Configuring ChatGPT for sales work is four moves: write Custom Instructions so every session knows your company and your rules, set the data controls before any deal information goes in, understand how memory behaves across prospects, and put your repeatable work into Projects. Budget forty-five minutes. On Plus, Pro, Business, Enterprise, and Edu you now get 5,000 characters of Custom Instructions — enough to stop re-explaining your job in every chat.
Before you start
Audience: individual seller — nothing here requires an administrator. On Business and Enterprise workspaces, data controls are owned by your admin and the defaults differ; notes below.
Plan requirements: Projects are available on every plan, free included, and you can create an unlimited number of them; file caps are per project — 5 on Free, 25 on Go and Plus, 40 on Pro, Business, Enterprise, and Edu. Custom Instructions run 1,500 characters on Free and Go, and 5,000 on Plus, Pro, Business, Enterprise, and Edu (a July 2026 change). Custom GPT creation is the paid feature.
Prerequisites: a ChatGPT account and the same ten minutes of honesty about what you sell, to whom, and in what voice.
Step 1: write Custom Instructions as your base layer
Custom Instructions live under Settings, Personalization. They are the base layer of a stack: they apply everywhere, and Projects and custom GPTs override them inside their own walls. The interface asks two kinds of question — who you are and what you do, and how ChatGPT should respond (the exact field labels shift as OpenAI iterates; the split does not). In the first kind, put your title and company, what you sell in two sentences, your ICP, differentiators, competitors, deal size and cycle, your CRM, and the tasks you run most. In the response side, put your rules:
Response rules — starting block
"Emails 100–150 words unless I specify otherwise, direct and peer-to-peer, never vendor-sounding; banned phrases: 'synergy,' 'touch base,' 'circle back,' 'I hope this finds you well.' Every email ends with one specific proposed next step. Summaries as bullets, ten maximum. Proposals follow situation, problem, solution, proof, investment. Never invent statistics, customer results, or product capabilities. When my request is ambiguous, ask one clarifying question before answering."
At 5,000 characters you no longer have to choose between company context and voice rules — write both fully. On Free and Go, spend the 1,500-character budget on voice rules and banned phrases; company context can travel per-conversation, but tone drift is what makes AI drafts unusable.
Step 2: set data controls before deal data goes in
Under Settings, Data Controls, review whether your conversations may be used for model training and set it deliberately — on individual plans this is your toggle; on Business and Enterprise, training is off by contract and your workspace admin owns the settings. Do this before pasting call notes, pricing, or anything customer-identifiable, and keep the general discipline: customer identifiers, unreleased pricing, and NDA material stay out regardless of toggles. The standing what-not-to-paste rules on the best-practices page apply here verbatim.
Step 3: understand memory before it surprises you
ChatGPT memory has two parts — saved memories it writes about you, and reference to your chat history — and both are useful and hazardous in sales work for the same reason: context from one prospect can surface while you work another. Three habits keep it clean. Review saved memories monthly under Settings, Personalization, Memory, and delete anything deal-specific that got saved as if it were permanent fact. Start a fresh conversation per prospect or per batch rather than working five accounts in one thread. And put durable facts — your ICP, your rules — in Custom Instructions, not memory: instructions you wrote are auditable; memories the model wrote are not. For genuinely confidential accounts there is a cleaner option than turning memory off entirely: create the project with project-only memory, which keeps its chats referencing only each other and leaves your saved memories out — the tightest per-prospect isolation ChatGPT offers. Set it at creation; existing projects cannot be converted.
Step 4: put repeatable work into Projects
Projects group conversations with their own files and instructions that override your base layer inside the project. Build one per recurring workflow: outreach (voice rules and one great example email), discovery and qualification (your framework and question bank), proposals (template and two past winners), account research. Upload five current documents per project rather than twenty stale ones (Free's five-file cap makes the choice for you) — a product overview, ICP definition, case studies, pricing, one strong proposal. Projects are available to everyone; custom GPT creation is where a paid plan comes in. If you want a shareable assistant with a fixed document set, a custom GPT is the packaged version of the same idea — build the Project first, and promote it to a GPT when a colleague asks for it.
Troubleshooting
If output ignores your instructions, check which layer you are in — a Project's instructions override your base, and an old Project with stale rules will faithfully apply them. If the voice drifts mid-thread, the thread is too long; start fresh, your instruction layers persist. If memory keeps resurfacing a dead deal, delete the specific saved memory rather than turning the feature off. If Custom Instructions reject your text on Free or Go, you are over 1,500 characters — cut context, keep rules.