To Stay Ahead on AI, Think Like a Designer

Why this is in the vault

A practitioner account of moving from "execution work" to a "design layer" (specs, constraints, diagnostic review) as the way to stay valuable once AI absorbs the visible output — a near-exact restatement of the spec-first discipline RDCO already runs as a hard rule.

Mapping against Ray Data Co

This is the outside-world version of the brigade pattern already built into ~/.claude/skills/station-spec-author, station-test-author, station-code-author, station-critic, and the hard rule logged in feedback_plan_tests_implementation_order.md ("Plan → tests → implementation... an assertion must trace to a source predating the implementation"). Reganti's argument — that the professional's job shifts from doing the work to writing the spec, constraints, and failure modes that let a human or an AI agent do it with your judgment baked in — is the same claim underneath the brigade's four-station split: spec-author and test-author write down judgment before code-author touches the artifact, so the critic station has something independent of the implementation to check against. Her five practical patterns map cleanly onto existing RDCO practice: (1) spec-before-build is literally what station-spec-author does; (2) targeted diagnostic questions instead of full code review is close to what station-critic's per-axis fan-out is trying to approximate cheaply; (3) turning recurring feedback into reusable prompt instructions is the whole thesis of Garry Tan's "thin harness, fat skills" pattern already filed and acted on via the /skillify and /improve skills; (4) tool selection based on whether an agent can actually use the tool is a live consideration in every MCP-server integration decision RDCO makes; (5) feedback loops against AI output drift is exactly what /vault-health, audit-newsletter-outputs.py, and the verify-* critic family exist to catch mechanically rather than by vibes. The piece doesn't add a new idea to RDCO's model, but it's a clean, well-worked independent confirmation from an AI-scientist-turned-founder that the spec-first constraint isn't RDCO-specific caution — it's the visible shape of how any solo operator has to run agents now.

The core argument

Reganti (former Amazon applied-AI lead, now founder of LevelUp Labs) argues that as AI increasingly produces the "execution" output that used to prove expertise — code, drafts, designs — professional value moves up to a "design layer": encoding judgment into constraints, specs, and systems before execution starts, so both human teammates and AI agents can make the decisions the expert would have made. She traces this through her own transition from "build fast, fix later" AI-scientist work to founding a company where she had to write down sales processes, client frameworks, and a five-stage "AI readiness" assessment so her team — and eventually agents like Claude Code and Codex — could operate with her judgment already encoded.

Five practical patterns: (1) write a spec before anything gets built — she walks through a worked example for a "friend tracker" personal-CRM app (overview, hero scenario, functional requirements, behavioral rules, non-goals, failure modes); (2) ask targeted diagnostic questions of AI output instead of reviewing every line (auth handling, token expiry, payment failure paths, rate limiting); (3) turn recurring taste/feedback into reusable prompt instructions rather than re-explaining it each time; (4) pick tools that solve a real, understood problem an AI agent can actually use — not novelty; (5) build feedback loops, since AI output quality drifts as models and context change. Named references: RAG, MCPs, Codex-native apps, voice-first knowledge work, Claude Code, Codex.

⚠️ Sponsorship

Two distinct disclosures in this issue, different in kind:

• Maven (content partnership, revenue share): Every discloses it "receives a share of revenue from new Maven course enrollments made through this partnership. Maven helped connect us with instructors and suggested potential topics; Every retained full editorial control." The piece links to Reganti's own paid Maven course ("Building Agentic AI Applications with a Problem-First Approach") with a 15% discount code. This is a structural sponsor, not a clean ad block — the author has a direct financial incentive to make the case that her specific expertise (spec-writing, AI-readiness frameworks) is the valuable layer, since that's also what she's selling. The underlying argument still holds up against RDCO's own independently-arrived-at brigade pattern, but read the framing as partly a sales pitch for the course.
• Svix (paid ad block, unrelated to the essay): standard "brought to you by" placement pitching webhook infrastructure ($12,000 in free credits, $50,000 for YC companies). Clean third-party ad, no relationship to the essay's content.

Standard Every footer self-promo (Sparkle, Cora, Spiral, Monologue) also present — house cross-promotion, not disclosed as a special relationship because it isn't one.

Related

• [[2026-04-11-garry-tan-thin-harness-fat-skills]] — the "turn recurring feedback into reusable instructions" pattern Reganti describes independently, already the basis for RDCO's /skillify and /improve skills.
• [[2026-05-12-zach-lloyd-warp-verify-then-build-test-harness-agentic-coding]] — the same verify/spec-before-build discipline from an agentic-coding-tooling angle, corroborating this from a different practitioner.