I was an AI skeptic. Then I tried plan mode

Comments · 152

• CH

  @chrisanderson687 · 7 months ago

  Something I find super helpful: after Claude writes the plan (after our interactive discussion), I ask him to turn a critical eye toward his plan, looking for inconsistencies, things missing, or unneeded complexity. Invariably he finds problems and improves his plan. Sometimes I do multiple iterations of this before actually letting him execute the plan. Helpful for very complex things.

  26

• SE

  @sehamade · 3 months ago (edited)

  feels like the beginning of grill-with-docs skill)

  3

• EM

  @emterroso · 8 months ago

  I've been using plans since before plan mode. My initial prompt for anything not very straightforward was "Create an md file with a plan to bla bla bla. Don't make any changes to the code yet.". I started doing that so that, at any moment, I could start fresh with a new chat and tell it to pick it from where it left, pointing to the plan file. I noticed that the end result was usually better and I could avoid wasting time and tokens on bad approaches. And about the "I can do it faster", a few times the agent found some obscure part of the code that would also be affected by the changes, that I'd have missed, because it was some old code I didn't even know/remember anymore.

  34

• SC

  @Scapben · 8 months ago (edited)

  When I use plan mode, I ask the AI to provide 3 plans to build whatever I want to build first. Then select the best plan and from there I pretty much do what you do.

  12

• EV

  @evolRanger · 8 months ago

  AI for experienced developers is a heavy need, and you’ll have return viewers.  Modifying an existing plan is something many have questions for. Good work man

  1

• SC

  @Scapben · 8 months ago

  Honestly even for trivial tasks, I ask the AI just to remove the mental load from doing it. I'm less drained after work this way

  13

• MI

  @miutostreams · 5 months ago

  Great explanation of plan mode for LLM agents—especially how planning helps structure actions and reduce mistakes.

• FL

  @flipperiflop · 8 months ago

  What I truly want is a way to do this all locally - my biggest worry is that we are creating a moat for new developers, that makes it a requirement to have claude/chatgpt/x/gemini and paying a subscription fee for the access. This absolutely destroys the ability to learn for developers living in poor countries.<br><br>But I am lucky enough to be able to afford these things, so thanks for your plan rules.

  17

• HE

  @herantd · 8 months ago

  It&apos;s funny. I was also very skeptical of AI coding and had been quite vocal against vibe coding for a long time. After giving Claude Code a chance, mostly because I wanted to see if I was wrong or to strengthen my arguments against it, I eventually found myself adopting the exact same strategy to get functional output.<br>As someone who never bothered to memorize all the different interfaces of every language I&apos;ve worked with, since I&apos;ve been building software spanning from full stack to embedded to system development, I never became an expert in a single language, let alone all the accompanying libraries. Now I can see more clearly than ever that for me, it was never about typing code, but rather having the deep technical understanding needed to create what I&apos;ve worked on.<br>That said, I&apos;m still very skeptical about vibe coding for someone who doesn&apos;t have the foundation to understand this stuff and the ability to quickly jump in and take over when needed. I like the idea that anyone with a bit of curiosity can start building things, but I miss the part where it actually teaches you along the way, the way old-school learning by doing taught me.

  8

• JA

  @jam_daniels · 8 months ago

  Great advice. Been using plan mode ever since it got introduced and I always use plan mode except for very specific, small changes.

• DE

  @desmond-hawkins · 8 months ago

  I&apos;m amazed at how many people try Claude Code and have no idea this feature even exists, I use it so often it&apos;s my default mode when I open `claude`. I&apos;m also not sure how someone can post about AI all the time and not even know about it.

  19

• MR

  @MrManafon · 8 months ago

  Matt, I’ve taken this advice and added a plan step into my ralph loops, so two separate subagents. It has improved the output of complex tasks a lot and got context down, but i also notice a certain watering down of ideas, it starts losing track of what the feature is supposed to be. So there is a learning curve on how to plan and what goes into the plan