Core files
soul.md · memory.md · posture.md
Three small files define how the agent sounds, what it carries forward, and what it does when the work is ambiguous.
Open-source kit · Agent systems
Give any agent a soul file, a memory file, and a better way to work.
Agent Soul Kit is a portable repo pattern for shaping tone, recall, and judgment without pretending the model has magic powers. The goal is practical: make agents feel more human, act more useful, and stay inspectable.
Boundaries
- Special backend needed
- 0
- Portable text files
- 3
- Runtime assumption
- Any agent
Repository design
A repo that treats character as infrastructure
This is not a prompt scrapbook. The repo is a practical kit: portable files, small adapters, examples, and a playground that makes the behavioral contract visible.
Soul
souls/
Voice, values, pacing, and temperament. This is where the agent becomes recognizable instead of generic.
Memory
memories/
Long-lived context, preferences, and known facts that should survive across sessions and tasks.
Posture
postures/
Operating rules for initiative, escalation, uncertainty, and how the agent should make progress under constraint.
Adapters
adapters/
Thin templates that translate the same files into OpenAI, Claude Code, local CLI, or generic chat wrappers.
Examples
examples/
Reference agents for coding, research, founder support, and personal operations so the repo feels immediately usable.
Playground
playground/
A no-magic browser surface where people can tune character sliders, edit files, and inspect the compiled agent packet.
Design principles
Portable, inspectable, and honest about limits
- No hidden backend state is required for the concept to make sense.
- Character stays in human-readable files that can be versioned, reviewed, and diffed.
- Adapters are thin on purpose so the same soul and memory can travel across runtimes.
- The repo pushes agents toward better taste and initiative without granting fake capabilities.
- The playground makes the behavior legible before someone wires it into a real product.
Proposed tree
agent-soul-kit/
README.md
souls/
pragmatic-builder.md
warm-operator.md
sharp-editor.md
memories/
founder-context.md
product-lead.md
personal-assistant.md
postures/
default.md
decisive.md
gentle.md
adapters/
generic-chat.md
openai-api.md
claude-code.md
local-cli.md
examples/
coding-agent/
research-agent/
studio-agent/
playground/
scenarios.json
evaluator.md
browser-demo.md
Playground
Tune warmth, pragmatism, and initiative in public
The point of the playground is simple: edit the ingredients, inspect the compiled packet, and see whether the agent feels grounded or fake before it ships anywhere.
Starter archetypes
Readout
Builder
The compiled agent should feel grounded, useful, and willing to take the first sensible step.
Warmth
steady
Pragmatism
high
Initiative
forward
Compiled packet
Starter system prompt
Where it fits
Three obvious starter modes
The repo becomes more persuasive if it ships with examples that solve real jobs rather than abstract prompt theater.
Coding agent
Sharper execution
Use posture files to favor concrete edits, explicit assumptions, and honest statements about what has or has not been verified.
Research agent
More signal, less drift
Use soul and memory files to keep the voice human while preserving durable context about the user’s domain and preferences.
Personal operator
Useful, not overbearing
Use initiative rules to decide when the agent should move on its own, when it should ask, and when it should stay quiet.
Next move
This is ready to become a real public repo
The repo can ship as an open-source playground with starter files, adapters, sample evaluations, and no proprietary backend assumptions. The local scaffold is already packaged under agent-soul-kit/, which makes the split-out repo tangible and forkable.