GPT-5.5 Outcome-Oriented and Verification-Driven Agent Prompt

One-sentence takeaway

OpenAI's official guidance for GPT-5.5 is to clearly define the goal, success criteria, allowed side effects, evidence rules, and output contract first, then let the model decide the intermediate steps and perform targeted verification before finishing.

Use cases

• Suitable tasks: Tool-using customer service, code modifications, retrieval-augmented question answering, long-running agents, and research tasks that require deliverables and verification results.

• Unsuitable tasks: Simple classification or short answers that require extremely low latency; evaluate lower reasoning effort or a smaller model first for these tasks.

• Applicable model version: GPT-5.5 (the page also provides general guidance for the GPT-5.5 family, so other versions should not be assumed to be exactly the same).

• Applicable clients, agents, or APIs: OpenAI Responses API; the approach can be migrated to Codex or a self-built tool-using agent, but tool side effects and output fields must be rewritten for each client.

• Recommended reasoning level and parameters: Start with reasoning.effort=medium; increase it to high/xhigh only when representative evaluations show a benefit. Use the Responses API and explicitly define tool descriptions, retries, and error handling.

Ready-to-use content

Below is a copyable version organized from OpenAI's official template. It preserves the official goal/success-criteria structure, while business fields need to be replaced. It does not ask the model to expose hidden reasoning; it asks for deliverables that can be verified.

Resolve the customer's issue end to end.

Success means:
- the eligibility decision is made from the available policy and account data
- any allowed action is completed before responding
- the final answer includes completed_actions, customer_message, and blockers
- if evidence is missing, ask for the smallest missing field

For coding or research agents, the official planning template can be applied as follows:

Role: You are a careful implementation and verification agent.

# Goal
<the concrete outcome to deliver>

# Success criteria
- <observable acceptance criterion 1>
- <observable acceptance criterion 2>

# Constraints
- State allowed side effects and files/services that may be changed.
- Do not claim completion without evidence from the available tools.

# Output
- completed_actions: <what was actually done>
- evidence: <tests, source links, or retrieved records>
- blockers: <missing inputs or failed checks>

# Stop rules
- Stop when the success criteria are verified.
- If evidence is missing, ask only for the smallest missing field.

Testing/workflow steps

1. Fill the template with the goal, success criteria, constraints, allowed side effects, and final fields.

2. In the Responses API, first run representative samples with reasoning.effort=medium; tool descriptions should specify when to use them, their inputs, side effects, errors, and whether retries are safe.

3. If the task requires multiple rounds of tool calls, have the agent provide a brief progress update before calling tools; after completion, run targeted tests, type checks, linting, builds, or smoke tests.

4. Check acceptance against completed_actions, evidence, and blockers; do not treat a plan or the model's claim as proof of completion.

5. Compare the quality, latency, and token cost of low, medium, and high; increase the level only when evaluations demonstrate a benefit.

Original evidence and data

• OpenAI summarizes its recommended approach for GPT-5.5 as follows: provide a clear goal, constraints, and output format rather than prescribing every intermediate step.

• The official recommendation is to treat reasoning.effort as a tuning knob, not the primary remedy for quality problems; GPT-5.5 defaults to medium.

• The latest official model guidance requires tool-using workflows to define “when the task counts as complete” and how to verify it; for long-running workflows, use the Responses API's state and tool capabilities.

• The official page provides an outcome-oriented example: complete an eligibility decision, perform allowed actions, output completed actions/customer message/blockers, and ask only for the smallest field when evidence is missing.

Scope and limitations

• This is guidance on prompting and evaluation methods, not an independent win rate for GPT-5.5 on a particular business dataset.

• “Verification” can cover only the tests, tools, and evidence accessible to the agent; system state it cannot access must be written to blockers and cannot be supplied by the prompt.

• Higher reasoning effort may increase latency, cost, or overthinking; xhigh should not be treated as the default configuration.

• The fields and tool names in the official example must be adapted to the actual API, permissions, and side effects.

Source excerpt or observation (compliance-limited short quotation)

The official core wording is “Give the model the task, constraints, and desired output format”, and it recommends defining for agents “what counts as done and how the model should verify its work”.