Source: raw/I_ve_added_a_few_things_to_my_AI_coding_workflow.md — YouTube tutorial 9vPyxCucxqI. Creator self-identifies as Chris, “I build productivity apps.” Productivity-apps focus (daily planner “Lee” mentioned, his agent open-sourced). Runs 200/mo Cursor Ultra ($400/mo combined AI tooling). Includes a sponsored mention of Gravile (AI code review).

A solo-developer’s seven-tool update to his Claude Code workflow, anchored by automated testing (Xcode Build MCP for iOS, Claude with Chrome / Cursor’s browser for web) and aggressive MCP+CLI usage for production debugging. Adjacent to the operator-track AIOS framings but distinct — Chris’s stack is for shipping mobile + web product, not for setting up second brains. Includes a direct Boris Cherny tip on remote-control auto-start as a config flag.

Key Takeaways

Models + tooling

  • Claude Code: Opus 4.7, max thinking mode (highest tier). On the $200/mo Claude Max plan. No rate-limit issues reported.
  • Cursor: GPT 5.5 extra-high with 1M context window. On the $200/mo Cursor Ultra plan.
  • Split: 70% Claude Code, 30% Cursor.
  • When to switch to Cursor: very complex bugs with many edge cases. Chris’s claim: “GPT 5.5 extra high with 1M context is better than Opus 4.7 max thinking at the majority of tasks, especially complex ones.” Caveat: too expensive to use for everything (credits evaporate in days). Default to Opus + max → fall back to GPT 5.5 when Opus is struggling.
  • Also pays $100/mo Codex but prefers Cursor for now; flagged that Codex has “an extremely good Mac app” + rumored iOS app coming.

1. Automated testing — Xcode Build MCP (iOS) + Claude with Chrome (web)

Xcode Build MCP — free MCP by Sentry. Lets Claude Code do ~90% of what Xcode does: build, run simulator, take screenshots, tap on screen elements, read logs.

Workflow with Xcode Build MCP:

  1. Tell Claude Code: “make this change, then verify in the simulator using Xcode Build MCP.”
  2. Claude takes control of the simulator, taps + tests, takes screenshots, pulls logs.
  3. Chris can keep Xcode completely closed while building iOS apps.
  • 10% gap: multi-touch gestures (e.g., reorder a list) can’t be tested by the MCP. Manual workaround: do the gesture yourself → MCP reads the resulting logs.

Claude with Chrome — equivalent for web. Install: claude-chrome plugin → run claude-chrome → Claude controls a Chrome instance, takes screenshots, reads console.

Cursor’s built-in browser does the same thing for Cursor users.

The before/after:

BeforeAfter
Tell Claude: “fix bug X” → manually verify → screenshot → paste console logs → “still broken, try again” → loopTell Claude: “fix the problem, verify via Xcode Build MCP or claude-chrome, keep going until fixed” → walk away for 20-30 min → return to a fixed bug

“It makes context-switching much easier because I can let Claude Code run for a longer period of time without intervention.”

2. MCPs + CLIs everywhere for production debugging

Chris previously used “barely any MCPs — just the Supabase MCP.” Now: “I’m fully addicted.”

The production-debugging example. User emails: “Hey Chris, the app keeps crashing when I open the settings page.”

Before MCPs/CLIsAfter MCPs/CLIs
Open Sentry → look up crashesTell Claude Code: “this specific user, here’s their email, they reported a crash on the settings page. Feel free to check Sentry, Axiom, Supabase, or Firebase.”
Open Supabase → look up user dataClaude pulls Supabase user info → checks settings → pulls Sentry for crashes for this user → checks iOS version + device anomalies → if needed, pulls Axiom for backend logs
Open Axiom → look at backend logsDone in ~3 minutes
Manually correlate (~45 min)

“Almost every major developer service has some sort of MCP or CLI available, and there’s really no reason not to use it.”

Tip on MCP vs CLI choice:

“After talking to engineers who know way more than I do, in most cases the CLI is more efficient in terms of context window. It eats a lot less tokens than MCP servers. And the agents seem to know how to use the CLI a little bit better than MCPs.”

Default to CLI when both are available; MCP as fallback.

3. Gravile — AI code review for solo developers (sponsored, but used independently)

Gravile = AI code review service. Disclosure: sponsored the video. Chris’s claim: he’d use it regardless, having A/B tested it against Cursor’s Bugbot on 60 PRs over a month — Gravile noticeably better.

