Guide
AI prompts vs plain words: we sent one task to our agent as a long prompt and as a normal request
Prompt templates promise better AI answers if you write “You are an expert…”, add a constraints block and ask it to think step by step. We tested that on our own agent. The same dinner plan went in twice, in two clean chats: as a 100-word structured prompt and as a plain request of 36 words. Below are both real answers, what the prompt actually changed, and the few details worth putting into any request.
The short answer
You do not need prompt templates to work with an AI agent. Write the way you would brief a colleague: what you want, who it is for, the numbers that matter and anything that must not happen. A couple of normal sentences with those details gets an answer as complete as a long prompt with a role, capital-letter headings and “think step by step”.
We checked it on one task. The same dinner plan went to our test agent twice, in two clean chats: once as a structured prompt of 100 words, once as a plain request of 36. The prompt changed the layout of the answer. The plain request got seven dinners of the same kind, the budget, the allergy checks and a costed shopping list. Then a short follow-up in normal words did what no template can: it fixed the plan after the fact.
Why prompt templates exist
Prompt libraries grew up around chatbots that start every conversation from nothing. If the model knows nothing about you and you get one shot, packing the role, context, constraints and format into a single message is a sensible way to make that shot count. That is where the familiar recipe comes from: “You are an expert…”, a task block, a constraints block, an output format, a tone, and a line asking the model to reason step by step.
An agent works differently. It keeps the conversation going, asks when something is missing, and remembers your preferences and projects between chats. On WAM Agents the memory lives in your own container and you edit it by conversation: “forget the old office address” is the whole interface. So the context a template tries to pack into one message builds up on its own, and the most useful part of any prompt, the facts, fits into an ordinary sentence.
The test: one task, two ways of asking
The task is an ordinary family one. Plan dinners for next week for four people, with a nut allergy in the family, a grocery budget of about 80 euros and little time to cook on weekdays. The family and the numbers are made up for the test. The agent's replies are real, from our test agent on the default model on 30 October 2026, quoted as they came and shortened in places, never added to.
Each version went into a fresh chat after /clear, so neither answer could lean on the other, and the agent's memory had nothing about this family. The plain request went first.
1. The plain request: two sentences, 36 words
This is how most people would actually ask, typed on a phone between two other things.
Twenty-four seconds from the message to the answer. Every condition from the sentence is there: four people, nut-free, weekdays under half an hour, the budget with room left over. Nobody asked for a shopping list with prices, a section on where nuts hide, or a note that the lentils need no soaking. The agent added them because that is what a person planning a family week needs next.
Look at the first paragraph too. The agent kept the week vegetarian “to match how you eat”. That came from its memory of the account it works for, not from the request, and it said so openly and offered the fix in the same line. This is the part a template has no slot for: an agent fills the gaps in a short request with what it already knows about you.
2. The same task as a structured prompt: 100 words
Now the version you find in prompt guides: a role, a task block, constraints, an output format, a tone and “think step by step”. Same family, same budget, a clean chat.
Thirty seconds this time. The prompt got exactly what it asked for: the table with prep times, a list grouped by store section, the weekend allowed to run longer. The content is the same kind of week as before, with meat this time, because the prompt described the family in full and the agent treated it as the whole picture. The vegetarian option came back as an offer at the end.
One line in the prompt did nothing visible: the 20 years of experience. In an earlier run, where the structured prompt went into an ongoing chat, the agent answered that line directly: “I'll skip the "20 years as a nutritionist" framing — I'm not one, and I won't pretend to be. What I can do is a practical, costed plan.” That is a good answer, and it tells you where the effort in a prompt should go. The role is decoration. The facts are the request.
Side by side: what the prompt changed
| Plain request | Structured prompt | |
|---|---|---|
| Length of the request | 36 words, two sentences | 100 words, six blocks |
| Seven dinners, nut-free | yes | yes |
| Weekdays within the time limit | yes, 20–25 min | yes, 25–28 min |
| Budget | about €57 of €80 | about €73.70 of €80 |
| Where nuts hide on labels | yes | yes |
| Shopping list with prices | yes, as a list | yes, by store section |
| Layout | list by day | table by day |
| Next step offered | a Google Sheet or swaps | a vegetarian version |
| Time to answer | 24 s | 30 s |
The difference sits in two rows: layout and grouping. Both came from the “output format” block, and both fit in a few plain words: “as a table, please” or “group the list by aisle”. Everything else in the long prompt repeated what the short one already said in normal language.
3. A follow-up in plain words does the rest
A template is a single shot. A conversation is not. After the prompted plan we added two facts that the prompt had left out, the way you would add them in real life, without restating anything.
Twelve seconds. The agent swapped one day, checked the whole week for the second condition and said so, recalculated the basket line by line and kept a fallback for a bad evening. The message had 19 words, no structure and two different requests in it.
This is where plain words beat a perfect prompt. Nobody writes the full brief on the first try, and with an agent you do not have to: the second message carries only what changed.
