Give ChatGPT and Claude Persistent Context with 3 Markdown Files

A developer outlined a lightweight method for giving ChatGPT and Claude persistent project context by maintaining three small Markdown files — product.md, voice.md, and rules.md — that are pasted or attached at the start of each session. The approach uses headings and bullet lists that models parse reliably, and includes an accuracy rule that has the model emit a [NUMBER] placeholder rather than invent statistics.

If you use ChatGPT or Claude for real work, you've probably typed some version of this dozens of times: "I'm building a tool for freelancers that... our tone is friendly but not cheesy... please don't use buzzwords..." Every new chat starts from zero. The model doesn't know your product, your audience, or the things you hate seeing in its output. So you either re-explain everything slow or skip it and get generic results worse . The fix is boring and effective: keep your context in three small Markdown files and paste or attach them at the start of a session. product.md — what you're building, for whom, and why voice.md — how you sound rules.md — hard constraints and output format Here's how to write each one, with templates you can copy right now. Built-in memory and custom instructions are useful, but they have limits: Markdown specifically works well because models handle headings and bullet lists reliably, and the files stay readable for you too. This is the single source of truth about what you sell. Keep it short — one screen is plenty. If it's too long, the important parts get diluted. Product Context One-liner InvoiceSnap turns a finished project into a sendable invoice in under a minute. Who it's for - Solo freelancers designers, developers, copywriters - Billing 2–10 clients per month - Currently using spreadsheets or Word templates Problem Invoicing is tedious, gets postponed, and delays getting paid. What it does - Creates invoices from a short project summary - Remembers client details and payment terms - Exports PDF and sends by email What it does NOT do - No accounting, tax filing, or payroll - No team features Pricing example Single plan, monthly. Current focus Getting the first 20 paying users. Priority: landing page copy and cold outreach. Two sections punch above their weight here: Without guidance, models default to a polished, slightly overexcited corporate voice. voice.md pulls them toward how you actually write. Voice & Tone In three words Direct. Friendly. Practical. Do - Write like you're explaining to a smart friend - Short sentences. Short paragraphs. - Use concrete examples over abstract claims - Address the reader as "you" Don't - No hype words: "revolutionary", "game-changing", "unlock", "supercharge" - No exclamation marks in headlines - No emojis unless explicitly asked - Don't open with "In today's fast-paced world..." Sample of my writing Invoicing shouldn't take longer than the last task on the project. Paste what you did, check the numbers, hit send. That's it. The sample paragraph is the most important part. Models imitate examples far better than they follow adjectives. Paste two or three sentences you've written and liked — an email, a tweet, a landing page line. This is where you put non-negotiables: facts the model must not invent, and how you want output formatted. Rules Accuracy - Never invent statistics, testimonials, customer counts, or partnerships. - If a number is needed and I haven't given one, write NUMBER as a placeholder. - Mark any sample figures clearly as " example ". - If you're unsure about a fact, say so instead of guessing. Output - Default language: US English - Use Markdown headings and bullet lists - Give 3 variations when I ask for copy headlines, subject lines, CTAs - Keep emails under 120 words unless told otherwise Process - If my request is ambiguous, ask at most 2 clarifying questions, then proceed. - When I say "tighten", cut length by ~30% without losing meaning. The accuracy rules alone are worth the effort. The placeholder convention NUMBER means you can scan output and immediately see what needs real data, instead of discovering a made-up "93% of freelancers" in something you already sent. 1. Create the files once. Put them in a folder like ai-context/ . Fill in the templates above with your own details. Resist the urge to write an essay; aim for under 300 words per file. 2. Start each session with a context message. In ChatGPT or Claude, attach the three files or paste them with a short opener: Here is context about my product, my voice, and my rules. Read all three files. Don't reply with a summary — just say "Ready." Then wait for my first task. The "just say Ready" line saves you from a long recap you'll never read. 3. Give a specific task. Write a cold outreach email to a freelance designer who currently invoices with a Google Docs template. Follow voice.md and rules.md. 4. Iterate, then feed learnings back. If you correct the model the same way twice "stop saying 'streamline'" , that correction belongs in voice.md or rules.md . Your files get better every week. 5. Use project features if you have them. ChatGPT Projects and Claude Projects let you attach files once so every chat in that project starts with them. The files stay the same — only the delivery gets easier. Persistent context isn't a feature you need to wait for. Three Markdown files — what you sell, how you sound, and what the model must never do — make ChatGPT and Claude noticeably more useful from the first message. Copy the templates above, spend five minutes filling them in, and use them on your next task. If you'd rather start from ready-made templates, PromptWorks has an AI Starter for Solo Founders https://promptworks-storefront.vercel.app pack: 40 prompts with usage context, a short guide, and checklists, as Markdown files.