neurodiveragents
Eighteen specialist AI agents, each grounded in a real cognitive style, not a list of rules.
Designed for Claude models and OpenCode. Works with Cursor and GitHub Copilot.
npx neurodiveragents install
One command installs everything. Detects your tools automatically.
Meet the Agents
Each agent is a different way of thinking. Pick the one whose posture matches the problem.
Agents are one layer
Each agent thinks clearly within its domain. But specialists are only as effective as the context, memory, and tooling you wire around them. Without that infrastructure, the agent reasons precisely about whatever lands in the prompt — and nothing else.
Agents don't know all the details of your project's conventions, architecture decisions, or domain rules unless you surface that context explicitly. They don't carry memory across sessions unless something feeds prior state in. They don't automatically adapt to your stack — you decide what gets wired in and how.
That infrastructure is your responsibility to build. The topics worth researching: harness layer design, context engineering, AGENTS.md and CLAUDE.md setup, and project-specific skills tailored to your codebase and workflow. neurodiveragents gives you the specialists. The harness is yours to configure.
How to use
The question before every session: "how much context do I have, and is it enough to act precisely?" Describe your situation below.
Cognitive Scaffolding
Agents route tasks to the right specialist. But how any step thinks — the cognitive orientation it brings — is a separate variable. Your AI scaffold tells the model what to do. Cognitive modules tell it how to think while doing it.
A review step that says "check for issues" and one that
loads
ndv-total-perception
produce different results — not because the instructions changed,
but because the cognitive frame did. Same scaffold, different
thinking.
These raise the ceiling. They don't raise the floor.
See the difference
# your-skill / step-review.md instructions: | Review this code for issues. Report what you find.
ndv-total-perception + ndv-adversarial
# your-skill / step-review.md instructions: | Review this code for issues. Report what you find.
ndv-total-perception — registers everything simultaneously before sorting by severity. Self-censoring a finding as "too minor to mention" is a perception failure, not a judgment call. ndv-adversarial — assumes the code is lying until proven otherwise.
Compose by phase
2–3 modules per phase. More dilutes the signal — the model balances competing frames instead of fully adopting one.
Probe intent. Scope is a wall.
Follow the pattern. Build on verified ground.
Everything registers. Code is guilty until proven innocent.
Model the reader. Every word earns its place.
Install modules
npx neurodiveragents install-skills claude
npx neurodiveragents install-skills opencode --global
→ Full guide: docs/ndv-skills.md
Claude Code and OpenCode only. Cursor and GitHub Copilot do not implement the Agent Skills spec.
Install
One command auto-detects your AI coding assistants and configures all agent prompts. No global state — pure local configuration.
All Tools (Auto-detect)
npx neurodiveragents install
Or install for a specific tool
Claude Code
npx neurodiveragents install claude
OpenCode
npx neurodiveragents install opencode
Cursor
npx neurodiveragents install cursor
GitHub Copilot
npx neurodiveragents install copilot