Reference

Considered — Claude Code design skill documentation

Considered is a portable design-reasoning skill that makes a coding agent frame the decision, structure the surface, and pass independent review before shipping UI — current version 0.6.0, installed with npx considered. Everything you need to install the skill, read its commands, and understand the modes and artifacts it produces is below. For the full argument and evidence, see the homepage.

Install

npx considered install --dest .claude/skills/considered

Point --dest at your host's actual skills directory — there's no universal default, because discovery paths vary by host and version.

Host style	Destination to verify	Capability needed
Agent Skills-compatible host	its documented project skills directory	reads `SKILL.md` and bundled files
Claude Code	`.claude/skills/considered`	skill discovery, shell optional
Cursor, Copilot, Codex, Aside	the host's documented skills directory	skill discovery; use the manual fallback if unavailable

Update: rerun the same command with --force after reviewing the release notes.
Uninstall: delete the destination directory. install never writes outside it.

Treat a suggested destination as unverified until it actually works in your installed harness — don't assume a path is supported just because it looks right.

How it works

Considered routes a design request through six phases — decision, questions, zones, hierarchy, composition, independent review — and each phase must write its artifact to disk before the next is allowed to start. A deterministic roll assigns the structural and visual direction rather than letting the agent pick its comfort zone, and the review phase hands a frozen packet to a reviewer who never saw how the surface was built.

See the full walkthrough on the homepage →

Commands

Command	Required outcome	Writes
`init`	Capture reusable product context	`.considered/PRODUCT.md`
`frame`	State the decision, object model, ranked questions, assumptions	`.considered/<surface-id>/FRAME.md`
`structure`	Record the assigned hand, zones, hierarchy, actions, contract	`.considered/<surface-id>/STRUCTURE.md`, `.considered/<surface-id>/CONTRACT.md`
`compose`	Build or specify the routed surface	the target surface itself, or a portable spec
`critique`	Run an adversarial review in a fresh context	`.considered/<surface-id>/REVIEW-PACKET.md`, `.considered/<surface-id>/REVIEW.json`
`subtract`	Make a removal-only pass	the revised surface, no new artifact
`harden`	Add real states, extremes, recovery behavior	state coverage folded into the surface
`document`	Emit the portable system record	`.considered/<surface-id>/DESIGN.md`

Modes

One surface, one mode. The mode sets density, focal position, action posture, and motion budget before any layout choice is made.

Persuade: Win belief before commitment — one claim, its proof placed beside it, a single primary action.
Operate: Put the task and its current state in direct view — direct actions, low motion, errors that recover.
Analyze: Lead with the finding that changes the decision, not with whatever chart the data happens to support.
Read: Shape a line of thought the reader can follow, scan, and return to without losing their place.
Experience: Let the work set its own pace — chrome and explanation recede so the work doesn't have to compete with them.

Artifact contract

Every durable artifact is workspace-relative and named the same way on every project:

.considered/PRODUCT.md                        # product context, written once by init
.considered/<surface-id>/FRAME.md            # decision, brief score, ranked questions
.considered/<surface-id>/STRUCTURE.md        # assigned hand, zones, hierarchy, actions
.considered/<surface-id>/CONTRACT.md         # the signed structural contract
.considered/<surface-id>/REVIEW-PACKET.md    # frozen evidence handed to the reviewer
.considered/<surface-id>/REVIEW.json         # the independent verdict the gate reads
.considered/<surface-id>/DESIGN.md           # the portable record of what shipped and why

The brief gate

Before any structure is chosen, frame scores the brief against eight dimensions — human named, decision named, consequence named, frequency named, success measurable, scope bounded, constraints known, current state known — each worth 0, 1, or 2, for a total out of 16.

Below 8 of 16, the deliverable isn't a design — it's ranked questions, asked in batches of at most three, re-scored after each round until the brief clears the threshold or the user explicitly accepts the risk of proceeding below it.