How to Give ChatGPT Context About Your Company

Odette Marchetti runs a 6-person bakery and catering business in Providence. She'd tried, more than once, to get ChatGPT to help draft vendor emails and seasonal menu proposals that actually sounded like her business — local-first sourcing, no artificial flavoring, a standing note that wedding orders need six weeks' lead time. Each attempt lived in a single chat: useful for that conversation, then gone the next time she opened ChatGPT for something else. She'd pasted the same three paragraphs of context into new chats a dozen times before asking, reasonably, whether there was just one place to put all of it once.

There isn't — not one place. ChatGPT splits "context about your business" across four features that do genuinely different jobs. Using the right one for the right piece of information is most of what "giving ChatGPT context" actually means in practice.

Start With Custom Instructions for Standing Preferences

Custom instructions are the right home for anything that should apply to every conversation, regardless of topic: tone, format, standing rules like Odette's six-week lead time note. They're written once, in Settings, and get pulled into every new chat automatically. OpenAI's own documentation sets the character budget at 1,500 for free and Go accounts, 5,000 for Plus, Pro, Enterprise, Business, and Education accounts — enough for genuinely durable preferences, not enough (and not designed) to hold a full business profile or a document.

The trade-off: custom instructions are static text you maintain by hand. They don't grow from what you tell ChatGPT day to day, and they don't hold files.

Odette's actual custom instructions, once she wrote them properly: "I run a bakery and catering business. We source locally where possible and never use artificial flavoring — flag anything that implies otherwise. Wedding and event orders always need six weeks' lead time; mention this whenever an order timeline comes up. Keep vendor emails direct and warm, not corporate." That's roughly 300 characters — well inside even the free-tier limit — and it now applies automatically to every new chat, instead of being retyped.

Use ChatGPT Projects to Group Files and Context by Client or Initiative

For anything scoped to one specific piece of work — a client, a seasonal campaign, an ongoing negotiation — ChatGPT Projects group chats, uploaded files, and project-specific instructions in one place, so every chat inside that project draws on the same material without re-explaining it. File upload limits scale with plan: 5 files on Free, 25 on Plus/Go/Edu, 40 on Pro/Business/Enterprise.

For Odette, this is where a specific wedding client's file would live — the venue details, the dietary restrictions, the agreed menu draft — separate from the general bakery vendor-email work. The trade-off is the same isolation that makes it useful: one project's files don't inform a different project's chats, even for the same business.

Turn On Memory for Facts That Should Follow You Everywhere

Memory is for individual, durable facts, not documents or process detail — the kind of thing you'd tell a new hire once and expect them to just know going forward, like "we never work with a supplier that can't confirm allergen sourcing." It comes in two forms with different reliability, covered in full in `ChatGPT Memory Not Working`: explicitly saved memories persist reliably; the passive "reference chat history" layer is selective by design and can drop things you never explicitly asked it to keep. For anything that actually matters, ask ChatGPT to remember it directly rather than mentioning it in passing.

Build a Custom GPT When You Need the Same Setup Reused by Others

Custom GPTs bundle instructions and uploaded knowledge files into something reusable. They're worth building in two cases: when more than one person needs the identical setup, or when the same configuration gets reused often enough that rebuilding it in a fresh chat would waste real time.

Commonly reported practical limits sit around 20 knowledge files per GPT. OpenAI's official figures don't confirm a single hard file count — they focus on the shared storage caps that apply across chats, Projects, and GPT knowledge together: 25 GB per user, 100 GB per org, per the File Uploads FAQ.

For a one-person business like Odette's, a custom GPT is usually more setup than the problem needs. A Project plus custom instructions covers the same ground with less maintenance.

Custom GPTs earn their complexity once a team is involved. Take a 12-person bookkeeping firm building one shared GPT trained on the firm's chart-of-accounts conventions: ten staff members would otherwise each be re-explaining the same conventions in their own separate chats. One person doing the same setup alone is paying that cost with no one else to split it across.

A worked example. Sofia Reyes runs client services at that 12-person bookkeeping firm, based in Tempe. Every new bookkeeper spent their first two weeks getting corrected on the same handful of things: how the firm codes recurring subscription charges, which client accounts use cash versus accrual, the specific phrasing the firm uses in client-facing variance explanations.

Sofia built one custom GPT to fix that. She uploaded the firm's chart-of-accounts guide and a dozen example variance write-ups, then set instructions covering the recurring corrections new hires kept needing. New team members now start from that GPT instead of a blank chat — the same conventions apply consistently across ten people's work, instead of getting re-taught one correction at a time.

