"I figured out the best workflow for Claude Code/Codex ever" - X / Alex Finn

Why this is in the vault

Alex Finn demos a workflow where Linear is wired into Claude Code and Codex via the plugin/integration, then used as the "second brain" task queue for autonomous coding agents. The pattern is structurally identical to how RDCO already runs the autonomous COO loop: Notion task board as the external brain, Ray pulls items via /check-board, executes, marks done. Worth capturing as a parallel proof point and as a comparison datapoint for the Notion-vs-Linear-for-coding-agent question that will eventually come up when squarely / sanity-check ship more code.

Not a sponsored video (Finn states this on camera). Finn is a known vibe-coding/AI creator with a paid community ("Vibe Coding Academy") - assume mild promotional bias toward whichever tool he's demoing, but the technique itself is the substance.

Episode summary

20-minute talking-head tutorial with live Claude Code + Linear screen recording. Finn builds a "prompt library" Next.js app from scratch to demonstrate the workflow. Core claim: pre-decomposing the build into Linear issues before any code is written keeps Claude Code from drifting, lets multiple agents (Claude + Codex simultaneously) work the same backlog, and survives across devices because the queue lives in the cloud. End of clip pitches an agent.md rules file and Slack integration for team-of-agents observability.

Key arguments / segments

• ~0:00-2:00 - Hook. Free tool will 10x your vibe-coding velocity, works across devices, lets multiple agents share state. Tool is Linear (not sponsored, per his disclosure).
• ~2:00-5:00 - Setup. Connect Linear via Claude Code's customize > connect apps. Same plugin pattern works in Codex. Create a "team" in Linear for the app being built.
• ~5:00-9:00 - Decomposition step. Prompt Claude Code: "Before we begin, build out all the issues into Linear so we can stay organized, as well as the projects inside [team]." Claude generates ~90 issues with priority, acceptance criteria, dates, dependencies. Finn frames this as the un-skippable step that prevents drift and slop.
• ~9:00-13:00 - Execution loop. "What should our first few issues be?" → Claude reads Linear, picks scaffolding tasks, moves them backlog → in-progress → done autonomously. Finn watches it knock out 17-18 tasks in one go without further prompting. App self-tests via Claude Code's browser testing.
• ~13:00-16:00 - Multi-agent. Open Codex pointed at same project, tell it "check our Linear board, move on to the next task." Two agents pull from the same queue, no manual sync needed.
• ~16:00-18:00 - Branch-per-issue + PR review. Each Linear issue maps to a GitHub branch, PR auto-comments back to the Linear issue with the branch URL. Founder reviews PRs as the human gate.
• ~18:00-20:00 - agent.md / claude.md rules file. A rules file that says "read the Linear issue before editing code, don't touch unrelated files, don't refactor what you weren't asked to refactor, create a PR when done." He promises to share the file in description (Google Doc link).
• ~20:00-20:38 - Slack bonus + community plug. Slack channel subscribed to Linear + GitHub events gives the human team a passive feed of what the agents are doing. Plug for his paid Vibe Coding Academy.

Mapping against Ray Data Co

Mapping strength: STRONG.

This is the exact pattern RDCO already runs, just with different surfaces:

• External queue = Notion Task Board (not Linear). Both serve the same role: durable, cloud-hosted state the agent reads via /check-board and writes back via notion-update-page.
• Rules file = ~/CLAUDE.md + SOUL.md + skill SKILL.md files. Functionally equivalent to the agent.md Finn ships. RDCO's version is more elaborate (per-skill rules, channel routing rules, summary rules) but the same shape: persistent agent constitution.
• Multi-agent on shared queue = the multi-device Ray pattern. The Mac Mini channels agent and ad-hoc desktop Ray sessions already share Notion state. Finn's two-agent demo (Claude + Codex on one Linear) is the same architecture.
• Branch-per-issue + PR review = the PR-only workflow on RDCO repos ([[feedback_pr_only_workflow]]). Already locked in as a hard rule.

What's actually novel for RDCO from this video:

1. The pre-decomposition step is the underrated move. Asking Claude to spec out all issues before any code is written is something Ray does inconsistently. Currently most coding work goes "founder ask → Ray executes" without an explicit issue-tree in Notion. For Squarely or the larger SC build queue, doing an explicit "decompose into Notion tasks first, then execute one-by-one" pass would likely improve quality and reduce drift. Worth piloting on the next non-trivial multi-step build.
2. Slack-as-passive-observability maps to discord notification channels for RDCO. Already partially in place; could be tightened.

What does NOT transfer:

• Linear-specific. RDCO is Notion-native and will stay that way - no reason to fork the task surface just because Finn likes Linear's UI.
• The "90 tasks autonomously" framing is theatrical. Real builds will have fewer, larger issues; the value is in the decomposition discipline, not the count.

Decision implication: Add a sub-skill or convention to /build-project and /build-landing-page that mandates a Notion-task-decomposition pass before code is written for any project >3 sub-steps. Currently the convention is implicit; making it explicit would close the drift gap Finn identified.

Related

• [[2026-04-01-every-saas-dead-linear]] - prior vault take on Linear as the AI-native task surface
• [[2026-04-04-claude-code-best-practices]] - Anthropic's official Claude Code workflow guidance
• [[feedback_pr_only_workflow]] - RDCO's locked PR-only convention; matches Finn's branch-per-issue pattern
• [[feedback_queue_work_to_the_board]] - the founder rule that codifies why external task queues beat internal TaskList for autonomous loops
• [[2026-04-22-garry-tan-skillify-it-workflow]] - related "fat skills, thin harness" pattern for agent rules files