System Prompts for Writers: What They Are and Why They Matter
Lee Harris·
A system prompt is not a prompt. A prompt is an instruction for a specific task. A system prompt is the standing context the model carries into every session: who you are, what you are writing for, what your voice sounds like, what the model should always do and never do.
Without a system prompt, every session starts from nothing. The model knows nothing about you, your audience, your voice, or your workflow. You either restate that context every time, which is slow, or you skip it, which produces generic output.
A system prompt solves this by establishing the persistent context once and maintaining it across sessions.
What belongs in a system prompt
The voice reference. Three to five paragraphs from your own writing that represent your style well. The model will write toward these examples. They are the most effective voice guidance available, more effective than any description.
The voice description. One page of specific notes about the voice: what it is not, the sentence length patterns, what the assumed reader knows, what register is appropriate. This supplements the examples and handles cases where the model might otherwise drift.
The standing instructions. Things the model should always do: run two editing passes rather than one, specify every factual claim in the brief rather than generating them independently, tell you when a brief is unclear rather than filling the gap with assumptions.
The standing prohibitions. Things the model should never do: add bullet-point summaries at the end of pieces, use phrases like "it's important to note" or "in conclusion," present balanced views on questions where the piece takes a position, generate statistics without sources.
The workflow context. What kind of content you produce, who the readers are, what the purpose of the content is. "Writing for a B2B software audience that is technically literate and skeptical of vague claims" gives the model a persistent frame for calibrating specificity and register.
What does not belong in a system prompt
Piece-specific context belongs in the brief, not the system prompt. The system prompt should contain only what is true for every session. Specific reader descriptions, arguments, and constraints for individual pieces go in the individual briefs.
Topic background does not belong in the system prompt. You are not informing the model about your industry. It already knows the industry. You are informing it about your voice and workflow.
Aspirational instructions do not belong in the system prompt. "Always produce excellent work" does not change model behavior. Instructions that are specific enough to produce observable differences in output belong. Instructions that are so general that they could apply to any task do not.
Maintaining the system prompt
The system prompt needs to be reviewed whenever your workflow or voice changes significantly. A voice reference built from work you did a year ago describes a voice that may have evolved.
A practical maintenance schedule: review the voice examples once per quarter. Replace any that no longer represent your current work. Review the standing instructions once every six months. Remove instructions that reflect problems you no longer have, add instructions for patterns that have emerged.
The system prompt is not a set-it-and-forget-it tool. Writers who build a system prompt once and never update it find that the output drifts from their voice as their actual writing evolves. The system prompt and the current work should be pointing in the same direction.
Tool differences
Not every AI tool implements system prompts the same way. Claude allows a persistent system prompt in Projects, which carries across all conversations in that project. ChatGPT has Custom Instructions, which are similar but apply globally rather than per-project.
For writers who produce different content types for different clients, per-project system prompts are more useful than global ones. The voice for a technical SaaS client differs from the voice for a lifestyle brand. A system prompt that handles one well handles the other poorly.
If the tool you use does not support persistent system prompts, paste the system prompt content at the start of each session. It is slower, but the content of the system prompt still works. The persistence is the convenience, not the functionality.
More from this topic
Context Is the Prompt: How to Give AI What It Actually Needs
Generic inputs produce generic outputs. What context means in practice: the audience, the argument, the angle, the constraints, the register. How to supply it efficiently.
Building a Personal Style Reference for Your AI Prompts
A style reference is a document that tells the model what you sound like. How to build one from your own writing, what to include, and how to apply it so outputs need less correction.
How to Prompt for Tone When You Have No Examples to Give
Most writers don't have a clean sample document to hand to a model. How to describe tone in terms AI can use: sentence patterns, register, what the piece should not sound like.
Prompt Iteration: How to Refine Until You Have Something Usable
Getting a bad first output is not a failed prompt. What iteration looks like in practice: which elements to adjust, in which order, and when to give up and change your approach.