Back to the shelf

048

Agentic Design Engineering (ADE)

A methodology and plugin for making AI-generated interfaces feel like a place, not a product.

Building · Started March 2026 · Tool · Vibe-Coded

That's the whole thing.

The site's sound is off. The narration still plays.

This reading didn't load. The transcript is below.

Read the transcript

AI can write code all day. What it cannot do, by default, is make that code feel like anything. Interfaces come out generic. What people now call AI slop.

That bothered me. I believe design is intentional. The same code, shaped deliberately, can feel like a place.

In March twenty twenty-six, I started building something I call Agentic Design Engineering. A methodology and plugin for Claude Code, Cursor, and Codex. Four frameworks: Build, Style, Move, Write. It starts with a portrait of the product. Each framework is scored before and after.

I have applied it to my own builds. It is still being built. You are welcome to follow along.

Narrated in my own cloned voice, generated with ElevenLabs v4.

AI can write code all day. What it cannot do, by default, is make that code feel like anything. Interfaces come out competent, functional and forgettable, the thing people now call AI slop. I think design should be deliberate, and the same code can be designed to feel like a place.

Agentic Design Engineering is my answer to that. It’s a methodology and plugin I coined in March 2026, installable from GitHub for Claude Code, Cursor and Codex. ADE makes the design pass deliberate instead of accidental.

It runs in four frameworks, applied in sequence. Build sets structure and clarity first: copy, layout, emphasis, accessibility, reward. Style introduces a physical metaphor, with lighting, colour as material, typography and animation, so the interface feels like somewhere. Move gives it life: core loops, feedback, physics, opt-in sound, a hidden easter egg. Write rewrites the copy so it speaks like one person, inside the metaphor.

Before any of this, step zero builds a Product Portrait. Every framework pairs with an auditor that scores before and after, keeps screenshots and leaves an evidence trail. One audit command reviews all four frameworks together.

I apply it across my own builds. Version 0.4.0 is on GitHub, installable and still being built.

01 · The problem

Before it existed

What hurt

AI agents write working interfaces that all look alike. Nothing made the design pass a deliberate step, so the result was competent and forgettable.

Who uses it

  • Designers who build with AI agents
  • Developers who use Claude Code, Cursor, Codex or Copilot
  • Me, on every build I direct
  • 33commits, March to November 2026
  • 4frameworks, run in a fixed order
  • 8skills at version 0.4.0
  • 4agent hosts it installs into

02 · The flow

6 steps, in the order they happen

  1. 01 / 06

    Install from GitHub

    You
    Adds the marketplace and installs the plugin
    It answers
    Puts the skills and agents into Claude Code, Cursor, Codex or Copilot
  2. 02 / 06

    Check the setup

    You
    Runs the setup skill
    It answers
    Reports the rendering stack, the evidence tier and any earlier ade_docs history
  3. 03 / 06

    Set the direction

    You
    Answers the taste questions
    It answers
    Writes a direction note that every framework reads first
  4. 04 / 06

    Call a command by name

    You
    Types the plugin prefix in the agent
    It answers
    Lists build, style, move, write, compound and audit as commands
    Claude Code with the slash-command list showing ade audit, build, compound, move, style and write
  5. 05 / 06

    Run the whole pipeline

    You
    Picks the compound skill
    It answers
    Runs Understand, Build, Style, Move and Write in order, each gating the next
    Codex listing the ade plugin skills with the Compound skill selected in the prompt box
  6. 06 / 06

    Read the review

    You
    Opens the review in ade_docs
    It answers
    Returns findings, two judgements and a status, with screenshots at 1440 and 390

03 · What was special

What it stands on

  1. Taste first

    A direction note records the owner’s thought and feeling before any framework runs, and every framework reads it.

  2. Render before code

    Auditors look at the page at two widths first. With no capture the review says source-inspected only.

  3. Findings, not a score

    Version 0.4.0 dropped the 200-point score for findings and a status, because a number let a strong part hide a broken one.

A design pass the agent cannot skip or fake.

Making-Of

The design pass AI skips, written down as four frameworks and a review.

The first commit in the public repository

  1. The world before

    AI agents produced interfaces that worked and looked the same. Design came after the code, if at all.

  2. One day

    29 March 2026. The first release: a README, a Claude Code marketplace file and a Codex plugin, with four frameworks defined.

  3. Because of that

    1. 30 March 2026. Twenty changes from a critical review, and Cursor support.
    2. 31 March 2026. The frameworks were renamed from acronyms to Build, Style, Move and Write.
    3. 1 April 2026. Rendering stopped being tied to one technology, and Style could send structural changes back to Build.
    4. 5 July 2026. Sound, a fidelity ladder and an evidence trail of screenshots.
    5. 13 July 2026. The repository itself became agent-readable, with a readiness check.
    6. 12 November 2026. Taste-first direction and a review in place of a score merged as version 0.4.0.
  4. Until finally

    Version 0.4.0 is on main, installable from GitHub for four agent hosts.

  5. Since then

    Fixtures with seeded defects are next, so the auditors can be proven.

The evidence Process, hands, and artifacts

Process

  1. Origin

    Four frameworks and one plugin, in the open under MIT

  2. Names

    Plain verbs instead of acronyms

  3. Ceiling

    Any rendering stack; structure may change for a metaphor

  4. Proof

    Screenshots and a declared concept to audit against

  5. Agent-native

    Root agent context and a vocabulary file

  6. Taste

    A direction step before any framework

  7. Review

    Findings and a status replace the score

Hands

Akshansh

  • The position: design is deliberate, and taste comes first
  • The four frameworks, their order and every rename
  • Review and merge of each release

Agents

  • Claude CodeWrote and revised the skills, agents and references
  • CursorRewrote the README for human and agent readers
  • AmpCo-author on some commits