Last Tuesday a friend forwarded me an email she’d been avoiding for three days. One line: “Hi, just following up on the proposal.” She’d asked a free AI chat to write it, hated the result, and closed the tab. The tool wasn’t broken. The request was.
What she typed was something like “write a follow-up email to a client.” That’s a description of a category, not an instruction. Free chatbots handle those badly because there’s nothing to grab hold of. Give the same model the actual proposal, the client’s last reply, and one sentence about what you want to happen next, and you get something you can send after a single edit.
That gap is what this guide covers. Seven steps, all of them doable inside free tiers, with the level of detail that separates a draft you rewrite from a draft you use.
Step 1: Match the Free Tool to the Job
Free tiers are not interchangeable, even when the interfaces look identical. The differences that actually bite:
- Context window — how much text you can paste before it quietly forgets your first paragraph.
- File handling — whether you can upload a PDF, spreadsheet, or screenshot at all.
- Web search — whether it can check current facts or is stuck at its training cutoff.
- Message limits — how many prompts before a cooldown or daily cap stops you mid-task.
For a 200-word rewrite, open whichever tab is already there. For “summarise this 40-page PDF and pull out the five commitments,” use the one with the biggest context window and file uploads. For anything about a launch from last month, switch search on or you’ll get confident nonsense.
If you’d rather not test all of this by hand, a breakdown of what free AI chat tools really give you for nothing covers the current tiers and limits side by side.
Step 2: Paste the Material, Don’t Describe It
This is the single biggest upgrade available to you, and it costs nothing. People describe their inputs to the model, then wonder why the output sounds like everyone else’s.
Bad: Write a product description for a burr coffee grinder.
Better: paste in your three most useful customer reviews plus the spec line, then add: Using only the details above, write a 90-word product description for someone replacing a blade grinder. Lead with the grind-consistency complaint.
The second version writes itself because every fact is already sitting on the table. The first one forces the model to guess, and guessing is exactly where generic phrasing comes from.
Step 3: Build the Prompt in Three Parts
Task, context, constraints. That order, every time.
Task: turn these raw notes into a 150-word project update for a client.
Context: the reader is non-technical, approved the budget in March, and has already heard the timeline slip once.
Constraints: past tense, no jargon, name the two blockers, finish with the single decision I need from them by Friday.
Constraints do more work than politeness. “Shorter” beats “please be concise” because you can count words. Useful categories to spell out:
- Length in words or bullets, not adjectives.
- Format: table, numbered list, three short paragraphs.
- Tone with a reference point (“like you’d explain it to a new hire”).
- Exclusions: no exclamation marks, no “leverage”, no invented statistics.
That last one matters more than people expect. A free chat model will happily fill a gap with a plausible number if you don’t tell it to leave the gap alone.
Step 4: Ask for the Skeleton Before the Prose
One message that saves three rewrites: Give me five headings for a 900-word guide on X, and tell me what each section has to prove.
Then argue with the list. Cut one heading, merge two, move the strongest point to the top. Once the skeleton is right, “now write section three in 180 words” produces far better text than “write the whole thing,” because the model is filling a defined space instead of inventing a structure and the content at the same time.
It also stretches your free message limit. Outlining burns a fraction of the tokens that a throwaway 900-word draft does.
Step 5: Chain Short Prompts Instead of One Giant One
Long prompts feel efficient. They usually aren’t. Three short prompts in sequence beat one sprawling request almost every time.
Here’s a chain I use on messy sources: extract every date, name and action from the block of text below as a three-column table. Then: group those rows by owner. Then: write a 120-word update for the one person who owns nothing and just needs the summary.
Each step is mechanical and a bit ugly. That’s the point. The polished writing happens once, at the end, working from clean input instead of a fog of half-sentences.
If you find yourself running chains like this most weeks, it’s worth building them on purpose rather than improvising. Setting up a lightweight control plane for AI agents is the same idea scaled up: fixed inputs, fixed steps, output you can actually review.
Step 6: Keep a Prompt Notebook
A plain text file is enough. Dated entries, one line each: which model, what you asked, whether the result was usable.
Mine has entries like “inbox triage / 3-bullet reply / works every time” and “contract summary / loses the party names / don’t bother”. After a fortnight you have six prompts you trust and you stop rebuilding them from scratch. The failures are worth logging too, because they’re the ones that waste twenty minutes before you notice.
A Worked Example: Messy Meeting Notes to Client Update
Say you have 600 words of rough notes from a 45-minute call. Scattered names, a few numbers, several half-finished sentences.
Prompt 1: Extract every decision, deadline, and open question from the notes below into a three-column table. Use only what’s there.
Prompt 2: From that table, list anything with no owner or no date attached.
Prompt 3: Write a 130-word update for the client. Decisions first, then the two deadlines, then the one question I need answered by Friday.
Four minutes of prompting, plus your edit. The trap is trying all three in one prompt, which lets the model blend a firm decision with a vague suggestion and hand you a sentence you can’t defend in the next call.
If call capture is a recurring problem rather than a one-off, getting Fireflies.ai running in an afternoon gives you clean transcripts to feed straight into those same three prompts.
When Free Chat Quietly Stops Being Free
Free tiers are genuinely capable, right up until you hit the pattern that costs you money anyway: a cooldown in the middle of a task, a 12,000-word document that needs chunking, or a deadline where a hallucinated statistic matters.
Two signals tell you it’s time to pay for something. First, you’re opening a second tool to finish what the first one couldn’t. Second, you’re paying for the same work twice: once in time spent working around limits, and again in a subscription that mostly sits unused. The argument that your AI bill behaves like a toll booth is worth reading before you add a third tool to the stack.
Until then, the free tier plus a decent notebook covers more than most people assume.
Run the Whole Loop on One Boring Task This Week
Pick something dull and recurring. A weekly status email. Summarising supplier invoices. Turning a support ticket into a knowledge-base answer. Run all seven steps on it once, properly, even if the first pass takes twenty minutes and feels slower than just doing it by hand.
It is slower, the first time. The second pass isn’t. By the fourth you have a prompt you trust, a line in your notebook, and a job that takes three minutes instead of twenty-five.
Nothing about that depends on a paid plan or a cleverer model. It depends on refusing to type “write me an email” and giving the thing something real to work with.