oracle
Get a second opinion by bundling a prompt + a curated file set, then asking another powerful LLM for debugging, refactor advice, design checks, or cross-validation.
When & Why to Use This Skill
Oracle is a specialized cross-agent skill designed to provide a 'second brain' perspective by bundling specific code context and prompts for validation by alternative high-performance LLMs. It enhances development workflows by enabling multi-model cross-validation, allowing developers to leverage the unique reasoning strengths of different agents (like Claude and Codex) for debugging, refactoring, and architectural checks.
Use Cases
- Logic Cross-Validation: Verify complex algorithms or business logic by sending the context to a secondary high-reasoning model to identify potential edge cases or hallucinations.
- Deep Debugging: Bundle a curated set of source files and error logs to get a fresh diagnostic opinion when the primary agent reaches a bottleneck.
- Refactor Assessment: Submit proposed code changes to a different model to receive objective feedback on maintainability, performance, and adherence to design patterns.
- Architectural Design Checks: Use the 'Oracle' to perform sanity checks on system designs or API contracts before implementation, ensuring the proposed structure is robust.
- Quality Assurance: Perform a final 'second opinion' pass on critical code paths to ensure high-security standards and code quality before merging.
| name | oracle |
|---|---|
| description | Get a second opinion by bundling a prompt + a curated file set, then asking another powerful LLM for debugging, refactor advice, design checks, or cross-validation. |
Oracle (cross-agent)
Use this skill when you want a fast “second brain” pass from the other agent:
- In Codex: bundle context, then ask Claude CLI (Opus 4.5) for review.
- In Claude: bundle context, then ask Codex CLI (GPT‑5.2,
xhighreasoning) for review.
This skill uses a tiny local bundler script.
Workflow
- Pick the smallest file set that contains the truth (avoid secrets by default).
- Verify the selected files / bundle look right.
- Run the Oracle script for your current agent.
Commands
When running in Codex:
- Preview selection:
"$HOME/.agents/skills/oracle/scripts/oracle-bundle" --dry-run -p "<task>" --file "src/**" --file "!**/*.test.*" - Preview bundle:
"$HOME/.agents/skills/oracle/scripts/oracle-bundle" -p "<task>" --file "src/**" --file "!**/*.test.*" - Run:
"$HOME/.agents/skills/oracle/scripts/oracle-to-claude" -p "<task>" --file "src/**" --file "!**/*.test.*"
- Preview selection:
When running in Claude or any other agent:
- Preview selection:
"$HOME/.agents/skills/oracle/scripts/oracle-bundle" --dry-run -p "<task>" --file "src/**" --file "!**/*.test.*" - Preview bundle:
"$HOME/.agents/skills/oracle/scripts/oracle-bundle" -p "<task>" --file "src/**" --file "!**/*.test.*" - Run:
"$HOME/.agents/skills/oracle/scripts/oracle-to-codex" -p "<task>" --file "src/**" --file "!**/*.test.*"
- Preview selection:
Tips
- Prefer a minimal file set over “whole repo”.
- If you need diffs reviewed, paste the diff into the prompt or attach the diff file via
--file. - Make the prompt completely standalone: include error text, constraints, and the desired output format (plan vs patch vs pros/cons).
- Never include secrets (
.env, tokens, key files). - Oracle can be slow while it reasons. Allow it several minutes to process.