11 comments

  • rimeice 1 hour ago
    I like this. I’ve been working on something similar. Good spec’ing is critical to getting good output and I suspect a lot of the “I’ve already got plan mode” comments are from technical HN folk, who do know the right questions to ask and know what good looks like. But as the success of Lovable shows there are millions of people out there who clearly want to build apps but don’t have the technical chops to do so and clearly don’t know what a good spec looks like. My experience of “plan mode” is that it won’t serve these people. I’d be keen to connect and share ideas around this. My email in my profile.
  • meander_water 1 hour ago
    I've been doing something like this for a while, and it provides decent enough results for agents to one-shot. The key is to specify the LLM with the idea, and then get it to ask you questions until you feel enough ambiguity has been eliminated from the product spec for you to pass it to an agent.
    • straydusk 1 hour ago
      I think the reverse Socratic method when it comes to LLMs is one of the best constructs, period - both in this context, but in many.
  • Aperocky 3 hours ago
    AI can generate specs and development plans but a tremendous amount of trash can slip through those specs and plans resulting in complete garbage as output.

    I'm using AI a lot, in planning but I take close manual oversight on specs and development plan and still read all active path code (give AI a little but not too much leeway on testing, since sometimes they start writing test asserting true == true).

    • straydusk 1 hour ago
      I think you're right. I think that's why I like this approach - it's keeping you in the loop for most of it - and you just get documents as output that you can edit, check, and iterate on yourself.
  • nvader 7 hours ago
    I'm happy for you! Some feedback as I walk through the app.

    I like the styling, it's really slick. I also like that you enable me to use the tool online without signing up. I was curious about how you're supporting this, and paying for inference, but I see now that you haven't really wired up anything. When I try to generate my 1pager, it returns a placeholder.

    > Turn your messy ideas...

    I'm not a fan of this framing. Messy has negative connotations, so it's not clear why you're insulting me when we just met. ;)

    The wizard:

    There's a bit of duplication, since you have "Tell the agent..." as well as "Tell me...", both conveying the same information.

    I can jump through steps without competing prior ones. Isn't that going to cause a problem?

    It's hard to truly evaluate this further without seeing it in action. As other authors have said, many agents already support Plan Mode, so it it's important for you to distinguish yourself from that.

    • straydusk 7 hours ago
      Thanks for the response! I have wired it up...it's using gpt-5-mini. I just have API-level usage limits - I figured if it hit those, I could worry about how to adjust costs, rate limit, etc.

      The jumping through steps is not intended - that's a regression.

      I agree on plan mode - this one is just a lot more featured. I should include some samples to demonstrate that. Here's an example, if you're interested, of the prompt plan output - https://github.com/benjaminshoemaker/data_graph_gap_report/b...

    • straydusk 6 hours ago
      I'm honestly not sure how you got the "turn your messy ideas" placeholder - how did you get to that?
      • adambb 5 hours ago
        I think the "messy ideas" was a reference to the homepage copy "Turn your messy ideas into crystal clear specs.", not continuing the previous thought about the placeholder. I'd agree that "messy" might have more negative connotations than you intended.
        • straydusk 5 hours ago
          Interesting, fair though. I think of my own ideas that way, but maybe not everyone does!
  • purge12 1 hour ago
    Is this still up? I tried to get some ideas on creating AGI, but no AI response came through?
    • straydusk 1 hour ago
      Uh... I have no idea why the chat isn't working! Looking now!
  • zahlman 5 hours ago
    > It is a wizard-style AI tool that will guide you from idea → vision → tech spec → implementation plan.

    (And then the implementation plan is fed to the same sort of AI that you were going to give the "idea" to in the first place.)

    If doing this gives good results, then it shouldn't be necessary.

    • majormajor 3 hours ago
      This the standard current approach for most models/agent tools because models can do well at "make a plan for this" and "execute this step" but are less good at generating a response string that includes both the plan and every step of the execution without intermediate prompting/redirection/focusing. Helps fight context drift and maximize effectiveness/efficiency of the predictions.

      Most advances in tools I've used in the last two years are exactly this sort of "automate the steering and feedback loop that the prompt goes through" automated-fairly-boilerplate-sequencing of refinement of initial idea -> plan -> execution -> feedback.

      • straydusk 46 minutes ago
        Thanks for saying this.
    • FuckButtons 4 hours ago
      Why? From first principles you can deduce why it makes sense this could work, they are auto regressive next token prediction engines. As to efficacy, well that you would need to try it and see, but I see no reason to dismiss the idea out of hand.
    • straydusk 5 hours ago
      I mean, it would be great if it wasn't necessary - but have you used these coding agents? They need it.
  • NamlchakKhandro 2 hours ago
    broken.

    there's no need for an app like this anyway.

    You want this as a series of prompts that handle the various stages.

    • straydusk 1 hour ago
      I mean, that is in large part what this is. I just turned it into an app because a) it was a lot of manual work for me and then b) this way other people can use it.
  • esafak 8 hours ago
    You can already do this in your coding agent ? You just need the LLM to index the code base, and the rest is easy. I need a before-and-after to understand what you are doing differently.

    Unless your tool has people skills, this engineer can just take the spec to the agent ;)

    • straydusk 8 hours ago
      Maybe that's something I should add, to more clearly illustrate it. This tool is creating a spec that the AI coding agent of your choice can execute on, it's not an AI coding itself. Or maybe I'm not sure what you're getting at?
      • wredcoll 8 hours ago
        But all ai models will generate a spec if you ask them to?
      • Avicebron 8 hours ago
        I'm not sure if it's intended behavior or not, but after it generated the spec it did offer to write the code and then generated it
      • esafak 8 hours ago
        I mean I can already iterate on a spec with the agent; I do not need a separate tool.
  • CuriouslyC 4 hours ago
    What are you doing that SpecKit/BMAD/etc aren't?
    • straydusk 1 hour ago
      I found SpecKit to be over-engineered, and BMAD to be quite complex. To me, this was a "thin" enough solution that I still felt like I was in control, and not at the mercy of an additional abstraction.
  • doppelgunner 3 hours ago
    [dead]
  • baculumguy 35 minutes ago
    [dead]