What most people get wrong when starting with ChatGPT

Executive overview

Most people using ChatGPT are getting roughly 10% of its value — defaulting to one model, ignoring customisation, and treating it like a search engine. The gap isn't the tool; it's how it's used.

Pay for Plus, build custom AI projects for your recurring tasks, and learn which model fits which job. These three shifts unlock the majority of the productivity gain.

The core insight: 70-80% of your ChatGPT use should be custom-primed AI, not the default chat interface.

Pay for Plus and why it matters

• Free plan gives no model choice and no customisation — both are critical
• Plus is ~$20/month; the productivity return far exceeds the cost
• Model selection and custom instructions are gated behind the paywall

Custom AI: projects vs. GPTs

• Projects are private, support larger knowledge bases, and are where most day-to-day work should happen
• Custom GPTs are shareable and suited for distributing a workflow to others
• Both use a system prompt (instructions) plus an uploaded knowledge base
• The system prompt should define persona, instructions, process, and output format
• Aim to run 70-80% of sessions inside custom projects, not base ChatGPT

Choosing the right model

• GPT-4o — fast, good for basic fact-finding and simple questions; the default for most users
• O4 mini — reasoning model, useful for quick concept lookups where you can wait a few seconds
• O4 mini high — slightly smarter and slower; good for AI-led interviews to extract and structure your thinking
• O3 — the most capable Plus model; use for complex analysis, coding, debugging, broad research, and data assessment
• GPT-4.1 — strong at following hyper-specific instructions; good for coding simple tasks and structured writing
• GPT-4.1 mini — cheap and fast; mainly useful when building apps via the API
• GPT-4.5 — best for creative writing and human-sounding prose; likely to be replaced by GPT-5

Features worth using

• Web search — gives the model access to real-time data beyond its training cutoff; pair with O3 for market analysis, product comparisons, or current events. Turn it off after the research phase to avoid repeated lookups mid-conversation
• Deep research — runs for 15-30 minutes and produces a long-form report; suited for in-depth topic analysis, aggregating multiple reports, or trend analysis over time. Use the O4 mini high interview technique first to sharpen the prompt before launching
• Canvas — a live document editor with AI; use for iterative writing (blogs, newsletters) and code. Highlight specific lines to change them without rewriting the whole document
• Image generation — best for social media graphics and blog visuals. For higher quality: have O3 research prompting best practices for the 4o ImageGen model, generate three prompt variations, convert to JSON, then feed into a fresh ImageGen conversation
• Agent mode — useful for research-then-action tasks (find a product and purchase it, research restaurants and book one, pull data and build a spreadsheet); not worth using for research-only tasks
• Voice mode — voice-to-voice; useful for therapy/advice sessions, sales practice against a simulated persona, and presentation or negotiation rehearsal

Prompting: outsource it to the AI

• Ask the model to research best practices for prompting a specific model, then have it write the prompt for you
• For critical tasks, run the draft prompt through a dedicated prompt generator (Anthropic Console or OpenAI Playground) to cut length by ~60% while keeping what matters
• Anthropic's generator is preferred for concision and XML delimiter structure
• Delimiters (XML or Markdown) separate sections of a prompt so the model can parse them correctly

Mindset: abundance over scarcity

• One bad answer is not evidence the tool doesn't work — it usually means insufficient context or a weak prompt
• Run the same question multiple times across different conversations; these models are probabilistic
• Try the same prompt on different models to find the best output
• Give the model explicit context; never assume it knows your situation

Going beyond ChatGPT

• Once comfortable with ChatGPT, experiment with other providers: Claude, Gemini, Grok, Perplexity
• No need to switch entirely — test alternatives for high-stakes use cases where the quality difference could be significant
• A new model from another provider may outperform ChatGPT on your specific task by a large margin