Workflow
PRD First — Product Requirements Document as source of truth
What it is
Write a full Product Requirements Document before any code is written. Every agent in the session reads the same PRD. Scope questions don't go to you — they go to the doc.
Try this
Start your next Claude Code session with a PRD prompt before asking Claude to build anything. One doc prevents a dozen conflicting agent interpretations.
"Before I write a single line of code, I need Claude to know exactly what we're building — and so do I."
Workflow
Confirm Understanding — 95% confidence threshold before any build starts
What it is
A custom Claude skill triggered with /confirm. Claude displays a confidence score. If below 95%, it asks targeted questions. Once it hits the threshold, it proceeds. Dave built this skill live from scratch using /skill creator.
Try this
Before starting any complex build, trigger /confirm. Build nothing with a shared confidence below 95%. This replaces most "wait, did you understand me?" interruptions.
"I don't need guardrails if Claude already asked every question I would have stopped it for."
Workflow
"Please Hold" — pause without losing context
What it is
Type "please hold" and Claude acknowledges the task, holds full context, and waits. No code starts. You switch to another terminal or task, return, and resume exactly where you left off.
Try this
When you need to switch focus mid-session: type it. Don't close the tab, don't start a new conversation. Just hold.
Workflow
"Thoughts?" — separate the thinking phase from execution
What it is
Append "thoughts?" to any prompt. Claude reasons and responds but writes no code. Dave used this before every major architectural decision — stack choice, UX approach, API strategy — to stress-test the thinking before committing.
Try this
Any decision with real consequences: ask "thoughts?" first. Execution comes after the thinking pass clears. Prevents half-built implementations of ideas that weren't ready.
Workflow
"Grill Me" — stress-test your PRD before any build starts
What it is
A Claude skill (from Matt Pocock's GitHub) that interrogates your PRD for gaps, contradictions, and missing specs. Claude plays devil's advocate — pointed questions that expose what's underspecified. Answers go back into the PRD before agents start.
Try this
Feed Claude your PRD and trigger the skill. Any ambiguity that surfaces here is ambiguity that would have surface mid-build as a broken assumption.
Workflow
Decision Documentation — a separate doc for every "why" call you make
What it is
A living document maintained alongside the PRD — logs every architectural or design decision: what was chosen, what was rejected, and why. Dave kept this open throughout the 12-hour session.
Try this
Start a decision log with your PRD. Every time you make a call, log it. Six months later — or when handing off to a new agent — this prevents relitigating settled questions.
"The PRD tells you what. The decision doc tells you why. Without the why, you're going to argue with yourself in six months."
Workflow
Skill Creator — build reusable Claude skills while the use case is live
What it is
A meta-skill (/skill creator) that walks you through building a new persistent Claude skill. Dave built the Confirm Understanding skill live during the stream — naming it, defining the trigger, setting the 95% threshold, saving it for reuse.
Try this
Every time you find yourself giving Claude the same setup instructions: turn it into a skill. Skills compound — each one is a reusable workflow primitive.
Workflow
Scaffold with Placeholders — build the architecture, plug API keys at deploy
What it is
Write the full integration scaffolding for Resend, Vercel, and Supabase with placeholder values — correct structure, correct function signatures, correct env var names — with YOUR_KEY_HERE standing in for live credentials. Real keys drop in at deploy time.
Try this
Keep the build moving without stopping to provision accounts. Decouple architectural decisions from credential management.
Multi-Agent
CC Chain — one human running 12 AI sessions simultaneously
What it is
One human stays in the loop with one orchestrator session (Claude.ai). That orchestrator directs multiple Claude Code terminal sessions running in parallel — all reading the same PRD. Agents update shared task docs (open / pending / closed) so they don't overlap.
Try this
Map your agent sessions with task docs. When you scale to multiple terminals, agents can't coordinate through you — they coordinate through shared docs.
"One human in the loop. Twelve AI sessions. That's not theoretical — that's what we're doing right now."
Multi-Agent
Agent All-Hands Meeting — agents read each other's session state and self-assign handoffs
What it is
Each agent has its own session notepad — a scoped .md file. During handoff, agents exchange these docs, read each other's state, and self-assign next actions based on what's open vs. closed in the shared task tracker. No human bottleneck for inter-agent coordination.
