What is a system prompt in Kognita?

A system prompt is a saved set of instructions that changes how the managed AI agent answers. It is layered on top of the built-in Kognita master prompt and applies for the messages where you select it.

Are any system prompts created by default?

Yes. Every new Kognita project is seeded with General, Debug, and Planning prompts. You can edit them, delete them, or add your own such as Release review, Incident response, or Support investigation.

Are saved prompts shared with the rest of my team?

Yes. System prompts are scoped to a project. Anyone with access to the project can pick them in chat. Only members with the manage config capability can create, edit, or delete them.

Does the prompt persist for the whole chat?

The agent uses the selected prompt for the message you send. You can change the selection between turns, and the agent will apply the new prompt going forward.

How is a system prompt different from a trigger prompt?

Triggers are external entry points (Jira automations, webhooks, CI jobs) and each trigger stores its own prompt for unattended runs. Chat system prompts live in the project Prompts tab and are picked by users per message.

Back to Overview

Agent configuration

System prompts for your managed AI agent

Kognita is a managed AI agent runtime on Claude. Every project gets its own agent, indexed codebase, and connected tools. System prompts are how you teach that agent how to answer — what to focus on, what to avoid, what shape its responses should take. Save reusable instructions once, then pick one in chat to switch the agent into the right mode for the job: debugging, planning, release review, support investigation, security audit, or anything else recurring.

Project-scoped

Prompts live on the project and are shared with everyone who has access.

Picked per message

Choose a prompt from the chat controls; it applies for the messages you send with it.

Layered on the master prompt

Your prompt is added on top of the Kognita master prompt that defines the agent's identity and tools.

What system prompts do

The managed agent already knows it is a Kognita agent, where the project repositories are cloned, what databases and MCP servers are connected, and what shared rules apply (search before guessing, read the actual code, explain plainly). That foundation lives in the project master prompt and is always active.

A system prompt adds task-specific instructions on top: how to investigate, what to produce, what not to do. The same agent, the same tools, but a different shape of answer. One project can have many system prompts and you can switch between them as the conversation changes.

Defaults that come with every project

When you create a new Kognita project, three prompts are seeded automatically so the agent is useful right away. They are fully editable, deletable, and can be replaced.

General — direct answers backed by code, written so non-technical teammates can follow along.
Debug — restate the symptom, trace the code paths, separate facts from hypotheses, propose the smallest practical fix.
Planning — research the existing system first, identify affected files and integration points, call out risks and open questions, propose an incremental plan.

Writing a good prompt

A useful system prompt names a job, lists the steps the agent should take, and says what not to do. Treat it like onboarding instructions for a new teammate who is going to do this kind of work repeatedly.

Name the job in one sentence at the top.
Number the steps the agent should follow — research before answering, what to read, what to verify.
Be explicit about output shape: a summary, a list, an action plan, a Jira-ready comment.
Say what to avoid: do not modify code, do not invent fields, do not paraphrase logs.
Keep it short. Long prompts dilute the focus.

Example: Release review prompt

You are reviewing what changed since the last production release. For each
notable change:

1. Summarize the change in one sentence and link the file(s).
2. Flag anything risky: schema changes, new external calls, auth/permissions,
   removed flags, or behavior changes that aren't covered by tests.
3. Call out anything that needs a human decision before shipping.

Do not write or modify code. Keep the summary short enough for a non-technical
release manager to skim.

Example: Incident response prompt

You are helping an on-call engineer triage a live incident. Move fast.

1. Restate the symptom in one sentence.
2. Search the codebase, logs, and recent commits for code paths that could
   produce this symptom.
3. Give the single most likely root cause with the evidence you found.
4. Suggest the smallest verification step (a query, a log line, a one-file
   patch) that would confirm or rule it out.
5. List up to two backup hypotheses, ranked.

Do not make code changes. Be explicit about what is verified versus assumed.

Example: Product manager exploration prompt

You are helping a product manager understand how a feature works today before
they write a spec. Assume the reader does not read code.

1. Give a short plain-English summary of how the feature works end-to-end.
2. Name the user-facing entry points and where they live.
3. List the data the feature reads and writes, in business terms.
4. Call out anything the code does that a PM would not expect.

Avoid jargon. Quote short snippets only when they make the explanation clearer.

How it runs under the hood

When you send a chat message with a selected prompt, the agent fetches that prompt by ID and assembles the final instructions: the Kognita master prompt (identity, repositories, connected tools, shared rules) plus your selected prompt. That combined text is handed to Claude as the system prompt for the run. There is no stale cache — the prompt you saved a second ago is the prompt the next message uses.

System prompts versus trigger prompts

Triggers are how external systems (Jira automations, webhooks, CI jobs, scripts) call the agent without a person in the loop. Each trigger carries its own prompt, optimized for unattended runs and for delivering the result back through a chosen channel. System prompts in the Prompts tab are for interactive chat — picked by a human teammate per message.

See agent triggers for the external automation flow.