DeepSeek Chat is free, opens in a browser tab, and never asks for a credit card. That low barrier is also the trap. Most people type “write me a blog post about productivity,” skim whatever comes back, and quietly decide AI is overhyped. The tool isn’t the problem. The method is.
What follows is the workflow I use for real work: untangling broken scripts, turning rambling meeting notes into task lists, and pulling research I can actually check. No API key, no paid plan, nothing you can’t do this afternoon.
Pick the Right Mode Before You Pick the Right Words
Once you’re signed in at chat.deepseek.com, you’ll see two toggles above the input box: DeepThink and Search. They change the quality of the answer far more than clever phrasing does.
- Both off — fast and cheap-feeling. Good for rewriting, brainstorming, formatting, translation, casual questions.
- DeepThink on — the model reasons through the problem before it answers. Slower, and clearly better at maths, logic, and planning a change across several files.
- Search on — pulls live web results. Use it for anything that happened after the training cutoff: prices, releases, this week’s news.
- Both on — for questions needing fresh facts and reasoning, such as “compare current pricing for three managed hosts and recommend one for a 50,000-visit blog.”
If the platform itself is new to you, there’s a plain-English explanation of what DeepSeek Chat is and why it costs nothing. Short version: it’s a general-purpose assistant, not a single-trick toy.
The Four-Part Prompt That Fixes Most Weak Answers
Vague prompts get vague answers. Every time, hand over four things in this order.
- Context — who you are, who reads the output, what they already know.
- Task — one clear verb. Summarise, rewrite, extract, compare, debug.
- Constraints — length, tone, what to avoid, which data to use.
- Format — a table, a numbered list, plain paragraphs, JSON.
Here’s the difference in practice. The weak version: “Write an email to my client about the delay.” The workable version: “Context: I’m a freelance web developer. My client runs a three-person bakery and doesn’t know technical terms. Task: tell them the new online ordering page slips by two weeks because the payment provider rejected our integration. Constraints: under 150 words, warm but not grovelling, no jargon, new delivery date of 14 March, offer a call. Format: subject line, then the email body.”
Same tool, same afternoon, completely different result. The second prompt takes 40 seconds to type and saves you a rewrite.
Three Tasks Worth Doing Today
Turn messy meeting notes into decisions and owners
Export your transcript or notes as a .txt or .docx, upload it, then ask: “From the attached transcript, list every decision that was actually made, with the person who owns it and any deadline mentioned. In a separate list, show topics discussed but not decided. For each item, quote the single line that supports it.”
That last instruction is the one people skip. Demanding a supporting quote makes fabrication obvious — if the line isn’t in the transcript, the row is invented. Also check your employer’s policy before uploading internal recordings to any cloud service.
Debug code without pasting your whole project
“My Python is broken” gets you a lecture. Give it the error, the smallest function that fails, what you expected, what happened instead, and your library versions. Then add: “Give me two hypotheses ranked by likelihood, and the single smallest test that would tell them apart.” You get a research plan rather than a guess, and you learn something reusable. Switch DeepThink on for this one.
Research you can cite
Turn Search on and ask for structure: “Summarise the current state of X. Give three sources with links, the date of each, and flag any point where the sources disagree.” The disagreement flag is the useful part — it tells you where a claim is genuinely contested rather than settled. Open the links yourself before you quote them.
Working With Long Documents
File uploads handle PDFs, spreadsheets, and images. For a 90-page report, don’t ask for a summary of everything at once; you’ll get a vague blur. Work chapter by chapter, keep the notes, then paste those notes back for a final synthesis.
Ask for output you can paste straight into a spreadsheet: “Return this as a table with columns: claim, source, date, confidence.” A note on the platform itself — DeepSeek’s V4 release pushed cost and capability further, but the technique stays the same whichever version you land on.
Where It Gets Things Wrong
It will occasionally state something false with total confidence, including invented citations and tidy-looking numbers that don’t add up. Two habits cut most of that risk.
First, give it permission to be uncertain: “If you’re not sure, say so and tell me what you’d need to check.” Second, make it show the working: for any calculation, ask for the arithmetic step by step. Then run one of the middle steps yourself. It’s a 20-second sanity check that catches most of it.
Be extra careful when you chain models together into automated workflows. Autonomous agents have already shown they’ll cooperate in unexpected ways when a goal rewards it, and a pipeline quietly going off-script is much harder to notice than a single bad answer in a chat window.
Running It Your Way
If you’d rather not send work to someone else’s servers, you have options. A self-hosted interface like Open WebUI gives you a familiar chat front end pointed at whichever model you choose. For something with no server at all, Jan AI runs models directly on your laptop and keeps working offline. The trade-off is honest: you give up DeepSeek’s largest models, and in return the data never leaves the machine.
Small Habits That Compound
- One task per chat. Mixed threads produce muddy answers, so start fresh when you switch jobs.
- Keep a plain text file of prompts that worked. After a month it becomes your real productivity tool.
- Ask “what would change your answer?” when a conclusion matters. It surfaces the assumptions you should test.
- Paste the finished draft back and ask for the three weakest claims in it.
- Rename chats by project so you can find the good one three weeks later.
A 20-Minute Exercise to Make This Stick
Pick one genuine chore from this week — a supplier comparison, a policy document you’ve been avoiding, a script that keeps failing. Run it through the four-part prompt, with the right toggles on. Then compare the result against the version you’d have written on your own.
The gap is usually in the constraints, not the model. Once you notice which constraint did the work, save that prompt. Three rounds of this and you’ll stop thinking about how to prompt DeepSeek Chat at all — you’ll just be reaching for the saved snippet that handles the job.