Custom instructions are persistent, user-defined directions that an AI assistant applies automatically to every new conversation, so that preferences about tone, format, background and constraints do not need to be repeated in each prompt. In prompting for artificial intelligence systems, they act as a standing layer of context placed before the individual request.
Most chat assistants expose custom instructions as a settings field or profile, sometimes split into what the model should know about the user and how it should respond. Their value lies in consistency across sessions; their main risk is that a forgotten instruction can silently distort later answers.
How custom instructions reach the model
Custom instructions work by injecting stored text into the context window at the start of each conversation, usually inside or alongside the system prompt that the application already sends. The model does not remember the user in a biological sense; it simply reads the same instructions again every time a new chat begins.
Because the text is processed as ordinary input tokens, custom instructions consume part of the context budget and are subject to the same attention dynamics as any other prompt content. Applications typically cap their length at a few thousand characters, and the platform own safety rules and system prompt take precedence over anything the user writes.
The table below shows where custom instructions sit among the layers that shape a typical assistant response.
| Layer | Who writes it | Scope | Priority |
|---|---|---|---|
| Platform system prompt | Model provider or application | All users | Highest |
| Organization instructions | Workspace administrator | All members of a workspace | High |
| Custom instructions | Individual user | All of that user conversations | Medium |
| Project or assistant instructions | User or team | One project or configured assistant | Medium, more specific |
| Conversation prompt | User, per message | Current turn | Most specific |
What custom instructions usually contain
Effective custom instructions describe stable facts and durable preferences, not one-off tasks. They usually fall into two groups: background about the user, such as role, expertise and working language, and response preferences, such as length, structure and level of formality.
| Category | Example instruction | Effect on answers |
|---|---|---|
| Role and expertise | Senior backend engineer, familiar with Go | Skips basics, uses precise terminology |
| Output format | Prefer tables and numbered steps | More structured, scannable responses |
| Length and depth | Keep answers under 200 words unless asked | Shorter default replies |
| Tone and style | Formal register, American spelling | Consistent voice across sessions |
| Constraints | Never suggest paid tools without saying so | Filters certain recommendations |
A well-written set reads like a concise briefing for a new colleague: specific, positively phrased and free of contradictions. Vague wishes such as “be helpful” add tokens without changing behavior. Concrete wording, such as naming the preferred date format or the target audience, gives the model a clear signal it can act on.
Custom instructions compared with related techniques
Custom instructions overlap with several neighboring techniques in prompting, but they differ in who controls them and how long they last. Role prompting and a prompt template, for example, are applied deliberately to a single task, while custom instructions apply by default to everything.
| Technique | Persistence | Typical author | Best use |
|---|---|---|---|
| Custom instructions | Every new conversation | End user | Durable personal preferences |
| System prompt | Every request in an application | Developer | Product behavior and guardrails |
| Prompt template | Each time it is filled in | Prompt author | Repeatable task structure |
| Role prompting | Single prompt or conversation | User | Adopting a specific perspective |
| Assistant memory | Learned facts across chats | Model and user together | Recalling details mentioned earlier |
The distinction from memory features matters in practice. Custom instructions are explicit text the user fully controls, whereas memory stores facts the assistant inferred or saved, which can be harder to audit. Context engineering treats both as inputs to be curated deliberately.
Writing and maintaining effective instructions
Good custom instructions are short, specific and testable. Each line should describe a behavior that can be observed in an answer, which allows the user to verify whether the model actually follows it.
- State defaults, not absolutes: “Prefer concise answers” leaves room for detail when a question demands it.
- Put the most important rules first, because long instruction blocks dilute attention.
- Avoid conflicts with likely task prompts, such as a strict word limit combined with frequent requests for full reports.
- Test with representative questions after every change and compare answers with and without the instructions.
- Review them periodically, since a changed role or project makes old background misleading.
In shared or enterprise deployments, organization-level instructions should cover compliance and house style, leaving personal fields for individual preferences. This separation of concerns prevents users from accidentally overriding policy text.
Limitations, risks and common mistakes
Custom instructions influence behavior but do not guarantee it. Models can drift away from a preference in long conversations, especially when later messages pull in another direction, and instruction-following quality differs between model versions.
| Symptom | Likely cause | Fix |
|---|---|---|
| Instruction ignored in long chats | Dilution by later context | Restate the rule or start a new chat |
| Odd answers on unrelated topics | Overly broad background details | Remove details irrelevant to most tasks |
| Inconsistent formatting | Contradictory format rules | Merge into one clear preference |
| Private data appearing in outputs | Sensitive facts stored in instructions | Delete personal identifiers from the profile |
A security point is often overlooked: custom instructions are not a security boundary. Text retrieved from documents or web pages can attempt prompt injection, and a user preference cannot reliably block it; protection belongs in the application design.
In simple terms
Custom instructions are like a note you hand to an assistant before every meeting. Instead of explaining each time who you are and how you like information presented, you write it once, and the AI reads that note at the start of every new chat.
The note shapes the answers but does not lock them. Keep it short, accurate and up to date, because the assistant will keep following old advice until you change it.
A practical example
A compliance analyst at a mid-sized bank uses an AI assistant daily to summarize regulatory updates. At first, every answer arrives as long prose with British and American spelling mixed together, and she spends several minutes reformatting each summary for internal reports.
She adds three custom instructions: her role as a compliance analyst in banking, a request for a five-bullet summary followed by a table of deadlines, and a rule to flag any uncertainty explicitly. Over the next weeks, summaries arrive in a consistent, report-ready format, and her editing time per document drops from roughly ten minutes to two. When she later moves to a risk modeling team, she rewrites the role line so that answers stop emphasizing regulatory deadlines.