Advanced Prompting: Workflows, AI Agents and Multimodal

From isolated prompt to agentic workflow

An isolated prompt is enough for a short task: summarize, rewrite, generate an idea or structure a draft. As soon as the task involves several steps, tools, external data, validation or actions to perform, you need to think in workflows rather than single prompts.

An agentic workflow describes who does what, in what order, with which sources, which tools and which control points. The goal is not to let AI act without a frame, but to turn a complex task into observable, verifiable and correctable steps.

When to move to a workflow

• The task requires several dependent subtasks.
• The model needs to consult documents, search information or use a tool.
• The result must be validated before publishing, sending, editing or deciding.
• The same process must be repeated regularly by a person or team.
• An error could affect customers, finances, legal matters, security or reputation.

Minimum structure of an AI workflow

1. Define the final goal and success criterion.
2. Break the task into short steps with an expected output for each one.
3. Identify allowed sources and tools for each step.
4. Define limits: forbidden data, forbidden actions and topics to escalate.
5. Add a control after critical steps: verification, review or human validation.
6. Document the workflow version, prompts used and errors observed.

Tip: before creating an agent, write the workflow by hand. If you cannot explain the steps, tools and validations, the agent will be difficult to control.

---

Multimodal prompting: documents, images and screenshots

Modern prompting is no longer limited to a written question in a chat window. AI assistants can work with documents, tables, images, screenshots, diagrams or presentations. The key skill is therefore to guide the analysis of the source, not only to phrase a general request.

A multimodal prompt should specify what the model should inspect, what it should ignore, the expected level of detail and the shape of the answer. The richer the source, the more you need to frame the task: summarize, extract, compare, verify, transform or suggest an action.

Structure of an effective multimodal prompt

• Goal: state the decision or deliverable expected.
• Source: name the document, image, screenshot or table to analyze.
• Focus area: specify the pages, columns, visual elements or passages that matter.
• Criteria: explain what will make the answer useful or usable.
• Format: request a summary, table, checklist, action plan or list of anomalies.
• Limits: ask the model to flag what it cannot confirm from the provided source.

Example exercise

Take a screenshot, document excerpt or business table. First ask for a simple description, then ask for an action-oriented analysis: things to fix, risks, priorities and questions to ask. Compare both outputs to understand how framing changes the value of the answer.

Tip: in a multimodal prompt, do not only say “analyze this image” or “summarize this file”. Say what you are looking for, why you are looking for it and how you want to use the answer.

---

Test a prompt across several AI models

A strong prompt should not only work in one specific tool. ChatGPT, Claude, Gemini, Copilot or an internal model may react differently to the same context, constraints or output format. A reference-level course must therefore teach how to test prompt portability.

The right approach is to separate what is stable from what is model-dependent. The goal, context, success criteria and useful data should stay clear. However, instruction length, how you ask for reasoning, the level of structure and style guidance may need variants.

Multi-model test protocol

1. Choose a real task and define the expected result before testing models.
2. Write a baseline prompt with goal, context, constraints, format and criteria.
3. Test the same prompt on two or three models without changing the initial request.
4. Compare the answers on five criteria: accuracy, clarity, structure, usefulness and risks.
5. Identify what should become a model-specific variant.
6. Document the reference version and required adaptations in your prompt dossier.

What to compare

• Does the model respect the requested format, or should the output constraint be stronger?
• Does it need more context to produce a usable answer?
• Does it answer too long, too short or with excessive confidence?
• Does it handle provided examples correctly, or does it mix them with the real task?
• Does it flag uncertainty, or does it invent details missing from the context?

Tip: keep one canonical version of the prompt, then add small model-specific variants. This avoids maintaining three completely different prompts for the same task.

---

Going further

These skills are built through practice. Resume the structured path in the free prompting course, and train with the interactive exercises: prompt chaining, mega-prompt, chain of thought.