Austin Marchese
Austin MarcheseApr 1
Tech

How Claude Code’s Creator Starts EVERY Project

12 min video5 key momentsWatch original
TL;DR

Boris Churnney, creator of Claude Code, starts every project in plan mode first, keeps claude.md minimal, builds in verification loops, runs parallel sessions, systematizes repetitive tasks with Claude skills, and optimizes for future model improvements rather than prompt tweaking.

Key Insights

1

plan mode firstBoris starts 80% of his sessions in plan mode, doing the thinking upfront so execution becomes almost automatic. Most people skip this and end up debugging mistakes that proper planning would've prevented.

2

delete and restartClaude.md should stay short—maybe a couple thousand tokens. When it gets bloated, Boris deletes it entirely and starts fresh because models improve daily and old rules become unnecessary baggage.

3

verification feedback loopsGiving Claude a way to verify its own work—like opening a browser to test a website or running a workflow to check output—can 2-3x the quality of results. It's a feedback loop that makes the model catch its own mistakes.

4

parallel sessionsRunning multiple Claude sessions in parallel on different tasks is a massive efficiency gain. Each fresh context window can see problems the first session missed because it's too deep in the weeds.

5

systemize inner loopsClaude skills let you document a process once and call it infinitely. Instead of re-explaining the same task repeatedly, you create a reusable play that runs perfectly every time with different data.

6

build for future modelsDon't over-optimize prompts. Models get better every month, so tweaks you make today will be pointless soon. Focus energy on your information architecture—what you feed the model—not micro-tuning language.

Deep Dive

Plan Mode Beats Panic Mode

Boris starts 80% of his Claude Code sessions in plan mode, and it's the opposite of how most people work. They jump in, type a vague request, and let the AI run wild. The problem: the solution Claude thinks it's solving and the solution you actually need are often completely different. Austin shares a client story where Claude was supposed to fix a display bug but instead changed database values, which broke five other things. By planning first, asking clarifying questions like what's the core problem, who's it for, and what should this not do, you put Claude in a position to succeed. The Navy SEAL wisdom applies here: move slow to move fast.

Keep Claude.md Lean or Blow It Up

Claude.md is a cheat sheet of instructions Claude reads every chat. Boris's counterintuitive advice: if it gets too long, delete the whole thing and start fresh. People load it with every rule imaginable thinking more instructions equals better results, but the opposite happens—Claude gets confused. Since models improve constantly, rules from six months ago are often baked in now. Boris recommends doing the minimal possible thing, then adding rules back only when you see the same problem twice. Austin takes a middle ground approach, running a cleanup command instead of full deletion, but acknowledges Boris's strategy of periodic resets is probably more effective.

Verification Is Your Quality Multiplier

Boris gives Claude a tool to see its own output and tells Claude about that tool. Claude figures out the rest. In practice: if you're building a website, open a browser and test it. If you're writing, ask Claude to review against brand guidelines. If you're automating, verify the output matches expectations. One pro tip Austin loves: before any work starts, ask Claude what its verification plan is. If there's no clear way to verify something, maybe there's a better approach entirely. After going back and forth multiple times, you can also ask Claude to verify all its work so far against best practices.

Parallel Sessions Are a Force Multiplier

Running multiple Claude Code sessions simultaneously on different tasks is where efficiency explodes. The key: each task must be partitioned with zero overlap. If two sessions work on the same thing at once, you get wasted effort and file conflicts. Boris's insight is sharp: two context windows that don't know about each other get better results. A fresh session with no baggage sees obvious solutions the first session missed because it's too deep in the problem. It's like restarting and having it just work.

Claude Skills Turn Repetition Into Automation

Inner loops are tasks you do repeatedly throughout the day. Claude skills let you document the process once and call it every time after. Austin uses a skill that generates local law 97 reports in identical format and style—the only variable is the data. A slash command on Boris's end, but Claude skills are more applicable for most users. Think of it like a sports play: a prompt is telling a player to dribble, but a Claude skill is the exact play to run every single time. These are genuinely powerful if you identify your actual inner loops first.

Never Bet Against the Model Getting Better

Boris keeps a framed copy of Rich Sutton's Bitter Lesson on the wall. The core idea: the more general model will always beat the more specific one. Translation: every micro-tweak and scaffolding you build today will probably be unnecessary in six months because the model improves daily. This doesn't mean you stop building, but it changes where you focus energy. Prompt optimization is pointless—Austin calls it stupid—because next month's model won't need your clever wording. Spend time instead on information architecture and the system that feeds the model. AI will never be as bad as it is today, so build assuming it gets better.

Takeaways

  • Start every Claude Code session in plan mode. Interview the model about your assumptions before it writes any code. Ask what the core problem is, who it's for, what success looks like, and what it should not do.
  • Keep claude.md short and delete it fresh when bloated. Add rules back one at a time only after you see the same mistake twice. Don't over-engineer with 50 rules you think might help.
  • Give Claude a verification method so it can check its own work—browser testing, output validation, brand guideline review. Ask for the verification plan before work starts.
  • Run parallel Claude sessions on different tasks simultaneously. Fresh context on separate problems prevents the baggage that slows down a single long session.
  • Create Claude skills for your daily inner loops—repetitive processes that need consistency. Document once, call infinitely.
  • Focus on information architecture, not prompt optimization. Models improve every month so tweaks expire fast. Spend energy on what you feed the model, not perfect wording.

Key moments

3:00Plan Mode Before Execution

Probably 80% of my sessions I start in plan mode. And once the plan is good, it just stays on track and it'll just do the thing exactly right almost every time.

7:00Claude.md Philosophy

If you hit this, my recommendation would be delete your quad and just start fresh. The more you give it, the more likely Claude is going to get confused and not apply the rules that are actually critical.

10:00Verification Doubles Quality

Give Claude a way to verify its work. If Claude has that feedback loop, it will two to 3x the quality of the final result.

12:30Parallel Sessions Strategy

Two context windows that don't know about each other tend to get better results. A fresh session could see something obvious that your first session missed because it was too deep in the weeds.

17:30Don't Optimize Prompts

Every piece of scaffolding or micro tweaks to improve the model's output that you do will probably be unnecessary in the next 6 months. AI will never be as bad as it is today.

Get AI-powered video digests

Follow your favorite creators and get concise summaries delivered to your dashboard. Save hours every week.

Start for free