GPT-5.4 Prompting Guide

GPT-5.4 is a major step up from GPT-5.2, with significantly improved agentic reliability, instruction adherence, and long-context reasoning. Most GPT-5.2 agents will work without modification, but GPT-5.4 rewards more structured prompting.

For the full reference, review the official guide:

OpenAI – GPT-5.4 Prompt Guidance

Below is a Synthflow-specific migration guide, focused on what matters when building or migrating voice agents and workflows on Synthflow.

Where GPT-5.4 Excels

GPT-5.4 performs strongest in areas directly relevant to Synthflow agents:

• Personality and tone consistency across long conversations
• Agentic workflow robustness with multi-step tool calling
• Instruction adherence, especially with structured prompts
• Long-context analysis for calls with extensive history or knowledge base context

If your agent relies on any of these, GPT-5.4 is a strong upgrade.

Output Contracts

GPT-5.4 responds well to explicit output constraints. For voice agents, this is critical for keeping responses short and natural.

1
<output>
2
Return exactly what was requested. Do not add preambles or filler.
3
Apply length limits strictly. You are on a phone call—keep responses under two sentences unless the caller asks for detail.
4
</output>

This replaces the simpler “Be concise” instructions from earlier models. GPT-5.4 follows granular formatting rules more reliably than GPT-5.2.

Follow-Through Policy

GPT-5.4 introduces a clearer model for deciding when to act vs. when to ask. This is useful for agents that perform actions (transfers, bookings, lookups):

1
<follow_through>
2
If the caller's intent is clear and the next step is reversible or low-risk, proceed without asking.
3
Only ask permission if the action is irreversible, has external side effects, or requires missing sensitive information.
4
If proceeding, briefly state what you did and what remains optional.
5
</follow_through>

This reduces unnecessary confirmation loops that slow down calls.

Tool Persistence

GPT-5.4 is better at chaining tool calls, but can still stop early if not explicitly told to persist. This matters for agents using custom actions, API lookups, or multi-step workflows.

1
<tool_persistence>
2
Use tools when they improve correctness or completeness.
3
Keep calling tools until the task is complete and verification passes.
4
If a tool returns empty or partial results, retry with a different strategy.
5
Do not skip prerequisite steps just because the final action seems obvious.
6
</tool_persistence>

If your agent combines knowledge base lookups with booking or transfer actions, this block prevents it from skipping the lookup step.

Completeness Contract

For agents handling batched or multi-part requests (e.g., “Check my appointment and also update my phone number”):

1
<completeness>
2
Treat the task as incomplete until all requested items are covered.
3
Keep an internal checklist of what the caller asked for.
4
If any item is blocked, state what is missing instead of silently skipping it.
5
</completeness>

GPT-5.2 would sometimes silently drop the second part of a multi-part request. GPT-5.4 handles this better with explicit completeness rules.

Verification Loop

Before finalizing a response or action, GPT-5.4 benefits from a lightweight self-check:

1
<verification>
2
Before giving your final answer:
3
- Does the response address every part of the caller's question?
4
- Are factual claims backed by the knowledge base or tool results?
5
- Does the response match the required format and length?
6
- If the next step has external side effects, did you confirm with the caller?
7
</verification>

This is especially useful for agents that query APIs or knowledge bases before responding.

Persistence & Agent Completion

The persistence pattern from GPT-5.1 and 5.2 still applies, but GPT-5.4 is more reliable at following it:

1
<persistence>
2
You are an agent. Continue working until the caller's query is completely resolved before ending your turn.
3
Only yield when you are confident the problem is fully solved.
4
Never stop due to uncertainty—choose the most reasonable assumption, proceed, and document it afterward.
5
</persistence>

GPT-5.4 is less likely than GPT-5.2 to stop early on “mostly done” tasks when this block is present.

Prompt Structure

GPT-5.4 is the most structure-sensitive model yet. We strongly recommend sectioned prompts using XML-style blocks:

• Role / Persona
• Primary Task
• Constraints & Tools
• Follow-Through Policy
• Output Format & Length
• Completeness & Verification
• Edge-Case Handling

Example skeleton:

1
<role>You are a friendly scheduling assistant for a dental clinic.</role>
2
<task>Help callers book, reschedule, or cancel appointments.</task>
3
<constraints>Only book appointments during business hours (9am-5pm). Never disclose other patients' information.</constraints>
4
<follow_through>If the caller's intent is clear and reversible, proceed without asking.</follow_through>
5
<output>Keep responses under two sentences. You are on a phone call.</output>
6
<completeness>Address every part of the caller's request before ending.</completeness>

Unstructured prompts that worked in GPT-5.1 or 5.2 may produce inconsistent results in GPT-5.4.

Personality Controls

GPT-5.4 supports a cleaner separation between persistent personality and per-response controls. This is particularly useful for Synthflow agents that need consistent tone across long calls:

1
<persona>Warm, professional, and direct. Never robotic or overly casual.</persona>
2
<output>
3
Emotional register: empathetic but efficient—not sycophantic.
4
Formatting: no bullet points or markdown. Speak naturally.
5
Length: under two sentences per turn unless the caller asks for detail.
6
</output>

Separating persona from output rules makes it easier to tune tone without accidentally affecting response length or format.

Reasoning Effort

GPT-5.4 supports tunable reasoning effort. For Synthflow voice agents, the right setting depends on the use case:

Before increasing reasoning effort, first implement output contracts, verification loops, and tool persistence rules. These often solve quality issues without adding latency.

Migration from GPT-5.2

When migrating an existing agent:

1. Switch the model first without changing the prompt
2. Pin reasoning effort to your current level
3. Test with real conversations or your evaluation set
4. Iterate based on results, one change at a time

Common adjustments after migration:

• Add an output contract block if responses are too verbose or too terse
• Add a follow-through policy if the agent asks for unnecessary confirmations
• Add a completeness contract if multi-part requests get partially handled
• Restructure into XML sections if behavior is inconsistent

Prompt Optimizer

The OpenAI prompt optimizer remains helpful for GPT-5.4:

OpenAI Prompt Optimizer

Suggested usage:

When asked to optimize prompts, explain which specific phrases should be added or removed to elicit the desired behavior or prevent undesired behavior.

Prompt: [PROMPT] Desired behavior: [DESIRED BEHAVIOR] Undesired behavior: [UNDESIRED BEHAVIOR]

Focus on minimal edits, not full rewrites.