AI coding tools could generate UI but lacked reliable, scoped design judgment.
Naksha Studio / Open sourcePlugin architecture × design systems
A design team
inside your terminal.
I turned design practice into an open-source system that gives AI coding tools specialist judgment, task-aware context, and repeatable quality checks—not another giant prompt.
Product concept, knowledge architecture, adapters, tooling, and maintenance.
Route tasks to bounded specialists instead of loading one giant prompt.
26 roles, 63 commands, six tool adapters, hooks, and executable quality gates.
300+ GitHub stars and 34 releases as of Jul 2026.
01 / The problem
AI could generate an interface.
It still lacked design judgment.
Coding agents were becoming capable of producing entire frontends, but their design decisions were still shallow and inconsistent. A single prompt could ask for polish; it could not reliably carry interaction states, hierarchy, accessibility, content, motion, design-system logic, or the reasoning that connects them.
Naksha began with a systems question: what would it take to make design knowledge callable, scoped, and testable inside the tools where software is already being built?
26Specialist design roles
63Task-specific commands
15,261Lines of design knowledge
02 / System model
A routing layer assembles the right design context.
The core design skill behaves like a routing layer. It interprets the task, selects the relevant specialists, and loads only the references required to complete the work. A landing page may need UI, content, motion, and design-system knowledge; a compliance audit should activate a different path.
Progressive context
Route before loading
Determine the task shape before bringing specialist knowledge into context.
Specialists, not personas
Each role contributes concrete rules, patterns, failure modes, and output expectations.
Artifacts over advice
Commands are designed to produce usable code, audits, systems, and documentation.
03 / Distribution
One knowledge core.
Six tool-native adapters.
The plugin could not assume every agent platform worked like Claude Code. I kept the design practice consistent while packaging it through each environment’s native convention: a full plugin and hooks where supported, persistent rule files elsewhere.
Claude Code
Commands, agents, hooks, skills, and a live symlink for updates.
Cursor
A native .cursor/rules/naksha.mdc integration.
Windsurf
Design behavior packaged through .windsurfrules.
Gemini CLI
Persistent project guidance through GEMINI.md.
Copilot
Repository-level behavior through Copilot instructions.
Autohand
AGENTS.md plus a project-scoped design skill.
04 / Quality infrastructure
Quality gates make the knowledge system testable.
As the command surface grew, copy-pasted counts, broken metadata, weak fixtures, and regression-prone behavior became product risks. I added executable checks around the plugin rather than relying on documentation to remain correct by hand.
verify-metadata.sh # manifest and stats agree
behavioral-smoke.sh # fixtures satisfy command assertions
design-lint.js # structural design rules
guard-legacy-branding.sh # removed identity cannot return
guard-counts.sh # README and filesystem cannot drift169 cases
Prompts and assertions cover expected behavior across the design command surface.
Detect context
Framework, tokens, dependencies, and project conventions are discovered before work begins.
Guard semantics
HTML writes are checked for language, landmarks, alt text, and genuine interactive elements.
Persist memory
Constraints and discovered patterns flow back into project memory for later sessions.
05 / Open-source evolution
The architecture grew through use—not a launch-day feature list.
Naksha moved from a role library into a maintained developer tool: cross-platform packaging, structured commands, project memory, browser research, health checks, evaluation infrastructure, and a reel pipeline all arrived through successive releases.
v3.3
Naksha identity and expanded specialist model.
v4.0
Project memory, pipelines, vision, and frontier roles.
v5.0
Persistent design agent and browser-grounded project context.
v5.1.1
Design Reel workflow and an end-to-end wipe correction.
300+GitHub stars · Jul 2026
23Repository forks
34Published releases
What I learned: encoding expertise is partly a content-design problem and partly a systems-design problem. The useful unit is not “more prompt.” It is a bounded capability with an activation rule, the minimum context it needs, a concrete output contract, and a way to detect regression.
End of selected work
Build the nextgood thing together.