The modes that change the output
Extended reasoning
For hard, multi-step problems, ask the model to think step by step (or enable an extended-thinking mode). It trades speed for accuracy on math, planning, and analysis — dramatically fewer careless errors.
Context & files
Paste the actual document, data, or code. An assistant reasoning over your material beats one guessing from a vague description every time.
Tools & code
Modern assistants can run code, do exact arithmetic, and generate files. For anything numeric, "show your calculation" turns a plausible guess into a checkable result.
Structured prompting
State the role, the goal, constraints, and the output format. "Act as an editor; tighten this to 150 words; keep the second paragraph" outperforms "make this better."
Weak prompt vs. strong prompt
Help me with my resume.Strong:
You are a technical recruiter. Here is my resume [paste]. The role is a backend engineer job [paste description]. Rewrite my summary and top 3 bullets to match, keep it truthful, and flag anything missing.Is this a good investment?Strong:
Think step by step. Given these three options and their fees/returns [data], compare 10-year outcomes, state your assumptions, and show the math. This is a high-stakes decision, so be conservative.The difference isn't magic words — it's giving the model a role, real data, a clear goal, and a format. Those four ingredients are what separate "20% of the tool" from the other 80%.
Where the extra effort pays off
Not every task needs the heavy machinery. A rough guide: quick factual lookups → plain chat. A decision that affects money, a deadline, or code that ships → turn on step-by-step reasoning and paste real context. Empirically, for multi-step reasoning tasks, prompting a model to reason explicitly can lift accuracy from roughly 18% to over 50% on hard benchmark problems — the exact gain depends on the task, but the direction is consistent: more structure and context, better answers.