The modern AI playbook: from chat to collaboration
Using AI well is no longer about testing a chatbot with trivia. The useful shift is treating AI as a capable but context-blind colleague: fast, flexible and helpful when the brief, source material and review process are clear.
Use the right tool for the job
Putting the wrong AI model into a workflow can move confusion faster rather than solve it. Different tools are better suited to different kinds of work, so start by matching the tool to the task rather than defaulting to the one already open in your browser.
| Platform | Best used for | Standout strength |
|---|---|---|
| Claude | Complex logic, long-form content, coding tasks and multi-file analysis. | Projects and artifacts that separate structured work from ordinary chat. |
| ChatGPT | Brainstorming, rapid prototyping, general drafting and conversational exploration. | Flexible interaction across text, voice and practical everyday tasks. |
| Gemini | Research, Google Workspace activity and large-scale document or data parsing. | Large context windows and close alignment with Google-based workflows. |
| Perplexity | Replacing rough web searches and compiling cited market or topic intelligence. | Search-led answers with linked sources for faster discovery work. |
Start with a prompt, but succeed with context
A single-sentence request usually produces a generic answer. Stronger AI work comes from context engineering: giving the model enough role, task, background and output structure to produce something useful on the first or second pass.
- Role: give the AI a useful point of view, such as senior business analyst, operations consultant or data quality reviewer.
- Task: define the objective, constraints and level of detail required.
- Context: provide background data, source material, audience, assumptions and any rules it must follow.
- Format: say whether the output should be a table, checklist, summary, JSON structure, email draft or set of recommendations.
Weak prompt: Write a report on our customer onboarding process.
Better prompt: You are an operations consultant. Analyse the attached onboarding data, identify the top three bottlenecks, format them in a markdown table, and suggest an automation-first improvement for each.
Use projects and source files to ground the work
The most useful AI workflows are grounded in the material that already matters to the organisation. Instead of copying the same background into every chat, create a project or workspace and load it with stable source files: standard operating procedures, tone-of-voice notes, data definitions, process maps, schemas, previous templates or approved examples.
This helps the AI stay closer to the way the organisation actually works. It also gives the human reviewer something concrete to check against, which matters when the output affects customers, reporting, internal tools or operational decisions.
Chain prompts instead of asking for everything at once
Large, multi-step requests often produce shallow or confused answers. Break the work down. Ask for an outline first, review it, then move section by section. If the task involves data, systems or internal process, add checkpoints where the AI explains assumptions before continuing.
For practical build work, show the AI what has changed at regular points. Treat it like a collaborator that needs orientation, not a system that should be left alone to invent the final answer.
Keep the human feedback loop active
The first output is a prototype. If it is off-target, do not simply start again. Steer it: explain what is useful, what is missing, what tone is wrong, what needs evidence, and what should be shorter or more concrete.
This is where AI becomes operationally useful. The value is not just faster drafting; it is faster iteration around a clear goal, with a person still responsible for judgement, quality and final decisions.