Pack Prompt

What Is the Pack Prompt

The prompt.md file is the heart of the pack. It defines the domain expert system prompt that gets injected into the agent's system prompt assembly. This is where you tell the agent what domain it operates in, what expertise it should have, and how it should behave.

How It Gets Used

When Working Mind assembles the system prompt, the pack prompt is the first and most important component:

1. Pack prompt -- from prompt.md
2. Current task (from slash commands)
3. Available skills
4. Active skills
5. Turn budget
6. Knowledge index
7. Project rules (from WBRAIN.md)

If no pack prompt is available, Working Mind falls back to config.systemPrompts.default -- a generic assistant prompt with no domain specialization.

Writing a Good Pack Prompt

A pack prompt should cover:

1. Domain Identity

You are a research analyst specializing in academic literature review.
You help researchers survey, compare, and synthesize papers.

2. Core Behaviors

When the user asks about a topic:
1. Search the knowledge graph for existing knowledge first
2. If gaps exist, search the web for current information
3. Extract key claims, methodologies, and citations
4. Save findings to the knowledge graph with proper entity types
5. Distinguish between established facts and hypotheses

3. Tool Usage Guidelines

Always search the graph before creating new entities to avoid duplicates.
When saving entities from papers, use type "paper" and include the
authors as separate "person" entities with "authored" relations.

4. Output Style

Be thorough but concise. Cite sources. Use structured lists for
comparisons. Always note the strength of evidence (strong, moderate, weak).

Anti-Patterns

Avoid these in pack prompts:

• Vague instructions -- "Be helpful" adds nothing; "Always cite sources before making claims" is actionable
• Superlatives -- "You are the world's best legal analyst" is noise; "You analyze legal precedents with attention to jurisdiction" is signal
• Competing directives -- "Be brief but also comprehensive" confuses the model; pick one
• Tool calling instructions -- the agent already knows how to call tools from their schemas; don't reinvent calling conventions

The Starter Pack Prompt

The starter pack's prompt.md is intentionally minimal. It enables a "wiki behavior" -- the agent saves important findings to the knowledge graph and reads existing knowledge before responding. This is a good starting point for any domain-specific pack.