All pages in this guide
Your Work
One procedure, four kinds of reader. You're going to load the thing you actually work on into an AI tool so it stops guessing and starts giving answers only your material could produce.
This is an optional side-branch, not part of the main path. The main guide ends at Running a Team of Agents. If you followed that far, you already know everything you need. Skip this page freely. It's here for one reason: the single best move on the whole site is loading your real material into the tools, and this page walks four different kinds of reader through doing exactly that.
Everything earlier in the guide taught you that generic output is a grounding problem: the AI wasn't told enough, so it guessed. This page fixes that for good. You gather what you already have, write a short note about who you are, upload it once, and then the tool sounds like it knows you. Do it properly and you never re-explain yourself again.
It works the same in every tool. In ChatGPT and Claude the home for this is a Project, a saved workspace with your files attached to every chat. In Gemini you keep the files in a Gem or lean on its reach into your Drive. The words on the buttons differ; the move is identical.
1. The common procedure
Four steps. They're the same whoever you are. The only thing that changes is which files you gather. Read this once, then jump to whichever reader below sounds most like you for the concrete list.
- Gather the raw material you already have. Not new writing, existing stuff. Put it in one folder on your computer, call it something like
ai-context. Plain text, a PDF, a Google Doc, screenshots: whatever you've got is fine. - Write a short "who I am and what I want" file. Half a page, plain English. Who you are, what you're trying to do, how you like answers (short or long, formal or casual), and any hard rules ("never invent a fact, ask me instead"). This is the instructions file the tool reads before every reply.
- Upload it. Make a Project (ChatGPT or Claude) or a Gem (Gemini), paste your half-page into the instructions box, and attach every file from the folder. Once, permanently.
- Test it with a question only your files can answer. This is the step people skip, and it's the whole point. Ask something whose correct answer requires your material. If the reply would have been identical without your files attached, the setup failed: your files are too thin or you're chatting outside the Project. Add more, and try again.
"Summarize good study habits." "Write a friendly cover letter." Those are questions the AI could answer with no files at all, so they tell you nothing about whether your setup worked. Your test question has to name something private: a fact, a phrase, a preference, that lives only in your uploaded material. If it gets that right, the grounding is real.
2. If you're a student
You want a study partner that knows your course and writes in your voice, not a generic tutor.
Gather:
- Course materials: lecture slides, the syllabus, assigned readings, your own lecture notes.
- Two or three of your own past essays or answers. This teaches it how you actually write, so its drafts sound like you and not like a stranger.
- The marking rubric, if you have one. This is the highest-value single file you can add: it lets the AI grade a draft against the exact thing your professor will grade it against.
Your half-page: "I'm a second-year student in [course]. Help me understand material and improve my own drafts, never write final work for me to hand in. Match the voice in my past essays. When I share a draft, mark it against the rubric and point at the specific sentence that loses marks."
The test that proves it worked: "Grade this paragraph against my rubric and quote the exact criterion it misses." A generic tutor can't do that. One with your rubric can, and it'll point at a real line.
3. If you're job hunting
You want tailored applications that stay honest: grounded in your actual history, not a pile of invented achievements.
Gather:
- Your resume, and a longer "brag document": every role, project, and number you're proud of, in more detail than a resume has room for.
- Two or three target job postings, pasted in full.
- Notes from any calls or interviews: what the recruiter said mattered, what the team is struggling with.
- The story of each past role in your own words: what you actually did, what went well, what you'd change. This is what lets it write about you without making things up.
Your half-page: "I'm applying for [kind of role]. Tailor my applications using only what's in my history, never invent a job, a metric, or a skill I didn't list; if something's missing, ask me. Write in a plain, direct voice. When I paste a posting, show me which of my real experiences match which requirement."
The test that proves it worked: "Match my experience to this posting's requirements, line by line, and flag any requirement I don't actually meet." If it invents a match, your history file needs more; if it flags a real gap honestly, the grounding is working.
4. If you have a project or hobby you care about
A novel you're writing, a woodworking practice, a garden, a game you're building, a family-history dig: anything you return to and want a thinking partner for.
Gather:
- Reference material you keep coming back to: articles, manuals, the rulebook, the source texts.
- Your own notes: plans, decisions, dead ends, the running log in your head that you've half-written somewhere.
- Previous work: earlier chapters, past builds, photos of what you've made, whatever shows where the project stands now.
Your half-page: "This is my [project]. Here's where it stands and what I'm trying to do next. Keep track of the decisions in my notes and don't contradict them. When I'm stuck, ask about my constraints before suggesting anything, and keep suggestions concrete."
The test that proves it worked: ask it something that depends on a decision buried in your notes, "given the direction I committed to in chapter two, what's a problem I'm about to run into?" If it references your actual choice, it's reading your material, not the internet's average.
5. If you run a small business
You want a tool that drafts in your voice and knows your real prices, services, and customers, so nothing it writes needs heavy fixing before it goes out.
Gather:
- Your website copy and your "about" page: how you already describe yourself in public.
- Five to ten past emails, captions, or newsletters you were happy with. This is the voice sample, and it matters more than anything else here.
- Your price list and services: packages, what's included, what's not.
- The questions customers actually ask, and how you answer them.
Your half-page: "I run [business]. My customers are [who]. Write in my voice, using the samples attached, warm and direct, no corporate filler. Use the real prices, links, and services from the files; never invent or guess them. If a fact's missing, ask before writing anything long."
The test that proves it worked: "Write a reply to a customer asking about [a specific service], and use my real pricing." If the price and the tone are yours, not a stock template's, the setup holds. If it hedges or invents a number, add your price list and try again.
Notice what didn't change across all four: not the tool, not the plan, not how clever your questions are. The only thing that turned a generic answer into a specific one was handing over real files. A thin folder gets thin answers. An honest, detailed one gets a tool that sounds like it knows you. Start with two files if that's all you can face today, and add more each time you catch it guessing.
6. Keep it clean as it grows
One workspace per thing you care about. When a genuinely separate project shows up, a new course, a side idea, a different hobby, make a new Project or Gem rather than piling it onto an existing one. Leftover clutter from an unrelated topic quietly drags down every answer, the same way an over-full chat does. A folder of four or five tidy, single-purpose workspaces beats one crowded catch-all every time.
This one habit, gather, write a short instructions file, upload, test with a question only your files can answer, is the difference between people who find AI vaguely useful and people who find it genuinely worth the time. Everything else is refinement.
Quick check
You've uploaded your files and want to know whether the setup actually worked. What kind of question tests it properly?
Answer: the whole point of grounding is getting answers only your material could produce. So the test has to depend on something private: a fact, a phrase, a rubric line, a price, that lives only in your uploads.
If a question could be answered just as well with no files attached, it proves nothing. A tool will happily claim it read your files whether it used them or not; the real proof is a right answer it couldn't have guessed.