ChatGPT Custom Instructions and Projects for Writers

Ask any regular ChatGPT user for writing and they'll tell you the same annoyance: every new chat starts from zero. You explain your tone, your audience, your formatting preferences, again. Custom instructions and Projects exist specifically to fix this, and most writers still aren't using them properly.

What custom instructions actually do

In Settings, ChatGPT lets you fill two boxes: what you want it to know about you, and how you want it to respond. This isn't a creative writing prompt — it's persistent context injected into every conversation. For writers, the highest-value use isn't "write in a friendly tone," which is too vague to change behavior meaningfully. It's specific, checkable constraints:

• "Never use em dashes; use periods or commas instead."
• "Avoid the words 'delve,' 'elevate,' 'unlock,' 'game-changing.'"
• "Default sentence length under 20 words unless I ask for a longer, exploratory paragraph."
• "I write for [specific publication/audience] — assume they already know [X], never explain [Y] from scratch."

Where Projects go further

Custom instructions are global — they apply everywhere. Projects let you scope instructions and reference material to one specific piece of work: a book manuscript, a client account, a newsletter. Inside a project you can upload past chapters, a style guide PDF, or brand voice documents, and the model treats that as grounding material rather than something it has to be told fresh each time.

This distinction matters practically. A freelance writer juggling three clients with three different voices should not try to cram all three into global custom instructions — that's a recipe for bleed-through. Three separate Projects, each with its own uploaded samples, keeps voices from contaminating each other.

Comparing the setup across tools

If you're deciding between the two platforms generally, our full ChatGPT vs Claude comparison covers more than writing workflows specifically, but for voice consistency over long projects, Claude's larger context handling is a real advantage worth testing against your own material.

A setup writers actually stick with

1. Write your global custom instructions once, focused on mechanical style rules you'd otherwise fix in editing anyway.
2. Create one Project per client, publication, or book.
3. Upload 3-5 of your best, most representative past pieces to each project — not everything you've ever written, just the ones that best match the voice you want reinforced.
4. Revisit the instructions every few months. Voice and audience shift, and stale instructions quietly degrade output without an obvious cause.

The limitation worth naming

Custom instructions bias the model's defaults; they don't guarantee compliance every single time, especially in longer conversations where earlier context can override later instructions. Don't treat this as a "set and forget" fix for voice — treat it as a strong first draft filter that still needs an editing pass. For more on picking the right writing stack overall, see best AI writing tools and the AI writing tools category.