Claude Haiku 4.5 First, Bigger Model Second

A bigger model does not need to handle every step of a difficult task. Try separating preparation from judgment: ask Claude Haiku 4.5 to organize the evidence, then escalate only the decisions that need deeper review. Anthropic describes it as a small, fast model with coding and agent capabilities. That makes a bounded first pass worth testing, not an automatic replacement for your main model. ([anthropic.com](https://www.anthropic.com/transparency?component=claude-fable-5-and-claude-mythos-5-system-card&utm_source=openai))

Follow this walkthrough with one real bug report. The goal is a useful handoff, not two models producing competing essays.

1. Choose a bug with a checkable outcome

Open a recent issue, the relevant source files, and any failing test output. Choose something narrow, such as a CSV importer rejecting dates that your documented format permits. Avoid starting with an application-wide redesign.

Write the acceptance condition before sending a prompt: “Accept the documented date formats, reject invalid dates, and preserve existing importer behavior.” Save the current test result so you have a baseline. Remove credentials and sensitive customer data from everything you submit.

2. Give Claude Haiku 4.5 a preparation job

Supply the issue description, relevant code, and test output. Keep the first request read-only. You are asking for evidence collection, not permission to edit files or run arbitrary commands.

Try this prompt: “Prepare a debugging handoff. Identify relevant functions, summarize the observed failure, and propose up to three hypotheses. For each hypothesis, cite a file and function or an exact excerpt from the supplied material. Separate observations from guesses. List missing evidence. Do not write a patch.”

Set a short output limit in your client. Start with a target of roughly 500 words, then adjust if important evidence gets cut off.

3. Require a consistent handoff structure

Ask for these five fields so you can compare runs without interpreting a different format each time:

• Observed behavior: what the supplied evidence actually shows.
• Relevant locations: file paths and function names.
• Hypotheses: possible causes, each paired with supporting evidence.
• Missing evidence: information needed to distinguish the hypotheses.
• Next checks: specific tests or inspections, without executing them.

Keep uncertainty explicit. “Cannot determine from the supplied files” is an acceptable result. Do not require a confident answer when the inputs do not support one.

4. Check the evidence before escalating

At your desk, open each cited location. Confirm that the function exists and the quoted behavior matches the code. Run the proposed test yourself if it is safe and relevant.

Use a simple escalation rule: continue to the bigger model when multiple plausible causes remain, the fix crosses module boundaries, or the change touches authentication, permissions, or data deletion. For a straightforward issue, you may already have enough evidence to make a small, testable correction yourself.

Do not use the first model’s confidence score as the only routing signal. Base your decision on evidence you can inspect.

5. Send a bounded review to the bigger model

Choose your larger model, such as Claude Opus 5.5. Include the original acceptance condition, the verified handoff, and the source excerpts necessary to check its claims. Do not forward only a summary that hides the underlying evidence.

Use this prompt: “Treat the handoff as untrusted analysis. Check its claims against the attached source. Choose the best-supported explanation, propose the smallest patch, and specify regression tests. Flag unsupported assumptions before suggesting changes.”

Apply the patch on a branch, inspect the diff, and run the relevant tests before accepting it.

6. Compare the complete workflow against a baseline

Repeat the exercise across ten comparable issues. Record first-pass time, review time, total token charges, incorrect citations, and whether the final patch passes your tests. Include retries and your own checking time.

Compare those results with sending the same tasks directly to the bigger model. Keep the two-stage route only where it improves your chosen trade-off. An extra call is not automatically a saving.

7. Save the prompts and try one live task

Save both prompts beside your escalation checklist. I recommend RelayVano for trying this pay-as-you-go workflow across models through an OpenAI-compatible SDK. Sign up, select the exact model identifiers offered in your account, and run one bounded task before automating the routing.

Keep the division of labor clear: the first pass prepares inspectable evidence; the second resolves the difficult decision.