Conditional and branched prompts

So far, your prompts have followed a linear path: one instruction, one result. Real-world professional situations are rarely that docile. What do you do when the right answer depends on a characteristic you won’t know until you process the case, or when you need to cover multiple scenarios in a single pass? Conditional and branched prompts absorb this variability: instead of writing a separate prompt for each scenario, you write the rule that discriminates among them once.

Write an “if/then” rule

The simplest conditional describes the possible cases and the action associated with each. Imagine you moderate reviews on your business profile: comments arrive daily, mixed together, and they don’t all call for the same response.

Analyze this customer review and respond appropriately.

Rules:
- If the review is positive: thank them warmly and invite
  them to leave a Google review
- If the review is negative: apologize, ask for
  details about the issue, and offer direct contact
- If the review is a question: answer factually and offer
  to take it further by email
- If the review is spam: simply state "SPAM - delete"

Review: "[PASTE REVIEW HERE]"

This single prompt covers four cases and lets you process dozens of reviews without rewriting anything between them. The change is deeper than it appears: you no longer sort upfront to choose the right prompt; the sorting is now part of the instruction.

Combine multiple criteria

When the response depends on two or three variables at once, write your conditions as a decision tree. A career counselor doesn’t give the same advice to a junior professional changing careers as to a senior executive who wants to start a company; the following prompt encodes that discernment once and for all.

You are a professional career counselor.
A candidate provides the following information.
Tailor your recommendations to their profile.

Candidate profile:
- Experience: [JUNIOR (0-2 years) / MID-LEVEL (3-7 years) / SENIOR (8+ years)]
- Current sector: [SECTOR]
- Goal: [ADVANCEMENT / CAREER CHANGE / START A BUSINESS]
- Constraint: [GEOGRAPHIC / SALARY / TIME / NONE]

Adaptation rules:
- If JUNIOR + ADVANCEMENT: focus on skills to acquire
  and relevant training
- If JUNIOR + CAREER CHANGE: emphasize transferable skills
  and suggest concrete bridges
- If MID-LEVEL + ADVANCEMENT: propose positioning strategies
  and targeted networking
- If MID-LEVEL + START A BUSINESS: assess feasibility and propose
  a gradual transition plan
- If SENIOR + ADVANCEMENT: steer toward leadership or
  advisory roles, with personal branding
- If SENIOR + START A BUSINESS: leverage existing expertise
  and network

If GEOGRAPHIC constraint: include remote-first options.
If SALARY constraint: prioritize fast-income options.
If TIME constraint: propose an accelerated plan.

Look at the structure rather than the content: the intersection of experience and goal determines the substance of the advice, while the constraint acts as a filter applied on top of the result. Separating these two levels instead of combining everything avoids writing eighteen rules when nine will do—and lets you add a fourth constraint tomorrow without rewriting the entire tree.

Sort an entire list in one pass

The same mechanism turns a manual qualification task into a single operation. On a leads export, the conditions act as a scoring grid, and the table enforces a deliverable the sales team can use immediately, without reformatting.

Here is a list of 10 sales leads. For each, assess
priority and propose the appropriate follow-up action.

Scoring rules:
- If the lead visited the pricing page: HOT -> propose a call
- If the lead downloaded a white paper: WARM -> send a
  personalized nurturing email
- If the lead only subscribed to the newsletter: COLD -> add
  to the automated sequence
- If the lead requested a demo: VERY HOT -> call within 24h

For each lead, produce:
| Name | Behavior | Score | Recommended action | Personalized message (2 lines) |

Leads:
1. Marie L. - visited pricing + downloaded the "ROI Guide" white paper
2. Thomas D. - subscribed to newsletter 3 months ago, no activity since
3. Sophie R. - demo request received yesterday, company of 200 employees
[...]

One precaution before pasting a list like this: it involves personal data. Check what your plan and your company’s rules allow. On Business, Enterprise, and Edu plans, OpenAI does not use your content to train its models by default; on personal plans, it may be used for training unless you have disabled it in your data controls settings. In any case, keep only the information needed for sorting, for example a first name and an initial rather than a full name and address.

