Oracle

Oracle — Second-Model Review

Bundle your prompt + relevant files into one "one-shot" request and send to another model. Treat all outputs as advisory: verify against the codebase and tests before acting.

When to use

• Stuck on a hard bug and want a second opinion without sharing full context
• Architecture decision that benefits from a fresh model's read
• Refactor plan you want cross-validated before executing
• Any situation where model A is too familiar with the problem to see it clearly

Approaches

Option A: code2prompt (recommended)

Rust CLI — single binary, token counting, git integration, Handlebars templates.

cargo install code2prompt

# Bundle dir → clipboard (default)
code2prompt src/

# Save to file
code2prompt src/ --output-file /tmp/bundle.txt

# Include/exclude patterns (glob)
code2prompt src/ --include "*.rs" --exclude "*.test.rs"

# Show token count before sending
code2prompt src/ --tokens

# Include git diff in bundle
code2prompt . --diff

# Custom Handlebars template
code2prompt src/ --template templates/review.hbs

Project-level config: .c2pconfig at repo root.

Run code2prompt --help for full flag reference. TUI available for interactive selection.

Option B: Agent tool with model override

Within Claude Code, spawn a subagent on a different model tier:

// In a Workflow script
const review = await agent(
  `Review this design: <context>. Question: <question>`,
  { model: 'opus', label: 'second-opinion' }
)

Writing a high-signal prompt

Any second-model review starts with zero project knowledge. Include everything needed to answer cold:

1. Project briefing — stack, build/test commands, platform constraints
2. Where things live — key directories, entrypoints, config files
3. Exact question — what you tried, the error text verbatim
4. Constraints — "don't change X", "must keep public API", "perf budget"
5. Desired output — "return patch plan", "list risky assumptions", "give 3 options with tradeoffs"

Exhaustive prompt pattern (long investigations)

When the problem may need multiple rounds, write a prompt that stands alone:

• Top: 6–30 sentence project briefing + current goal
• Middle: concrete repro steps + exact errors + what you already tried
• Bottom: all context files needed (entrypoints, configs, key modules, docs)

Make the prompt re-runnable — the model has no memory of prior exchanges.

Discipline

• Tight file set: fewest files that contain the truth. Whole-repo dumps add noise, not signal.
• No secrets: never attach .env, key files, auth tokens. Redact before sending.
• Advisory only: verify outputs against the actual codebase before acting. The model cannot run your tests.