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.
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 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.
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.
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).
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.
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.
This is really good work! Love the UX and the design. Even though I work with AI agents and building them a lot, I found the way you structure the steps very helpful for thinking about the project itself. The agent asks the right questions which makes me feel confident it's understanding my goals.
If you're looking for feedback, you could include a tiny section on the homepage about how to run the output docs. e.g. put them in a folder, point Claude Code/Codex to it and give it the prompt.
Thanks a lot! I agree (obviously) - this workflow really helped me structure my projects.
> If you're looking for feedback, you could include a tiny section on the homepage about how to run the output docs. e.g. put them in a folder, point Claude Code/Codex to it and give it the prompt.
This is great feedback - I totally agree - going to add this asap
I've never understood why anyone would use a wizard to install software: they are unstable old men with magical powers and they are NOT to be trusted!!!
Now you are going to have one write out instructions for an AI?! I guess we know how the AI apocalypse gets started!
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.
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.
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 ;)
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?
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.
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.
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.
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...
Hey there, I changed the messy framing. Hope you like it!
I also fixed the wizard text, I agree.
I also fixed the step jumping.
Would love to hear your feedback if you try it again with the fixes in place :)
I'm honestly not sure how you got the "turn your messy ideas" placeholder - how did you get to that?
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.
Interesting, fair though. I think of my own ideas that way, but maybe not everyone does!
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.
Shot you an email!
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).
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.
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.
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.
This is really good work! Love the UX and the design. Even though I work with AI agents and building them a lot, I found the way you structure the steps very helpful for thinking about the project itself. The agent asks the right questions which makes me feel confident it's understanding my goals.
If you're looking for feedback, you could include a tiny section on the homepage about how to run the output docs. e.g. put them in a folder, point Claude Code/Codex to it and give it the prompt.
Thanks for building this!
Added a new modal for this - I really like it! Thanks for pointing that out - it was a bad experience before! https://imgur.com/a/VLbjuW5
Thanks a lot! I agree (obviously) - this workflow really helped me structure my projects.
> If you're looking for feedback, you could include a tiny section on the homepage about how to run the output docs. e.g. put them in a folder, point Claude Code/Codex to it and give it the prompt.
This is great feedback - I totally agree - going to add this asap
I've never understood why anyone would use a wizard to install software: they are unstable old men with magical powers and they are NOT to be trusted!!!
Now you are going to have one write out instructions for an AI?! I guess we know how the AI apocalypse gets started!
> 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.
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.
Thanks for saying this.
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.
I mean, it would be great if it wasn't necessary - but have you used these coding agents? They need it.
My point is that it shouldn't be necessary because the agent should already implement it.
I wish it did as well. I am sure someday, this tool won' t be needed. Until then... maybe it will!
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 ;)
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?
But all ai models will generate a spec if you ask them to?
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
I mean I can already iterate on a spec with the agent; I do not need a separate tool.
What are you doing that SpecKit/BMAD/etc aren't?
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.
Is this still up? I tried to get some ideas on creating AGI, but no AI response came through?
AGI? In a single prompt?
I guess we should tell thousands of AI researchers to stop what they're doing right now since you're a single prompt away from solving the problem??
That was meant to be a satire dude. I just meant to say the chat isn't working :)
I fixed it, and a few other issues. It was because I hit my limits in the API... OpenAI billing is weird.
Uh... I have no idea why the chat isn't working! Looking now!
broken.
there's no need for an app like this anyway.
You want this as a series of prompts that handle the various stages.
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.