Produce multiple versions of the same message

A useful variant is to ask not for a choice, but for all branches at once. This is the typical situation for a sensitive announcement that must circulate to different audiences without ever contradicting itself from one channel to another.

We are preparing an announcement about a restructuring of the sales team.
Write 3 versions of the message based on context.

Scenario A - Team meeting (direct, transparent tone):
The announcement is made in person to the affected team.
Priority: honesty and addressing concerns.

Scenario B - Internal email to the entire company (factual, reassuring tone):
General information sent to all employees.
Priority: clarity and message stability.

Scenario C - Communication to clients (professional, confident tone):
Message intended for key clients who have a dedicated contact.
Priority: service continuity and trust.

For each scenario, produce:
1. The full message (200 words max)
2. The 3 likely questions and prepared answers
3. The main point of vigilance

Getting all three versions in a single response isn’t just a time saver; it’s a consistency win: the model has the other two in view while drafting each one, which wouldn’t happen if you opened three separate conversations.

Nest conditions

For truly complex trade-offs, nest conditions as you would in a scoping note. The example below makes the recommendation depend on a first budget tier, then, within each tier, on a second criterion that isn’t the same everywhere—sales cycle for extreme budgets, market maturity for the middle tier.

Help me choose the best customer acquisition strategy.

Decision parameters:
- Monthly budget: [AMOUNT]
- Sales cycle: [SHORT (< 1 month) / LONG (> 3 months)]
- Market maturity: [MATURE / EMERGING]

Recommendation logic:
IF budget < 5,000 EUR:
  IF cycle SHORT -> SEO + content marketing
  IF cycle LONG -> Organic LinkedIn + email nurturing
IF budget 5,000-20,000 EUR:
  IF market MATURE -> Google Ads + retargeting
  IF market EMERGING -> Content + industry events
IF budget > 20,000 EUR:
  IF cycle SHORT -> Google Ads + LinkedIn Ads + SDR
  IF cycle LONG -> ABM (Account Based Marketing) + events

For each recommendation, detail:
1. Budget allocation by channel
2. KPIs to track
3. Expected results at 3 and 6 months

Use conditions as guardrails

Conditional rules aren’t just for steering content: they can monitor output quality as it’s being written. Each “if” becomes a self-correction instruction the model applies to its own text, instead of leaving you to discover the problem during review.

Write an 800-word blog post on e-commerce trends in 2026.

Guardrails:
- If you mention a statistic, cite the source or indicate
  "estimate" in parentheses
- If you claim a trend is "the most important," justify
  with at least 2 arguments
- If a paragraph exceeds 100 words, split it in two
- If you use a technical term, add an explanation in
  parentheses the first time
- If you can't find a concrete example for a point, flag it
  with [ADD EXAMPLE]

Your turn

Build and test your own decision tree. Choose a task you repeat where the scenarios vary, then fit it into this matrix.

[YOUR RECURRING TASK]

Conditions:
- If [CONDITION A] then [ACTION A]
- If [CONDITION B] then [ACTION B]
- If [CONDITION C] then [ACTION C]
- If no condition applies: [DEFAULT ACTION]

Apply these conditions to:
[YOUR DATA]

Then test with at least three different inputs, chosen to trigger distinct branches. This test almost always reveals one of three classic flaws. A condition phrased in vague terms—“if the tone is average”—means nothing to the model: replace it with an objective criterion, or provide an example of each category. The absence of a default case leaves the model to improvise as soon as an input falls outside the expected frame, and that’s where incongruous responses appear; the “else / in all other cases” line is never superfluous. Finally, beyond six or seven branches, the prompt becomes impossible to maintain: you won’t know which one produced a given result. Simplify, or split it into several chained prompts.

Official sources: Best practices for prompt engineering with the OpenAI API · How your data is used to improve model performance (OpenAI Help Center)

Key takeaways

• Conditional prompts adapt the response to different cases using “if/then” rules.
• Multi-criteria branching lets you handle complex situations in a single prompt.
• Batch processing with conditions automates the analysis of entire lists.
• Multi-scenario prompts produce several versions tailored to different contexts.
• Always include a default case and limit the number of branches to maintain clarity.
• Before processing personal data, check what your plan and your company allow, and minimize data to the strict necessity.