Four habits that beat prompt formulas: one outcome per prompt, a plain first attempt, feedback specific enough to act on, and a real check on every number, date and quote the AI gives you.
Most wasted time with AI comes from the way people ask, not from the tool. Long bundled requests, borrowed formulas and vague corrections all feel efficient and all cost you a second editing pass. The fix is a working rhythm of short steps: ask one thing, start plain, correct precisely, and check the facts that matter.
This guide is hands-on and task-level. It assumes you already use an AI chat tool and want fewer rounds of rework.
When three things are on your mind, it is tempting to write all three into one message and assume one trip will save two. It rarely does. As the source article puts it, stacking three requests into one prompt does not save three trips: it usually costs five.
A colleague can ask a clarifying question or tell you what they skipped. A single AI response does not do that by default. It tries to satisfy everything in one pass, and when the asks pull in different directions, something gets shortchanged without any signal of what. The reply still reads as complete, with full sentences and a confident tone, so the gap only shows later, usually in the part that mattered most.
Untangling is the real cost. If the reply blends three unrelated changes, accepting the two that are right means accepting, or stripping out by hand, the one that is not. Whatever you stated last or least specifically is the most likely to get generic treatment.
Three working rules:
Batching is fine when it is the same job repeated. "Summarise each of these five reports using the same three headings" is one instruction, so there is nothing to prioritise between. The dividing line is not the number of items. It is whether the asks are the same job repeated or different jobs competing for one response.
Seven-step frameworks and magic phrases circulate constantly. Most were true for a specific model at a specific moment. Models change every few months, and a phrase that mattered for one version often does nothing for the next, because the weakness it compensated for has been fixed. You cannot tell which formulas still apply just by reading them.
A memorised formula answers last year's model. A quick test answers today's.
Instead:
This is not the same as having no method. You still notice what changed between a good result and a bad one, and carry that forward as a working sense of what a given tool needs. The difference is that your habit comes from evidence rather than from a rule written for another model.
Some structures are worth keeping because they close a real information gap: state the audience, name the format, give one example. A simple test: if you can explain in plain terms why a technique should work, it is probably durable. If you cannot, it was probably specific to a version that has gone.
"Make it better" is the least useful feedback you can give a person, and it is just as useless for AI. The model cannot read your mind about what is wrong, so the quality of the next draft depends on what you say about this one. Vague feedback gets you a different draft. Specific feedback gets you a better one.
A good test: could you hand this feedback to a new hire and expect a specific change?
Two further habits help. Keep a running note of corrections you make often, such as phrases you always cut or a structure you always ask for, and paste that list in as a starting instruction so you stop repeating yourself. And if a draft is wrong in its basic approach, not just its execution, do not patch it for three rounds. Say plainly that the approach is wrong and why, then let it restart.
AI research tools are confident whether or not they are right, and confidence is not the same signal as accuracy. Treat every specific claim as a lead to check, not a fact to cite.
| Fine to delegate | Always verify |
|---|---|
| Finding starting points: sources, angles, search terms | Any statistic |
| Summarising long material you will read yourself anyway | Any direct quote |
| Outlining a write-up from findings you already trust | Any named study, date or number |
The two-minute check: ask the tool for its source, then open that source. If it does not exist, does not say what was claimed, or cannot be found, that is not a rare glitch. It is the model doing what it was likely to do with an obscure or specific fact. If you cannot verify a claim independently, treat it as unconfirmed.
Speed on the search, rigour on the facts.
Once a task is repeated or built into an agent, what content reaches the model matters more than how the prompt is worded, and retrieved content must be treated as untrusted.
Custom Agents & Tools · 4 minGuide · 20 September 2026Four structural prompt patterns that hold up on messy real documents, including showing the actual template, naming what to ignore, requesting uncertainty flags and chunking long files.
Team AI Training · 3 minGuide · 21 September 2026Four targeted checks to run before AI-assisted work leaves the building, covering sources, boundary cases, judgement calls and omissions, without redoing the whole task by hand.
Team AI Training · 3 minIf this is the question on your desk, a thirty-minute call tells you whether the service fits, or that you do not need us yet.