Set your professional context
Describe your role, industry, and expertise level once. Every response is calibrated to that context without you having to ask — no more explaining what an M&A adviser does to a general-purpose model.
How it works — Custom Instructions
Tell NuBench AI about yourself once. Have it apply to every conversation automatically.
Custom Instructions give you a persistent, personal system prompt that NuBench AI reads before every conversation. Describe your role, your context, and how you want responses structured — and never repeat yourself again.
The repetition problem
Without custom instructions, every new conversation starts from zero. A financial analyst who always wants concise bullet-point summaries has to say so every time. A lawyer who needs responses to assume familiarity with legal structures restates that assumption in every new chat.
The context that makes an AI response useful — your role, your audience, your preferred format — is invisible to the model unless you provide it. And providing it manually on every turn adds friction that compounds across a team.
Custom Instructions remove that friction. Write your context once. The pipeline injects it on every turn, invisible to you, without taking up your conversation window.
How it works
- 01
Write once
Open Settings → Custom Instructions and fill in two fields: what you want NuBench AI to know about you, and how you want it to respond. Each field accepts up to 1,500 characters. Save once and your instructions are stored against your account.
- 02
Applied at the gateway
When you start a new conversation, the NuBench gateway pipeline runs step 03a before any project instructions or model routing. Your instructions are read from the database (with a 60-second Redis cache for speed) and prepended as a system message ahead of every other context the pipeline adds.
- 03
Consistent across every model
Custom instructions apply regardless of which model you use — Claude, GPT, or any Bedrock-hosted model. Because they are injected as a system message before provider dispatch, the instruction is translated into whatever format each provider expects. Mid-conversation model switching preserves the instruction.
Technical specification
- Storage
- gateway.user_custom_instructions table (Aurora PostgreSQL), scoped per org_id + user_id. Two text fields: about_user and response_style, each limited to 1,500 characters. A per-row enabled flag lets users pause instructions without deleting them.
- Pipeline injection
- Step 03a (user instructions) runs after rate-limiting and share-capability denial, and before step 03b (project standing instructions). This ordering ensures user instructions are always the outermost system context — project instructions layer on top, and provider routing and budget checks happen after both.
- Caching
- Instruction rows are cached in Redis (gateway DB 4) for 60 seconds per user, keyed as user_instr:{orgId}:{userId}. Cache is invalidated immediately on any PUT or DELETE. A Redis failure falls through gracefully to a direct database read.
- Classification
- Custom instructions are stored as plain text and have no classification of their own. They are injected before the classification middleware runs, so they are subject to the same DLP and output guardrail pipeline as the rest of the conversation.
- Erasure
- The user_custom_instructions row is deleted as part of Phase 2 (physical deletion) of the NuBench two-phase erasure workflow. There is no PII in the row beyond the instruction text itself — the user controls the content.
What your team can do with it
Enforce response format
If you always want concise bullet points, executive summaries at the top, or responses that avoid jargon — say it once in your style preferences. Every response follows the format.
Pause without deleting
Toggle custom instructions off when you want the model to respond without constraints — for creative tasks, brainstorming, or client demos. Re-enable with one click and your instructions return intact.