Agentic skills, commands, and plugins for designing AI products — from interaction patterns to alignment, evaluation, agent orchestration, and prompt architecture.
44 skills and 18 commands across 6 plugins, available for both Claude Code and Gemini CLI. Same skills, same shape, both supported as first-class agents.
Agentic Experience Design (AXD) is a new discipline. It has its own vocabulary, definitions, and practices, and it is emerging in real time. Mixed-initiative flow. Harm anticipation. Handoff protocol. Error personality. These are real, useful, often-academic terms. They have not yet been collected into a form designers can reach for in real work.
This repo is the collection. Six plugins, mapped to six layers of the discipline: model interaction, alignment reasoning, system behaviour, evaluation, agent orchestration, prompt architecture. Inside each, seven skills and three commands your AI agent can load when you are designing or auditing agentic experiences. The work in the underlying ideas was done by alignment researchers and HCI scholars over the last three years; the translation into installable skills is the contribution. See REFERENCES.md for the mapping from skills back to source papers.
Pick the install command for your agent.
Add the marketplace, then install plugins from it.
claude plugin marketplace add Owl-Listener/ai-design-skills
claude plugin install model-interaction-design@ai-design-skillsTo install all six plugins at once:
for p in model-interaction-design ai-alignment-reasoning system-behavior-shaping evaluation design-agent-orchestration prompt-architecture; do
claude plugin install "$p@ai-design-skills"
doneGemini CLI installs one extension per directory and expects the manifest at the install source's root, so for this monorepo, clone and install each extension from its local path.
git clone https://github.com/Owl-Listener/ai-design-skills
cd ai-design-skills
gemini extensions install ./gemini-extension/model-interaction-designTo install all six:
for ext in gemini-extension/*/; do gemini extensions install "./$ext"; doneFor development (symlink instead of copy, so edits take effect immediately):
gemini extensions link ./gemini-extension/model-interaction-designBoth agents load skills the same way: each skill has a description field in its frontmatter, the agent matches your wording against those descriptions, and the relevant skill loads automatically. You do not pick the skill — the agent does.
If you have not used either Claude Code or Gemini CLI before, the path is short.
- Pick an agent. Both work with this skill set; both are well-documented; both are free for personal use. Set-up instructions for Claude Code and Gemini CLI.
- Open a terminal in any project folder.
- Run the install command for your agent (above).
- From then on, the agent will load the relevant skills automatically when you ask agentic-experience-design questions.
Try this as a first prompt to see the difference:
"I am designing an AI assistant for customer support. Help me write the error states for when the assistant does not understand the user's question. Walk me through the trade-offs."
You will see the agent reach for error-personality, tone-calibration, and (depending on the framing) harm-anticipation automatically. Compare the answer to the same question without the skills installed — the difference is what this repo is for. A worked example of that exact comparison is at examples/error-states-walkthrough.md.
You can install all six plugins at once, but if you want to start small, here is a guide based on what you are working on.
- Building any agentic feature: start with
model-interaction-designandprompt-architecture. These are the foundation. - Shipping a feature soon: add
evaluationandai-alignment-reasoning. You need the failure modes mapped before launch. - Working on a multi-agent system: add
design-agent-orchestration. Handoff protocols and orchestration anti-patterns are non-negotiable here. - Designing the agent's voice: add
system-behavior-shaping. Persona architecture, error personality, tone calibration. - Just curious: install all six. They are small.
Skills are domain knowledge units (nouns). They teach the agent about designing AI products — like crafting conversation patterns, specifying guardrails, or structuring system prompts.
Commands are workflows (verbs). They chain multiple skills together to accomplish a task — like designing a complete AI persona or auditing a prompt for effectiveness.
When you ask your agent a question, it matches your wording against the description field in each skill's frontmatter and loads the relevant ones automatically.
| Plugin | What it covers | Skills | Commands |
|---|---|---|---|
model-interaction-design |
How humans and agents take turns | 8 | 3 |
ai-alignment-reasoning |
Outside-of-the-model alignment work | 8 | 3 |
system-behavior-shaping |
How the agent shows up — persona, tone, error personality | 7 | 3 |
evaluation |
Failure taxonomies, output rubrics, agent-specific heuristics | 7 | 3 |
design-agent-orchestration |
Multi-agent role design, handoffs, observability | 7 | 3 |
prompt-architecture |
Chain-of-thought, constraint specification, system prompts | 7 | 3 |
Each plugin has its own README.md with a full skill-by-skill table.
The skills are translations of academic and industry work into a form coding agents can use. See REFERENCES.md for the mapping from skills to source papers — Gao et al. on mixed-initiative interaction, Schreiber et al. on persona-based interactions, and others.
Skill files are plain markdown. Editing one is a sentence and a commit. New skills welcome — see CONTRIBUTING.md for the file conventions and what makes a skill substantive.
If you build something on top of these — a workflow, a higher-level command, a derivative plugin for a specific domain — open an issue or a PR. The discipline gets better when more of us write it.
MIT — fork it, modify it, ship it. Attribution appreciated, not required.
Built by MC Dean. Read about why AXD is a discipline now in the launch essay.