Derived view for design tooling (impeccable). Canonical sources win on conflict: product strategy/scope =
THESIS.md, design context =.impeccable.md, visual system =DESIGN.md. Update those first, then re-derive this file.
product
- Primary: AI-agent developers and platform engineers running long, unattended autonomous coding-agent sessions (overnight runs, CI agents, background fleets) against real repos and real infrastructure. They live in terminals, judge tools in 60 seconds by README/SDK/error-message quality, and are one bad run away from an agent that force-pushes over main or drops a table. Job to be done: "Let my agent act in production without it doing something expensive, irreversible, or embarrassing."
- Secondary: governance / compliance / security stakeholders who audit agent behavior, approve risky actions, and produce evidence. Job to be done: "Show me, verifiably, what the agents did, why it was allowed, and who approved anything sensitive."
Context of use is always professional, focused, and consequential — an agent is running in prod, or a live integration is being wired. No one is idly browsing.
DashClaw provides policy checks, remote approvals, and execution evidence for unattended AI agents. Supported hooks, OpenClaw, and the bounded capability-invoke path can stop a call before execution; bare SDK/API callers must honor the result. The lattice is allow < warn < allow_contained < require_approval < block. Protocol-1 clients obtain a single-use execution claim under current policy, then report the outcome. Signed receipts verify their recorded contents, while liveness probes provide time-bounded, client-reported evidence of the installed seam. Neither proves exactly-once external execution. The primary surface is the Approvals inbox, supported by Setup, Policies, and Decisions. The canonical product scope lives in THESIS.md.
Serious · Precise · Trustworthy. Voice is direct, technical, declarative — verbs like intercept, enforce, record, verify; no hype, no exclamation marks. Operational surfaces target quiet confidence ("things are under control" — a calm instrument panel, not an alarm board). Marketing surfaces target confident competence: the adults in the AI-safety room. Reference register: Vercel, Linear, Resend, Supabase, Datadog, Grafana.
- Generic SaaS dashboard — flat blue accents, Bootstrap-era cards, "empower your team" copy.
- Consumer AI / playful LLM wrappers — gradients, emoji, chat bubbles, pastel sparkle.
- Heavy enterprise compliance UI — dense gray-on-gray tables, mid-2000s corporate energy.
- Crypto / web3 dark theme — neon gradients, glassmorphism overload, glow spam (the most tempting failure mode given the dark + orange palette).
If a surface could be mistaken for any of the four, it's wrong.
- Evidence over decoration. Every element on operational surfaces earns its pixels by communicating state, causality, or enabling action.
- Signal, not noise, with brand orange.
#f97316is reserved for active state, required attention, primary action, brand identity — never ambient wallpaper. - Calm under pressure. Prefer typographic hierarchy, whitespace, and semantic color over motion and alarm patterns; reserve motion for truly live events.
- Token-first, never hardcoded. All color/spacing/typography flows through
app/globals.csstokens and the Tailwind theme; ad-hoc hex in JSX is a regression. - Developer-reader first. Declarative, technical copy; keyboard-friendly, density-tolerant components. Would it help a senior platform engineer at 1am mid-incident?
WCAG 2.1 AA is the floor: 4.5:1 contrast on primary text, 3:1 on large text and meaningful UI components, full keyboard navigability, visible focus rings (brand orange or border-active), prefers-reduced-motion honored on every animation, semantic HTML and ARIA labels on data widgets. Never convey status by color alone — always pair with icon or text. Dark-mode only by explicit product decision.