Design Dash

Take a fuzzy problem all the way to a complete, ready-to-build product plan — with research, cross-functional input, and ethics/equity checks built into the path, not bolted on after.

Design Dash is an open-source, AI-facilitated UX workflow you can clone and run today, for any product domain. You bring a problem; it walks you through nine phases (intake → plan), enforcing quality gates so what you produce is grounded in evidence and matches how your users actually think.

For designers. This is your starting point. The machine/agent entry point is AGENTS.md. Contributor info is in CONTRIBUTING.md.

The explainer site is live at https://abenjamin765.github.io/design-dash/.

Try it:

1. ChatGPT — open the Design Dash skill and describe the problem in plain language.
2. Cursor or Claude Code — clone this repo and run ./install.sh, then start with /design-dash.
3. Any other setup — follow the tool-neutral Getting Started guide.

You can run a dash in a general AI chat, Cursor, Claude Code, another file-capable agent, or a human-facilitated workshop. The portable method contract lives in method/method.yaml. A small finished example is the synthetic Express dash in examples/reading-list/.

---

When to run a dash

Run a dash when the work has real stakes or real uncertainty — a new workflow, an unclear problem, anything where assumptions could sink a build. For a one-off copy tweak or a throwaway sketch, you don't need the full machinery (that's what the Express tier is for).

You don't pick your rigor — the work does

Tier is rule-derived, not designer-chosen. The risk, reversibility, and reach of what you're building determine how many gates are mandatory. This is deliberate: it stops you from "tier-shopping" your way out of doing the evidence work.

Tier	When it applies	What's mandatory	Cross-functional voices
Express	Low risk · easily reversible · narrow reach (one role, small cohort). Never grading, payment, PII, or compliance.	Ethics/equity floor only. Skipped gates become tracked evidence debt — deferred, never deleted.	Panel may simulate all disciplines
Standard	Moderate risk · reversible with effort · a real workflow with bounded blast radius (>1 role or meaningful user count).	Evidence · Reconciliation · Selection · Ethics · Learning (Selection when concepts compete; Learning when it ships — both required by the machine contract for Standard).	Real sign-off from each Responsible discipline
High-stakes	Hard/irreversible · broad reach · or touches regulated personal data, financial data, or user safety.	All five gates. None waivable.	Real sign-off required; simulation never sufficient

Any feature touching regulated personal data (PII, financial, health, minors) forces High-stakes and adds a Privacy & Compliance gate with appropriate legal/privacy sign-off.

---

How it works — the nine phases

Phase	What you do	What comes out
P0 · Preconditions	Classify tier; set session mode; confirm research access	A dash you can actually run (dash-config.yaml)
P1 · Opportunity & Evidence	Write a falsifiable problem statement; size the opportunity; log assumptions	assumptions.md · success-metric hypotheses
P2 · Intake & Object Modeling	Dual mental models + ORCA; write object guides to library/objects/	scope.md · NOM · CTA Matrix · Object Guides
P3 · Framing lock	Scope the design surface — provisionally, tied to the assumption register	design-spec.md §1–3
P4 · Flow & Reconciliation	Derive scenario flows + page architecture; reconcile system ↔ mental model	flow.md · page list · resolved divergences
P5 · Divergence & Selection	Generate 2–3 real concepts; score on user + business value	A defended concept choice
P6 · Wireframe & Ethics	Wireframe; cover edge states; ethics, equity, a11y	wireframe.html · ethics review
P7 · Optional Build	Coded prototype when a workspace is configured; otherwise stub note	Prototype page or p7-build-note.md
P8 · Validate & Learn	Research plan → pitch site → Learning gate	pitch/index.html · research-plan.md · plan artifacts

---

The gates — questions you have to answer

Gates aren't bureaucracy. Each one forces "is this real, or is it assumed?"

• Evidence Gate (P1) — Is this problem real? What's the evidence, and how big is the opportunity? No statement passes on zero evidence unless explicitly tagged as an assumption.
• Reconciliation Gate (P4) — Where does our system model diverge from how users think — and how is each divergence resolved?
• Selection Gate (P5) — Did we weigh 2–3 genuine alternatives against both user and business value?
• Ethics Gate (P6) — Have we designed the empty / loading / error / permission-denied / at-scale states? Checked for dark patterns, equity issues, and privacy exposure?
• Learning Gate (P8) — How will we know if this worked? Is a usability test plan in place?

---

What you walk away with

Every dash produces five portable deliverables:

1. library/objects/*.md — Object guides that accumulate across dashes. One authoritative source of truth per domain object.
2. dashes/{slug}/pitch/index.html — Stakeholder pitch site (primary walk-away deliverable).
3. dashes/{slug}/requirements.md — Complete product requirements: context, goals/non-goals, users, objects, flows, states, acceptance criteria, open questions.
4. dashes/{slug}/wireframe.html — Monochrome, design-system-agnostic wireframes with annotated interaction patterns.
5. Evidence trail — scope.md, flow.md, assumptions.md, metrics.md, glossary.md, optional thin summary.html index.

---

Get started

No terminal required

Open GETTING_STARTED.md, choose the general-assistant path, and paste the generic agent prompt into your tool. Missing capabilities have explicit fallbacks, and the outputs stay in portable Markdown, YAML, and HTML.

Cursor or Claude Code

# 1. Clone the repo
git clone https://github.com/abenjamin765/design-dash.git
cd design-dash

# 2. Install skills (symlinks into ~/.cursor/skills/ and ~/.claude/skills/)
./install.sh

# 3. Open Cursor or Claude Code and start a dash
# Type: /design-dash

Options:

./install.sh --cursor       # Cursor only
./install.sh --claude       # Claude Code only
./install.sh --dry-run      # Preview without making changes
./install.sh --uninstall    # Remove all symlinks from this repo
./install.sh --update       # Idempotent re-link (add new, remove stale)

After install, a single edit to any skills/**/SKILL.md propagates to both agents via symlink — never fork content.

---

Related projects

• Many Hats — a cloneable AI product team that retains Design Dash OOUX/ORCA methods as skill references

---

For contributors & agents

• AGENTS.md — cross-agent entry point (Cursor + Claude Code parity, stage map, skill reference).
• CONTRIBUTING.md — how to add, port, or improve skills.
• method/ — tool-neutral phases, gates, outputs, and capability fallbacks.
• adapters/ — conformance rules and tool-specific starting paths.
• GETTING_STARTED.md — accessible setup for terminal and no-terminal environments.
• examples/reading-list/ — the canonical synthetic Express example. examples/README.md is the quality bar for future examples.
• landing-page/ — source and editorial illustrations for the public explainer site.
• commands/README.md — all / slash commands and what they do.
• rules/README.md — .mdc rules and when each applies.

Skills are organized by workflow stage under skills/ (0-orchestration through 7-critique-testing, plus _cross-cutting), with shared rules/, commands/, templates/, and a local object library at library/objects/.