# Claude Code New Features, Explained

Greg Isenberg's solo setup for turning Claude Code (Desktop) into a 24/7 "AI employee": give it the same things a new hire gets (workspace, memory, brief, ticket, eyes, review, schedule, permissions), then skills/connectors/hooks — demoed on a missed-lead responder for med spas.

Tags: youtube-inbox, source, claude-code, greg-isenberg, agents, workforce, plan-mode, routines, anthropic

## Summary

The pitch is that most people use Claude Code as a one-off chat, while an "AI employee" is an operating loop: a self-explaining repo, plan mode before edits, one reviewable ticket at a time, desktop-preview "eyes," layered review against `review.md`, scheduled routines (morning brief / weekly ops / PR review), and explicit safe vs ask-first vs human-owned permissions. He builds it live around an IdeaBrowser missed-lead responder for med spas, then offers a seven-day (or seven-hour) plan to copy.

## Key points

- **AI-employee map**: Workspace (repo) + memory (project context) + brief (plan mode) + ticket (definition of done) + eyes (desktop preview / click-through) + review (diff vs standards) + schedule (Claude **routines**) + permissions. Once set, Claude Code is meant to feel like an operating layer, not a chat box.
- **Prefer Desktop over terminal**: He runs Claude Code inside the Claude Desktop app (less overwhelming than CLI) and selects a project folder as the workspace.
- **Repo brain**: Suggested layout — `/app` (product), `/context` (business brain), `/customers` (calls, support notes, objections, customer language), `/spec`, `/demos` (Loom scripts, screenshots, flows), `/routines` (recurring prompts). Three root manuals: `CLAUDE.md` (how to work), `roadmap.md` (what matters now), `review.md` (how to judge work before it ships). Takeaway: Claude is much more useful when the project explains itself.
- **Demo idea**: IdeaBrowser "missed-lead responder" for med spas — form/DM/after-hours pricing inquiries that get a late reply. Prompt: "Help me set up this repo as an AI employee workspace" plus buyer, pain, and current goal.
- **Plan mode first**: For meaningful product work, inspect the app + `CLAUDE.md` + `roadmap.md` + `review.md`, then propose files to change, smallest clean implementation, UX, risks, verification, and what V1 leaves out — **wait for approval before editing**. He treats this as measure-twice-cut-once. Planning can use a cheaper/faster model (he mentions Fable over Opus). Example ticket: waitlist form on the landing page (name/email/company + success state); he can then constrain ("frontend only," "don't touch auth/payments/DB yet," or "connect to Supabase").
- **Tickets / done**: One task, one finish line, one reviewable change. Good: add waitlist form; create pricing page from the existing design system; fix onboarding redirect after email verification; turn five customer objections into a landing-page section. Bad: "make the app better," "make this more viral," "add AI," "build the whole thing." Vague tickets make you clean up guesses. (He notes frontier models are better at guessing, but still wants explicit done.)
- **Eyes**: After a build, Claude should open the app in **desktop preview**, click the flow, check console/errors/mobile, and report what a buyer actually experiences — not only edit files. Build → run → use → test → improve.
- **Review / diff**: Desktop diff view shows before/after; click files, leave comments, ask for revisions. After editing: run checks, open preview, summarize what changed / what was tested / what still needs a human. Small tickets keep diffs trustworthy.
- **Routines ("night shift")**: Recurring work that keeps the company queryable while you sleep.
  - Morning brief: read the business + current work; sharper starting point for the day; **do not** change the product or touch production.
  - Weekly ops (e.g. Friday 3pm): review open issues + recent customer notes, group/duplicates, suggest the single highest-leverage fix, post to `/context/weekly-ops.md`, do not edit code.
  - On PR open: review against `review.md`; comment only on bugs, broken flows, security, or confusing behavior; short summary of what looks good / needs attention / ready for human review.
- **Parallel agents + worktree isolation**: Desktop Code tab can run separate sessions with isolated changes. Example morning trio: (1) onboarding-redirect bug, (2) landing-page hero clarity in five seconds for a med-spa owner, (3) demo script from customer notes. Each session gets the same project context and a specific handoff (root cause + files + checks; before/after hero + customer language; pain/product-moment/payoff script). Goal is inspectable packets, not one giant pile of AI work to untangle.
- **Permissions as delegation**:
  - **Safe**: read files, inspect repo, propose plans, run local tests, edit a small feature branch, update docs, create a draft PR.
  - **Ask first**: install dependencies, change migrations, touch auth, change payment logic, delete files.
  - **Human-owned**: production deploys, customer-data decisions, billing, security-sensitive changes.
  - Start conservative; use plan mode for bigger changes; avoid YOLO. Speed up only as the repo brain, review checklist, and ticket scope get stronger.
- **Skills / connectors / hooks** (company-specific "powerups"):
  - **Skills** = repeatable prompts (landing-page teardown as a med-spa owner: 5-second clarity, vague copy, trust signals, CTA; customer-notes skill for exact language/objections/buying triggers; demo-script skill).
  - **Connectors** = context plugins (GitHub, Linear, Google Drive, Slack).
  - **Hooks** = guardrails (format after edits; tests before a PR summary; checks before a change ships).
  - Combined with roadmap/review/customers/routines, this is the claimed moat: bespoke outputs, not a generic coding chatbot.
- **Seven-day plan** (also 7 hours / 70 minutes / 31 days depending on time):
  1. Create the repo brain (`CLAUDE.md`, `roadmap.md`, `review.md`, `/context`, `/customers`).
  2. Plan mode on one small product task (plan, file list, risks, verification).
  3. Build one visible improvement (waitlist, pricing page, demo flow, or onboarding bug).
  4. Preview loop (desktop preview, click-through, mobile, clarity).
  5. Review: diff view + `review.md`.
  6. Send the Loom/demo/landing page to ~10 people; dump replies into `/customers`.
  7. First routine (morning brief → one recommended build task).
- **Closed loop**: customer feedback → `/customers`; direction → `roadmap.md`; working style → `CLAUDE.md`; quality → `review.md`; small tasks through plan mode; changes through preview + review; recurring work as routines.
- **Sponsorship**: Anthropic / Claude sponsored the episode. Description has Claude Code setup links and IdeaBrowser.

