OpenAI publica la guía de prompts de GPT 5.5: 7 cambios clave

Background: why legacy prompts need a rewrite

Django co-creator Simon Willison noted that OpenAI recommends rewriting from scratch rather than assuming tuned legacy prompts still work. If you are moving to GPT-5.5 (or trying similar capabilities on LimaxAI), treat prompts as a new product design—not a model-name swap.

Change 1: from hand-holding to outcome-first

Estilo antiguo (proceso primero):

Comprueba A, luego B, compara campos uno a uno, razona cada excepción, decide qué herramienta llamar y explica todo el proceso al usuario.

Estilo nuevo (resultado primero):

Resuelve por completo el problema del usuario. Usa política y datos de cuenta para decidir qué hacer primero; informa qué se hizo y qué queda. Si falta información, haz solo la pregunta más crítica.

Describe el destino, no cada giro del mapa. GPT-5.5 planifica su propio camino; un guion rígido puede bloquear mejores rutas.

Change 2: fewer absolutes—reserve them for real red lines

Guardrails like “always JSON,” “never exceed 500 words,” or “must search before answering” were common. GPT-5.5 is highly sensitive to absolutes—when two conflicting “must” rules appear, the model may try to satisfy both and fail both.

Situation	Recommended style
Safety rules, required fields, hard bans	Use always/never
Whether to search or ask the user	Use conditional rules

Example: replace “must search before answering” with “After receiving the question, check whether existing context is enough to answer; if yes, reply directly without searching.”

Change 3: roles return—but stay short

• Steady: reliable collaborator—calm, direct; advance without unnecessary check-ins.
• Expressive: conversational, curious; ask when ambiguous, decide when context is enough.

Change 4: acknowledge multi-step work before executing

Products like Codex already use this pattern (e.g. “Analyzing your codebase, starting with test files”) for APIs and long ChatGPT tasks alike.

Change 5: define stop conditions

Si solo dices «busca información relevante y responde» sin definir cuándo parar, el modelo puede entrar en bucles de recuperación.

Dos estrategias habituales:

1. Presupuesto de búsqueda: una ronda por defecto en Q&A normal; responde si la evidencia de la primera página basta. Segunda ronda solo si la primera no tiene la respuesta, faltan datos clave o el usuario pide cobertura exhaustiva.
2. Parada de bucle: tras cada ronda de herramientas, pregúntate si la evidencia actual responde la pregunta central; si sí, para.

Change 6: the official seven-part prompt skeleton

Section	Purpose
Role	1–2 sentences: function and context
Personality	Tone, style, collaboration
Goal	User-visible deliverable
Success	Definition of done
Constraints	Safety, business, evidence, side effects
Output	Format, length, voice
Stop rules	When to retry, escalate, refuse, ask, or stop

OpenAI developer console also offers Optimize to reshape prompts to this structure.

Change 7: format control and concise defaults

GPT-5.5 responde de forma concisa y directa por defecto. En la API usa text.verbosity (medium por defecto; prueba low para respuestas más cortas); para formatos complejos prioriza Structured Outputs.

Ejemplo:

Por defecto párrafos naturales; títulos, negrita y listas solo cuando ayuden a comprender.

En reescrituras, indica qué conservar antes de cómo editar:

Conserva estructura, longitud y voz originales; mejora solo la claridad—sin nuevas afirmaciones o párrafos salvo petición del usuario.

Sin reglas de conservación, el modelo suele expandir y reestructurar por defecto.

Migration tip: tune reasoning effort from the low end

Beyond prompts, re-evaluate reasoning.effort. GPT-5.5 defaults to medium; many tasks work at low. Higher is not automatically better—conflicting instructions or weak stop rules plus high effort can mean over-searching and detours. Run A/B tests on a fixed eval set; start with low / medium, then move to high / xhigh only when quality gains justify latency and cost.

Summary: one line to remember

Say what you want—not how to do every step.

Dimension	GPT-5.4 and earlier	GPT-5.5
Process	Detailed steps	Outcomes only
Tone	Many absolutes	Absolutes for red lines; conditionals elsewhere
Role	Long background	1–2 sentences + personality
Multi-step	Execute immediately	Preamble first
Stopping	Often omitted	Should be explicit
Structure	Free-form	Seven-part skeleton (trim as needed)
Format	Defaults	API params + prompt guidance

After rewriting, run side-by-side tests on real tasks. To try GPT in chat, use the button below on LimaxAI with your preferred model.