  • Open a PR → Gravile reviews automatically.
  • Workflow: “Claude Code, wait for Gravile to finish reviewing, then go through every comment, fix anything that’s real, keep looping until Gravile gives you a 5/5 score.” Gravile auto-grades PRs.
  • Setup: sign up → link repo → toggle on. ~$30/mo. 14-day no-credit-card trial.
  • Use case rationale: solo dev with no human reviewer + shipping ~daily PRs → AI code review is the safety net.

4. Remote control — continue Claude Code sessions on your phone

Type /remote-control in Claude Code → continue the session in the Claude mobile app under the Code tab. Syncs both ways across devices.

Boris Cherny tip (Chris talked to him directly): run /config → enable “every time you start Claude Code, automatically start a remote-control session.” Chris has this on now.

Why Claude Code’s remote control beats Cursor’s cloud agents:

  • Remote control gives the phone session access to all the local MCPs and CLIs Chris has installed on his Mac.
  • Cursor’s cloud agents only have a handful of remote MCPs available; workarounds to get local-only MCPs working are “a hassle.”

“Probably a little unhealthy, but it has been massive in terms of being able to continue sessions while I’m on the go.”

5. CMUX — terminal app replacing Cursor terminal for running Claude Code

Three reasons Chris switched from running Claude Code in Cursor’s terminal to using CMUX:

  1. Sidebar with all projects on the left — open many windows on the right, easy to juggle multiple projects.
  2. Built-in notifications — when a specific Claude Code instance needs attention, CMUX pings + highlights the relevant window. Critical when running ~10+ instances concurrently.
  3. Lightweight memory footprint — running 5+ Claude Code instances in Cursor terminal slows his M4 Max + 64GB RAM machine; CMUX handles 20+ instances “just fine.”

6. No-flicker mode (Claude Code beta render)

Beta rendering mode at time of recording. Flag-enabled. Two visible quality-of-life wins:

  • Text input bar stays pinned to the bottom while scrolling.
  • Click directly in text input field to move cursor to that position (default Claude Code requires tabbing).

Plus smoother scroll. Recommended.

7. Settings tip: max-thinking flag

  • New Claude Code sessions default to medium thinking. Chris has a flag set so every new session auto-starts with max thinking enabled.
  • Side observation on community complaints (“Claude Code doesn’t get it”): half the time, the complainer has it set to Sonnet + medium thinking. “That’s why.”

Open Questions

  • Chris’s full name + channel — first name only in transcript. References to a daily planner app “Lee” + open-sourced AI agent (URL not given) + Instagram/TikTok presence.
  • CMUX repo / publisher — described but not linked.
  • No-flicker flag exact name — referenced but the flag string isn’t in the transcript.
  • Max-thinking flag exact name — same.
  • Boris Cherny’s exact /config setting name for auto-remote-control — referenced but not transcribed verbatim.
  • Chris’s open-source AI agent repo URL — Gravile demo is “go check out the repo and look at the PR tab”; URL not in transcript.

Try It

  1. Install Xcode Build MCP (if shipping iOS) — free, from Sentry. Tell Claude: “fix this bug, verify in the simulator via Xcode Build MCP, keep going until tests pass.”
  2. Install Claude with Chrome (claude-chrome plugin) — same automation pattern for web. Or use Cursor’s built-in browser.
  3. Audit your MCPs + CLIs. For every major service you use to debug (Sentry, Supabase, Axiom, Firebase, Stripe, Vercel, Railway), check if there’s an MCP or CLI. Default to CLI; fall back to MCP.
  4. Try Gravile for AI code review if you’re a solo developer shipping fast. 14-day free trial. Alternative to Cursor’s Bugbot.
  5. Enable remote control/remote-control for a one-off, or set the Boris Cherny config flag (/config → auto-start remote) to make every session phone-accessible.
  6. Try CMUX if you’re running multiple Claude Code instances and Cursor’s terminal is hogging RAM.
  7. Flip on no-flicker mode — beta rendering with cursor-click-to-position + pinned input bar.
  8. Set max thinking as default — set the new-session flag instead of changing it manually every time.
  9. Default model rule of thumb: Opus 4.7 + max thinking for 90% of tasks, switch to GPT 5.5 extra-high (1M context) only for very complex multi-edge-case bugs. Track cost on both plans.