Four common mistakes stopping you from getting results from ChatGPT

Executive overview

Most people blame AI quality when they get poor results. The real causes are model choice, weak prompts, missing context, and wrong features.

Fixing these four mistakes — not learning prompt engineering — is what gets consistent, high-quality output from ChatGPT.

Choosing the right model

• Auto handles 60–70% of tasks; it delegates reasoning level to OpenAI.
• Instant is for quick facts under time pressure — 5–10% of use.
• Thinking (standard or extended) is the default choice for high-stakes tasks; extended reasoning gives the most quality for complex work.
• Pro is for narrow, highly specialised research (biopharma, physics); 8–10 min wait times make it rarely worth it for business use.

Prompting: answer what, why, and how

• State what you want — be explicit about the goal.
• State why — the end state and intent give the AI useful context.
• State how — set constraints: output format, writing style, process steps.
• If the output is poor, scroll back to the original prompt and fix the root cause rather than patching the response.
• Iteration is expected; use trial and error to build intuition over time.

Providing the right context

• Missing files are the most common cause of wrong output format or structure.
• For recurring reports, provide diverse sample outputs so the AI learns the format — this is few-shot learning.
• Too many or too-large files cause bloated context, which degrades AI intelligence.
• When providing a large file, explicitly direct where the AI should focus — state your role, which sections matter, and why.

Using the right features

Hallucinations

• Enable web search to ground responses in citations.
• Add an explicit instruction: require a citation for every fact, and specify trusted source types (government, research institutions).

Unwanted rewrites

• When asking for targeted edits, use the canvas feature.
• Canvas lets you select specific text and request changes to only that section, avoiding full rewrites that waste time and bloat context.

Repetitive instructions

• Memory is for broad cross-conversational preferences (location, dietary needs) — not task-specific constraints.
• Use custom GPTs or projects for recurring, constrained tasks like report writing or data analysis; these support system prompts and a knowledge base.