GPT-5.4: Result Contracts and Verification Loop Prompt

One-sentence takeaway

GPT-5.4 reliability comes primarily from clear result contracts, tool persistence, and a verification loop. reasoning effort should be the final adjustment knob, not a high setting used to conceal an ambiguous request.

Use cases

• Suitable tasks: Multi-step coding, research synthesis, tool calls, cross-file changes, long-running Agents, and professional work with explicit acceptance criteria.

• Unsuitable tasks: Using xhigh by default for short, structured conversions that require no reasoning; in high-cost or low-latency scenarios, test none/low first.

• Applicable model versions: gpt-5.4; gpt-5.4-pro can be used for harder problems, but this template targets standard GPT-5.4.

• Applicable clients, Agents, or APIs: OpenAI Responses API, Codex, and Agents with custom tools.

• Recommended reasoning levels and parameters: Start with none or low for actions/extraction; start with medium for research or multi-document synthesis; try medium/high for long-running Agents; enable xhigh only when evals demonstrate a benefit. Set output verbosity separately to low/medium/high.

Ready-to-use content

The following template combines OpenAI's official GPT-5.4 prompting guidance (it is not an official verbatim system prompt):

<task_contract>
Goal: <one-sentence description of the final deliverable>
Success criteria:
- <observable functional/factual/file result>
- <tests or checks that must pass>
- <output format, scope, and stopping condition>
Allowed side effects: <files, tools, and external actions that may be modified or used>
Prohibited actions: <things that must not be changed, sent externally, or guessed>
</task_contract>

<tool_persistence_rules>
- Keep progressing until the success criteria are met; do not stop at analysis, a plan, or a partial result.
- When a tool fails, diagnose the failure and choose a safe alternative path; do not pretend the tool succeeded.
- Write key tool results back into the current context and cite the actual results in the next step.
</tool_persistence_rules>

<verification_loop>
At every major stage: implement → run tests/checks → read the results → fix → verify again.
Before the final response, check each success criterion one by one; clearly mark unverified assumptions rather than substituting placeholder results.
</verification_loop>

<user_updates_spec>
Update the user only when starting a major stage or when the plan changes; each update should state the result in one sentence and the next step in one sentence, without narrating routine tool calls one by one.
</user_updates_spec>

If you want a short reason displayed before a tool call, add the following developer instruction:

Before calling a tool, briefly explain why it is being called and how it advances the success criteria; do not expose internal reasoning.

Test/workflow steps

1. Run action-oriented tasks with none/low first, recording task success, tool errors, and latency.

2. If implicit requirements are missed, recovery fails after a tool is canceled, or the task stops early, first complete the result contract and verification loop.

3. Then raise the same task to medium/high and compare quality, token usage, number of tool calls, and latency.

4. For long-running tasks, compact at major milestones while keeping the contract, tool rules, and acceptance criteria unchanged.

5. Use xhigh only when business evals show a sufficient quality gain; record the full cost and context length.

Original evidence and data

• OpenAI's guidance defines reasoning.effort as controlling the number of reasoning tokens, recommends treating reasoning as the final tuning knob, and says to improve the result contract, verification loop, and tool-persistence rules first.

• Official recommendation: none is suitable for execution, extraction, and support triage; medium/higher is suitable for research, multi-document synthesis, and conflict resolution; xhigh is suitable for long-running, reasoning-intensive Agents, but should not be used by default.

• OpenAI's official GPT-5.4 autonomy/persistence template requires continuing as far as possible within the current turn through implementation, verification, and result explanation, rather than outputting only a plan.

• GPT-5.4's tool preamble can be enabled with a one-line developer instruction; OpenAI says this helps with tool accuracy and debugging, but it is not chain-of-thought output.

Scope and limitations

• The XML tags, task contract, and verification workflow in the template are a reorganization based on official guidance, not the only format guaranteed by OpenAI.

• Higher reasoning effort does not automatically mean better results. It can bring more tool calls, latency, and overthinking, so it must be evaluated by task type.

• A verification loop requires real tests or tool access; without external evidence, a model cannot establish facts solely through prompting.

• High-impact actions should still be protected by tool permissions, human confirmation, and server-side validation; do not rely on the prompt alone.

Source excerpt or observation (for compliant short quotation only)

The official guidance recommends “Treat reasoning effort as a last-mile knob” and calls for clear result contracts and verification loops to improve quality; the template retains both core principles.