avoid broad or vague answers.
  • Include relevant context or code snippets in the prompt when needed, but keep prompts focused.
  • Use the read-only sandbox for informational guidance only; do not expect file edits.
  • Compare the Oracle's output with other sources and apply judgment before adoption.
  • If the user omits a prompt, summarize the current context first, then consult the Oracle.

Example use cases

  • User asks: "/oracle How would you refactor this function to be more maintainable?" and receives refactoring suggestions.
  • User requests a second opinion on architecture choices and gets pros/cons from the Oracle.
  • User is uncertain about API error handling and asks the Oracle for recommended patterns.
  • Agent summarizes conversation context and asks the Oracle for alternative debugging approaches.
  • User asks for concise pros/cons to compare two implementation strategies.

FAQ

Can the Oracle modify files or system state?

No. The Oracle runs in a read-only sandbox and provides informational responses only.

What should I include in the prompt?

Include a focused question, relevant context, and any small code snippets needed to make the issue clear.