What the setup actually took, roughly an afternoon:

• Pulling together the firm's existing style guide

• Selecting the dozen variance write-ups that best represented the firm's voice — not the twelve most recent, but the twelve new hires had most often gotten wrong

• Writing instructions specific enough that the GPT would flag a cash-versus-accrual mismatch, not just format the answer nicely

That upfront cost is the real trade-off custom GPTs carry. A Project or custom instructions can be usefully set up in minutes. A GPT built well enough to actually correct a new hire's mistakes takes closer to the effort of writing a real onboarding document.

Six weeks in, the firm has needed exactly one correction on GPT-drafted variance language — down from a near-daily occurrence in the two weeks before any new hire got up to speed the old way. That's not because the underlying model changed. It's because the firm's actual conventions now live somewhere every new hire's first conversation can draw on, instead of living in the institutional memory of whichever senior bookkeeper happened to review their work that week.

Setting This Up in the Right Order

For most small businesses, the practical sequence is simpler than the four-feature landscape above might suggest:

First, custom instructions. Five minutes, immediate effect on every future chat. Write the two or three things that should always be true — tone, standing rules, things to never do or say.

Second, a Project for anything client- or initiative-specific that's coming up now. Don't create Projects speculatively for work that might happen someday; create one when a specific piece of ongoing work actually needs its own files and history.

Third, explicit memories for facts as they come up. Rather than trying to front-load every fact ChatGPT might ever need, save things the moment they turn out to matter — after the first time a missing fact causes a wrong answer is a reasonable trigger to go back and save it properly.

Fourth, a custom GPT only once a real, recurring, multi-person need for it appears. Building one speculatively, before there's a team actually re-explaining the same thing repeatedly, is effort spent on a problem that doesn't exist yet.

What About Claude and Gemini?

The same four-way split shows up on other platforms, with different names and different limits, so the decision logic carries over even if the settings menu looks unfamiliar.

Standing preferences — Claude: Custom instructions inside a Project · Gemini: Personal Context handles this passively — less manual control

Scoped workspace — Claude: Projects — larger effective window (~200K tokens per Anthropic, comparable to a 500-page book) · Gemini: No direct Projects equivalent

Durable individual facts — Claude: Memory — on for every user since March 2026, unified across chat and Cowork this week · Gemini: Personal Context — on by default for eligible personal accounts, unavailable on work/school accounts

Reusable, team-shared setup — Claude: No direct custom-GPT equivalent · Gemini: No direct custom-GPT equivalent

Claude also rolled out persistent memory to every user, free and paid, in March 2026, and just this week unified it across Claude's chat and Claude Cowork products, so an update from one surface reflects in the other. There's no direct equivalent of a custom GPT on Claude in the same reusable, shareable sense.

Gemini splits things differently again. Its Personal Context feature — the rough equivalent of ChatGPT's memory — is on by default for eligible personal Google accounts rather than something you opt into, and it isn't available at all on work, school, or supervised accounts.

That's worth knowing specifically if a team is planning to standardize on Gemini through a Google Workspace account: the memory behavior a founder sees testing on a personal account may not carry over once the team moves to a business account.

Common Mistakes When Giving ChatGPT Context

Putting a whole business profile into custom instructions. Custom instructions have a character limit for a reason — they're meant for standing preferences, not a company handbook. Cramming in pricing tables, full client lists, or long process documents either gets truncated or crowds out the preferences that actually need to be there every time.

Assuming an upload in one chat carries into the next. A file attached directly to a regular chat (outside a Project) is available for that conversation and generally doesn't persist into a fresh one. If a document needs to be referenced repeatedly, it belongs in a Project, not a one-off attachment.

Mentioning something important once and assuming it's "remembered." This is the single most common gap. Saying something in passing may or may not get picked up by the passive memory layer — it's designed to be selective, not exhaustive. Anything that actually matters needs an explicit "remember this," not a hopeful mention.

Building a custom GPT for a workflow only one person will ever use. The setup cost of a custom GPT is worth paying when it's amortized across a team. For solo use, it's usually more maintenance than a Project plus custom instructions, for the same result.

Why None of This Adds Up to One Business-Wide Picture

Set up correctly, these four features solve real, specific problems. What they don't do, even combined, is give a business one current, shared understanding that every conversation and every person automatically draws on — Odette's custom instructions don't know what's in her wedding client's Project, and neither knows what her memory has saved.

Sofia's custom GPT solved the new-hire onboarding problem, but it doesn't know anything happening inside any individual client's Project, and it doesn't update itself when the firm's conventions change — someone has to remember to edit the GPT's instructions by hand. Each piece of context lives exactly where you put it and nowhere else.

`Why AI Forgets Your Business` covers why that's a structural limit of chat tools, not a setup mistake, and what a system built to hold one shared picture — the way a teammate does — actually looks like instead.

Tired of deciding which ChatGPT feature should hold which fact about your business? Sign Up for Free — no credit card required, free to start, cancel anytime.