Try this
Give each agent its own session .md. Include an open/pending/closed task list. Agents negotiate handoffs themselves.
Multi-Agent
Remote ENV Numbering — make agent sessions addressable and repeatable
What it is
Each remote Claude Code session is tagged with an ENV number. The orchestrator uses that number as a trigger/selector when activating or referencing a specific agent instance — instead of manually opening tabs or terminals.
Try this
Name and number your agent sessions before they start. Makes large multi-agent builds navigable without losing track of which terminal is doing what.
Multi-Agent
Caffeinate — one command keeps your Mac awake for the entire build
What it is
macOS caffeinate -disu prevents display sleep, idle sleep, system sleep, and disk sleep simultaneously. Dave ran this at the start of the 12-hour stream to ensure agents kept working without interruption.
Try this
Run caffeinate -disu & before any overnight or long-running agent session. Trivial to run, catastrophic to forget.
Live Build
A/B Testing Landing Pages — cookie-based variant assignment built in from day one
What it is
Three different hero angles for the Golem site. On first load, the site randomly assigns the visitor to a variant and stores the assignment as a cookie. Cross-reference with Supabase conversion events to find what converts.
Try this
It costs nothing extra to build in at project start. Don't wait for "later." Three hero angles, random assignment, cookie-stored. The conversion data is worth it.
Live Build
Homebrew for Everything — install any Mac software in one terminal line
What it is
Dave demonstrated brew install paintbrush live — no browser, no DMG, no drag-to-Applications. Every tool in the stack installs via a single command, making environment setup scriptable and repeatable.
Try this
If Homebrew isn't installed: /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
Carter's Pick
iMessage Automation via Blue Bubbles — first-touch iMessage at lead capture, no API fee
What it is
Blue Bubbles server runs on a MacBook. An old iPhone connects as the iMessage relay. Tressle IQ verifies phone numbers (landline vs. mobile, carrier, reputation) before sending. Lead form submission → Supabase → trigger → auto first-touch iMessage. No third-party API, no monthly fee.
Try this
If you have an old iPhone collecting dust: this stack costs nothing beyond it. iMessage open rates destroy email and SMS. First-touch under one minute from lead capture is a significant conversion edge.
"iMessage is the highest-response-rate channel you have. If your lead form isn't firing an iMessage in under a minute, you're leaving money on the table."
Carter's Pick
21st.dev — drop in pre-built Next.js components and skip designing from scratch
What it is
A component library Carter recommended live during the stream. Pick standard UI patterns (hero sections, cards, nav, forms), plug them directly into the Next.js project. Zero time spent designing components that already exist.
Try this
For any landing page or app using Next.js: check 21st.dev before writing any UI component from scratch. Standard patterns are already there, polished, and ready.
10 Things You Can Do Tomorrow
Pulled directly from the 12-hour session. No setup required for most of these.
1
Write your PRD before you open Claude Code. Even a rough one. Every agent you spin up reads it — one doc prevents a dozen conflicting interpretations.
2
Install the Confirm Understanding skill. Trigger it with /confirm. Build nothing with a shared confidence below 95%. This replaces most of your "wait, did you understand me?" interruptions.
3
Use "please hold" to pause without losing context. When you need to switch focus mid-session, type it. Don't close the tab.
4
Start a decision log alongside your PRD. One doc for what to build, one doc for why you made each call. Keep both open in the same project folder.
5
Run caffeinate -disu & before any overnight or long-running agent session. One command, prevents hours of wasted work.
6
Use "thoughts?" before "build it." Any significant decision — stack choice, UX, architecture — ask for thoughts first. Execution comes after the thinking pass clears.
7
Check 21st.dev for component scaffolding. Pre-built Next.js components. Don't design from scratch what already exists.
8
Build A/B variants into your landing page from day one. Three hero angles, random assignment, cookie-stored. It costs nothing extra at build time.
9
Map your agent sessions with task docs (open / pending / closed). At scale, agents coordinate through shared docs — not through you.
10
Use /skill creator to capture any repeatable workflow. Every time you give Claude the same setup instructions, turn it into a skill. Skills compound.