digestVidSign in
summary

Austin Marchese - How Claude Code’s Creator Starts EVERY Project

Thumbnail of the video “Austin Marchese - How Claude Code’s Creator Starts EVERY Project”
Community upload. AI output has not been independently verified.
Author: Austin Marchese
Categories: Education Productivity
Tags: education, productivity
Views: 2 · Last viewed: 2026-10-15 07:48 UTC

Short Summary

Main recommendation: start every Claude Code project in plan mode, do the thinking and questioning up front, and keep your CLAUDE.md file minimal, since Boris Cherny says plan mode sessions "just stay on track" and build almost correctly every time. The video pairs this with giving Claude a way to verify its own work, running parallel sessions on partitioned tasks, and turning repeated tasks into Claude Skills.

Key Points

  • Plan mode first: Boris starts about 80% of his sessions in plan mode, and once the plan is good, the build step is nearly automatic.
  • Vague requests cause wrong solutions. The speaker's example: a client's numeric display bug was "fixed" by Claude changing the underlying database value, which broke about five other things in the app.
  • Keep CLAUDE.md short. Boris's file is roughly a couple thousand tokens, and he advises deleting a bloated file and adding rules back only when the model goes off track.
  • Verification is high leverage. Boris's claim is that giving Claude a feedback loop will "2-3x the quality of the final result."
  • Parallel sessions work when each task is partitioned so sessions don't overlap, and a fresh context window can catch things a deep session missed.

Details Worth Noting

  • Concrete instructions:
    • Before building, press Shift+Tab twice in the terminal to enter plan mode, then go back and forth until the plan is right.
    • Use this planning prompt: "Interview me about this. What are the core problems this solves? Who is this for? What does success look like? And what should this not do? Summarize it back to me before you write any code."
    • Add a verification plan to CLAUDE.md before work starts, so Claude states how it will check its work before building. If there is no clear way to verify, rethink the approach.
    • Run a check after work: "Please go back and verify all of your work so far. Make sure you used best practices, were efficient, and didn't introduce any issues."
    • To prune CLAUDE.md, either delete it entirely and rebuild, or use the speaker's middle ground: "Update my CLAUDE.md to remove anything that's no longer needed, contradictory, duplicate information, or unnecessary bloat impacting effectiveness."
    • When a mistake recurs, ask: "Based on this conversation, can you update CLAUDE.md so this doesn't happen again?"
    • Give Claude a verification tool (for example, a browser for web work), then tell it that tool exists.
    • Use slash commands for repeated daily workflows; for a repeatable process, create a Claude Skill. To find candidates, ask: "Based on the project I'm working on, what Claude Skills should I create?"
  • The speaker's example skill is /l97, which generates a Local Law 97 report in the same format each time, changing only the input data.
  • The speaker's own CLAUDE.md practice differs from Boris's: he admits he feels scared to delete it outright and prefers the pruning command.
  • The "Bitter Lesson" principle (Rich Sutton) is cited as the reason to avoid over-investing in micro-optimized prompts, since general models keep improving.

Bottom Line

The real leverage is in planning, verifying, and systematizing repeated work rather than in clever prompt wording, because the underlying model keeps improving.