4. A messy message works too
Real requests rarely look like a brief. They come in lowercase, with a sigh at the start and the important numbers somewhere in the middle. This one is about rent.
Fifteen seconds. The agent picked out every fact that matters for the letter: the 15%, January, four years, always on time, the 5% counter-offer. It wrote in the register the request asked for, polite and ready to send. Then it added two pieces of negotiation advice and a sensible check before sending. No role, no format block, and the result is a letter you can paste into a messenger.
5. We asked the agent itself
Before the test we put the question directly to our agent, in a separate chat.
Nine seconds, and it is the most useful paragraph in this article. The agent names the three things that make a request good: the task, the person or the deadline, and the shape of the result when it matters. Everything else it handles by memory and by asking.
What actually makes a request good
The test points to a short list. None of it needs special syntax.
- Say the task as a goal. “Plan dinners for next week” beats “generate a meal plan” only because it is how you think about it, and that is enough.
- Give the numbers and the hard limits. Four people, 80 euros, a nut allergy, 30 minutes on weekdays. These shaped both answers in our test, the plain one and the prompted one.
- Name what must not happen. “Nothing with nuts”, “no mushrooms”, “don't apologise for the price”. One line like this saves a whole round of corrections.
- Ask for the shape only when you care. “As a table”, “three short options”, “short enough to send as a text”. That is all the output format block of a template really does.
- Add details as they come. The second message can carry only what changed, as the Wednesday swap showed.
- Tell the agent once what to remember. “I'm vegetarian”, “my son has a nut allergy”, “sign letters as Maya”. Next week the request gets shorter still.
For the writing jobs a working week is made of, see AI writing assistant in Telegram: five text jobs, from a payment reminder to a post in Spanish. For a wider set of errands sent in the same plain way, read AI chatbot in Telegram: 10 real tasks we gave an agent, with its replies.
How an agent differs from a prompt box
Chatbots like ChatGPT and Claude also have memory and also understand plain requests well. Prompt guides are useful there for one-off work: a long, carefully structured message is a fair way to get a precise first answer from any model.
An agent on WAM Agents adds the part around the model. It lives in your Telegram chat, remembers you and your projects between chats, and acts on the request: with Google connected, the shopping list above becomes a Google Sheet in your own Drive, and a reminder to buy the fish on Thursday becomes a message on Thursday. It works in DMs, groups and forum topics. It starts in Telegram, and other messengers can be connected. The longer you work with it, the less each request needs to say, because the context already sits in its memory.
What stays with you
What it costs
Free gives 200 in the first month and 100 every month after, with no card. Pro costs $13.90 a month with 400 , or $1.39 for the first month if you subscribe within 7 days of signing up. Both plans run the same agent, and plain-text work like this is included in both.
On the default model, text costs 1 per 20,000 tokens, and a message in a thread costs a fraction of a magnet, as the pricing page lists. In our test, by the balance in the agent's status line, each of the two dinner plans came to about 4 , the Wednesday swap to 3 and the landlord letter to 1 . The long prompt cost the same as the short request.
By default requests go through DeepSeek infrastructure in the People's Republic of China, and you confirm this during onboarding. On Pro you can attach your own model, Claude or Codex.
Messages worth copying
Can you help me plan dinners for next week? There are [number] of us, [allergy or diet], about [budget] for groceries, little time on weekdays.
Swap [day] for something I can make in 15 minutes.
Same plan, but as a table, please.
Remember that [fact about you or your family] for next time.
Write something polite I can send to [person]: [what you want and the numbers].
Create your agent, write your first request the way you would say it out loud, and add the details in the next message.
Try it with your own agent
Free, no card. The agent starts working in the chat straight away.
Create your AI agentQuestions
Do I need to learn prompt engineering to use an AI agent?
No. In our test the same dinner plan came back just as complete from a plain two-sentence request as from a 100-word structured prompt. What matters is the facts you give: what you want, for whom, the numbers and anything that must be avoided.
Then what do prompt templates actually change?
Mostly the layout. Our structured prompt asked for a table and a shopping list grouped by store section, and that is what came back. The plain request got a list of seven dinners by day, the same kind of allergy checks and a costed basket. If you want a table, say “as a table” in normal words.
Does “You are an expert with 20 years of experience” help?
It adds nothing the agent needs. In an earlier run our test agent answered the role line directly and said it would give a practical, costed plan instead of pretending to be a nutritionist. The facts in the request did the work.
What if I forget an important detail?
Add it in the next message. The agent keeps the conversation, so “Wednesday I work late, swap it for something quick” changed one day and the shopping list in 12 seconds, without repeating the rest of the request.
Does the agent remember what I told it last time?
Yes. It remembers your preferences, contacts and projects between chats, and you can ask it in plain words to add, change or forget a fact. That is why the next request can be even shorter than the first one.
How much does one request like this cost?
Text on the default model costs 1 per 20,000 tokens. In our test each dinner plan came to about 4 by the balance in the status line, the follow-up to 3 . Free gives 200 in the first month and 100 every month after, with no card.