## Quotes / memorable lines

- "I think there's a lot of people using Claude Code wrong."
- "If you want Claude Code to act more like an employee, you need to give it the same basic things you would give a person joining your company."
- "Claude gets just way more useful when the project explains itself."
- "A ticket is a small, clear assignment with a visible finish line."
- "Once [Claude] starts guessing, you're no longer managing the work. You're cleaning up the work."
- "Give Claude one clear ticket at a time… one task, one finish line, and one reviewable change."
- "That's this whole concept of like the night shift… the work is going to continually get organized."
- "You don't want a giant pile of AI work at the end of the day that you have to untangle."
- "There's some things that Claude can do freely. There's some things that Claude should ask about. Some things they got to stay with the human."
- "If you find yourself typing the same prompt over and over again, chances are that should just be a skill."
- "Once you start building like this, you're not going to build like the old way."

## Sources

- YouTube: https://www.youtube.com/watch?v=SkY-tR9kf-k
- Raw ingest: `raw/youtube-inbox/SkY-tR9kf-k/` (metadata.json, transcript.txt, notes.md, source.url)
- Channel: Greg Isenberg https://www.youtube.com/channel/UCPjNBjflYl0-HQtUvOx0Ibw
- IdeaBrowser (demo idea source): https://www.ideabrowser.com
- Transcript: Supadata v1 mode=native text=true, lang=en, 44339 chars. ASR spellings are not authoritative.

## Related pages

- [[youtube-ai-agent-workforce]] (same host; Allie K. Miller's 34-agent org vs this single Claude Code employee loop + routines)
- [[youtube-agentos-claude-sdk]] (custom Agent OS / throwaway containers / inbox vs Claude Desktop Code tab + worktrees)
- [[youtube-hermes-kanban-multi-agent]] (durable multi-agent bus + human gate vs plan-mode tickets + PR routines)
- [[grok-bot]] (managed shared computer vs a self-explaining local/repo Claude Code employee)
- [[hermes-agent]] (self-hosted agent OS vs Anthropic Claude Code Desktop)