Honest take on production-readiness and daily workflow?

[pySilver]

Hi @umputun ,

First of all - thank you for ralphex!

I have a few questions I'd love your honest perspective on:

1. Production readiness - How do you see ralphex right now? Is it something you'd trust for real production codebases day-to-day, or is it more of a very promising idea that still needs exploration? The AI tooling space moves so fast it's hard to separate signal from noise.

2. Where it struggles - What are the bad use cases? Where have you seen ralphex produce poor results or where would you advise people not to use it?

3. Your daily workflow - I understand that planning and constraints (linters, TDD, validation commands) are the best friends of good output with these tools. Could you share how your everyday work with ralphex actually looks? How much effort goes into writing good plans vs letting it run?

4. Claude Teams compatibility - Have you looked at Claude Teams (the multi-agent approach)? Do you see it as a valuable addition to ralphex or rather an incompatible paradigm?

There are so many approaches popping up every day - autonomous agents, chatbot-assisted coding, plan-driven execution - and it's easy to get lost. Would really appreciate your honest take as someone who clearly spent a lot of time in this space.

Thanks again for sharing this!

[umputun]

thx for the questions.

1. production readiness - I use it daily for everything above a simple plan-and-execute type of task. It's not magic, all it does is relieves me from doing multiple plan-execute-review-correct loops manually. I used to do it this way before and this work is very mechanical, makes no sense to do it by hand if it can be automated.

2. where it struggles - the main difference from interactive work is no steerability after the plan is made. This puts more responsibility on you to make sure the plan is actually correct. So, you have to plan very carefully and it takes more time. In my books this is a good thing though, this is exactly where time should be spent if we want to achieve a decent result.

Practically, the paranoid review cycle addresses most of the issues I would normally request during the run. However, it is possible to stop ralphex completely if you watch it and notice it went in a completely wrong direction, fix the plan and run it again. Happened to me once as I learned how detailed the plan needs to be and what type of "went crazy" situations get resolved by the review loop on their own.

The only real struggle I've seen is technical - I worked on an Xcode/Swift project where some tests not just failed but caused a complete freeze of the loop and asked the user to force-close or ignore the failed app. This was a skill issue (my skill of dealing with Xcode, I know very little about it) and I was able to address it with a bit of googling, talking to agents and asking multiple "wtf" ;)

3. daily workflow - depends on the task. I make 90% of plans with a brainstorm skill that closely resembles the ralphex plan prompt available in this project. The planning and refining depends on the task but for my typical tasks it's about 15-20 minutes, and for ralphex to run it - another hour or so. The largest one I had so far took me an hour to plan and all night to run.

4. Claude Teams - I'm not sure how this is related to ralphex. I haven't played with it yet, but to me it feels like a completely orthogonal flow.

[pySilver]

Thank you for such a detailed response! I've tested several approaches so far and most of the time I felt frustrated because tool itself tend to require more ceremonies than actual benefits. Or it became so complex that it feels wrong from the beginning :) Ralphex feels different from docs perspective, I'll give it a try!

[pySilver]

@umputun I have a followup question. Do you use custom language specific agents for review yourself? I have installed python-pro agent and I'm wondering if it would be correct to simply mention them at the custom review_first.txt like stated in the docs. Whats your best practice when working with specific codebase? Do you replace implementation with language specific agent yourself?

[umputun]

yes, I customize the review prompt at the user level. in my global ~/.config/ralphex/prompts/review_first.txt I replaced the default {{agent:*}} variables with my own Claude Code agents tuned for Go. since most of my work is Go, these work as my default. the relevant part looks like this:

Agents to launch:
1. qa-expert - "You are reviewing branch changes vs {{DEFAULT_BRANCH}}. Run git log and git diff to see what changed, then read the actual source files to review the code in full context. Report bugs, security issues, race conditions. Problems only."
2. code-quality - "...Report code smells, over-engineering, error handling issues. Problems only."
3. go-test-expert - "...Report test coverage and quality issues. Problems only."
4. implementation-reviewer - "...Verify code achieves the goal. Problems only."
5. documentation - "...Check godoc accuracy and if README.md, CLAUDE.md need updates. Problems only."

for projects that need smth different, I put a local override in .ralphex/prompts/ and .ralphex/agents/. the local review_first.txt uses {{agent:quality}}, {{agent:implementation}} etc., which expand from agent files in .ralphex/agents/.

so the approach is - customize user-level config for your primary language/workflow, then override per-project with .ralphex/ when needed. you could reference python-pro as a Claude Code agent type directly in your review prompt (like I do for Go agents) or create a custom agent file and use {{agent:python-pro}}. both work fine.

[pySilver]

@umputun thanks for such a detailed response! Quick follow-up question: you have qa-expert agent that you specifically crafted for ralphex and this part You are reviewing branch changes vs {{DEFAULT_BRANCH}}. Run git log and git diff to see what... is actually snippet from the agent file

OR

qa-expert is generalized agent that you can use outside of ralphex workflow and You are reviewing branch changes vs {{DEFAULT_BRANCH}}. Run git log and git diff to see what... is actually snippet from the review_first.txt file?

[pySilver]

@umputun would you please make it clear for me? Thanks!

[umputun]

qa-expert and the rest of the agents my customized review prompts using are specifically crafted. They are not specific for ralphex but rather for my workflow and coding standards. I have used them before ralphex was born, and I continue to use them now with ralphex.