Humane Agentic Design

21 design-method skills for people working with agents. Frame a real job, make a testable interface, and keep the evidence behind the next decision.

humane.salient.community — method and skill guides.

Humane packages discovery, prototyping, design constraints, and review as skills an agent can use when the task needs them. It includes executable tools for tokens, contrast, typography, interaction maps, and story maps. It is not a UI framework or an automatic workflow engine. The person supplies context and judgment; the agent produces artifacts and records what it actually checked.

Start with the task

For a new idea, capture the job with jtbd, name an open question, and use prototype to answer it at a useful fidelity. Reuse the existing design system, or establish one with design-tokens; use design-frameworks when fitting the result to an existing component system. Apply layout-rules and ux-writing, then polish settled English or Russian text with typography.

Add semantic-micro-interactions for bounded state feedback, narrative-scrollytelling for an ordered scroll story, and sound-effects only when optional audio serves a clear purpose. walkthrough attempts the task; review consolidates domain findings and coverage into one verdict.

This is a suggested path, not a requirement to run all 21 skills. Start directly with a review or another bounded task when context already exists. using-humane reads conditional handoffs from skill metadata. Those routes do not automatically execute the next skill.

before-after expresses a transformation hypothesis or messaging grid. It is not proof of observed improvement. Synthetic personas and respondents can expose questions to investigate; they do not replace research with real people.

The 21 skills

Skill	Owns
using-humane	Task routing and the declared handoff graph
setup	Configuration, dependency checks, artifact paths, and copy-drift diagnosis
jtbd	Jobs, switching forces, outcomes, and the evidence corpus
prototype	A cheap artifact answering an open design question
persona-review	Synthetic stakeholder objections to a document
respondent-panel	Isolated synthetic reactions to a user-facing artifact
design-tokens	Token validation, aliases, CSS/brand exports, and measured contrast
type-specimen	Typeface comparison using actual product copy
layout-rules	Structural and interaction defects in tool/dashboard interfaces
ux-writing	The wording of interface strings and documentation
typography	English/Russian text mechanics, with editorial changes kept separate
semantic-micro-interactions	Meaningful feedback for bounded state transitions
narrative-scrollytelling	Story order, scroll mapping, and readable linear fallbacks
sound-effects	Optional sound palettes, provider adapters, and generation receipts
nielsen-heuristics	Evidence-located usability inspection
walkthrough	Attempting a task, including unhappy paths and capability limits
review	Domain orchestration, coverage, deduplication, and one verdict
before-after	Felt transformation hypotheses and messaging grids
brandkit	Exploring identity directions before the token contract is settled
brand-illustrate	Image batches under an existing token/brand contract
design-frameworks	Native framework discovery, fit plans, and separate compliance checks

Install and host support

The repository contains the consolidated v0.23.0 package. A local checkout does not update installed copies or publish a release.

Claude Code

/plugin marketplace add glebis/humane-agentic-design
/plugin install humane@humane-agentic-design

Invoke installed skills as humane:<skill>. These commands use the published repository, so unpublished local additions are not included until released.

Codex and other Agent Skills hosts

npx skills add glebis/humane-agentic-design

Choose the scope and installation method deliberately. To work with this local version, point the host at humane/skills/ or install from the local checkout using that host's supported mechanism. Keep required scripts and references with each skill; some workflows also read sibling skills in the complete pack. In Codex, use the skill selector or $<skill>; other hosts determine invocation.

ChatGPT

OpenAI documents a skills-only plugin conversion path. humane/ is the candidate package root. This repository does not establish a published ChatGPT listing or a verified ChatGPT installation. Conversion and fresh-environment testing remain separate release steps.

Instructions are portable; execution capabilities differ. Python helpers use the standard library. Browser checks require a browser tool. Image generation requires an optional backend; the initial sound provider requires ElevenLabs access and a process-injected key. Typeface previews may load Google Fonts. Isolated synthetic panels require separate contexts. Missing capabilities must be reported rather than silently treated as successful checks.

Evidence and maintenance

The evaluation harness tests limited, machine-checkable parts of the method. It does not establish whole-method superiority or universal host compatibility. A review verdict applies only to the inspected scope.

The packaged agent guide is the source of truth for ownership, authoring, and validation. Root AGENTS.md and CLAUDE.md both load it. The changelog records capabilities and package versions.

The method is also taught at the Agentic Design Lab. See that site for current cohort details.

MIT © Gleb Kalinin