Considered

An Agent Skill for AI coding agents

Your agent can render. Considered makes it design.

Considered is a portable design-reasoning skill: it makes an agent frame the human decision, commit to a structure, and hand frozen evidence to an independent reviewer — before the surface ships.

npx considered install --dest .claude/skills/considered
Read the docs →

How it works

Six stages. Every one leaves an artifact.

Decision

Every surface exists so a person can decide or act. Considered starts by naming that decision — not the layout, not the components.

FRAME.md
  • Human goal
  • Observable finish
  • Current workaround or alternative artifact considered

Questions

Thin briefs fail an eight-dimension gate. The skill asks ranked questions in small batches until the brief is designable — or records explicit assumptions with validation paths.

FRAME.md
  • Brief score /16
  • Question inventory
  • Assumptions

Zones

The surface is divided into zones with jobs, before any component is chosen. A rolled "hand" constrains the visual direction so defaults can't take over.

STRUCTURE.md
  • Zone map
  • Assigned hand
  • Order rationale

Hierarchy

One focal point. Type, weight, and position are rationed: what earns the largest treatment is argued, not accidental.

STRUCTURE.md
  • Hierarchy table
  • Action table
  • Layout and state plan

Composition

Only now is the surface built, routed by mode and surface type, honoring the recorded contract.

CONTRACT.md
  • THESIS
  • ZONES
  • HIERARCHY
  • ACTIONS

Independent review

Evidence is frozen into a packet and reviewed in a fresh context that never saw the build. The gate can block the ship.

REVIEW.json
  • gate.decision
  • findings
  • fixList

Why agents need it

Structure by accident

Information architecture emerges from generation order, not from the user's decision. The most important value on screen is whatever came first in the data.

Template gravity

Left to its defaults, a model reaches for the same components every time. Competent, interchangeable, unconsidered.

Self-certified quality

The agent that built the surface declares it done. No fresh eyes, no evidence, no gate a bad build can fail.

A vocabulary, not a black box.

init

Capture reusable product context

 frame

State the decision, ranked questions, assumptions

 structure

Commit to zones, hierarchy, action scopes

 compose

Build the routed surface

 critique

Adversarial review in a fresh context

 subtract

A removal-only pass

 harden

Real states, extremes, recovery

 document

Emit the portable system record

Measured, not vibes.

24–6

Blind paired preference, Considered vs. baseline builds, eval m0-002.

Read the eval tally →

8 v 9

Unsupported claims, Considered vs baseline — honesty at parity, up from 2:1 against.

Read the honesty record →

190

Rules in the corpus, distilled from design practice and measured failures.

Browse the corpus →

Self-run blind evaluations; scenario briefs, review forms, and full records are public in the repository. github.com/mamjed/considered/evals

Get started in one command.

npx considered install --dest .claude/skills/considered
View GitHub → View npm → Read the docs →

Claude Code

Install to .claude/skills/considered — skill discovery, shell optional.

Any Agent Skills-compatible host

Install to its documented project skills directory; the host reads SKILL.md and the bundled files.

Cursor, Copilot, Codex, Aside

Install to the host's documented skills directory and use the manual fallback if skill discovery isn't available.