Give AI the context it actually needs.

A vague question is missing your world

Context is everything the AI can see when it answers: your message, earlier parts of the chat, anything it has saved about you, and any document you paste or it looks up. It can't see your fridge or your week unless something puts them in front of it. It builds its reply by predicting likely text from what's in front of it (see How AI Actually Answers), so a bare question gets the most average answer there is.

Ask any chat assistant "What should I make for dinner this week?" and you'll get something like lasagna, beef stir-fry, mushroom risotto and chili. Fine dinners for nobody in particular. Now ask with a small context pack:

• Request: four weeknight dinners for two, one shopping trip.
• Stable preferences: 30 minutes max on weeknights; one of us won't eat mushrooms; we like spicy food.
• Recent: bought a big bag of spinach on Sunday; had pasta twice last week.
• Reference: this week's sale flyer from our grocery store, pasted in.

This time the plan uses the spinach, skips the risotto and the pasta, leans on what's on sale, and fits in 30 minutes. Nothing about the AI changed. What it could see did.

Four kinds of context

Building LifeQuest forced me to separate my information into layers. Four matter for everyday use:

1. The immediate request. What are we deciding right now? "Today's workout." "This week's dinners."
2. Stable preferences. Things that stay true about you: goals, constraints, dealbreakers. For me: martial arts, strength training, and talking like a human instead of filling out tiny boxes every time I eat lunch.
3. Recent history. What happened lately: the workouts I logged this week, the spinach you bought Sunday. This keeps advice from repeating yesterday.
4. Reference material. A trusted source for this question: a training note, an official page, a sale flyer. You paste it, or the AI looks it up.

For long histories there's a fifth helper: a summary that squeezes weeks of entries into a few lines. It only works if it keeps the signal and the uncertainty. (If you're unsure what to ask in the first place, that's a different step: Ask Why Before You Ask How.)

Memory is not the library

Some chat assistants remember things about you between conversations. That's handy for stable preferences. But what a chat remembers is a loose collection of things you once said, possibly incomplete or out of date. Looking something up means reading a trusted source at the moment you ask.

In LifeQuest I kept three piles apart: the current chat, my saved records, and a library of training notes. A logged meal was checked and saved to a record, not left to chat memory. When a question needed a training note, the app fetched the one that matched and showed where it came from. That step is called retrieval (you'll also see RAG, retrieval-augmented generation: look up the right document, then answer). In the comic's words, RAG is the librarian, not the boss.

So when a fact matters (a price, an opening time, a rule), don't trust what the chat "knows." Paste the official page and ask which source each fact came from.

Less, dated, and updated

Old facts can ambush you. In the comic, my assistant thinks I'm still on last year's plan. Nothing is broken; an old fact is sitting in the context and nothing says it was replaced. So date your pack, prefer recent facts that apply, and delete the ones that no longer hold.

Supply only what changes the decision. For each line, ask: would the answer change if I deleted this? Your favorite movie doesn't change dinner. The mushroom rule does.

Update the pack instead of piling on. Say you start a night class on Tuesdays. Don't tack "actually, ignore the 30 minutes on Tuesdays" onto an old list. Edit the line to "Tuesdays: 15 minutes max," change the date, and rerun. One current version beats a stack of corrections.