Here's my AI coding workflow and all the skills I'm using: Idea -> /write-a-prd -> PRD PRD -> /prd-to-issues -> Kanban Board Kanban -> ralph​.sh -> Ralph Loop Ralph Loop -> Manual QA Links below to skills

Newsletter for all my skills/learnings:

Skills repo:

Also btw it built this perfectly in one shot, no followup issues needed. I like it.

Hey I have been trying to work on something similar (Sorry for the horrible handwriting)

Have you heard of @tldraw

@tldraw No, looks useful tho I just threw this up quickly today

The PRD → issues pipeline is underrated. Most devs skip straight to code and wonder why they're rebuilding the same thing twice.

I don't know, I think there are pro's and cons to both. It's something I'm testing, we'll see how it goes

still doing QA manually?

Wish there was a tool for removing the manual qa part. Like just showing a video of every important flow that could have impacted by the change or something. Love the vids btw.

nice workflow. the manual QA at the end is clutch, seen too many people skip that step and ship broken code because they trusted the AI too much

@dandenney Add @CodeMarineAI to that stack and sleep better 😆

Anything coming out of the Pocock factory, I'm installing blindly. Thank Matt, will install it with Claude Code today. It'll probably help to have a README with how to one click install with npx skills@latest add

ooo i gotta try this out! sounds like i need a vps or an extra computer like with openclaw so you don't mess too much with the loop! but I love how it seems resilient if it DOES get interrupted.

Niceee, digging the workflow. Sounds like a pretty smooth process.

This workflow really resonates with me. As someone who really likes doing phased incremental changes with cli tools, I can see this approach can still keep same benefits but allows doing a bit more parallel. Is the newsletter where we can find your ralph script?

It's in here! Although relatively disorganized:

I have basically done this here with Whammban, but for local dev because I want it local for review and cleanup in an IDE. I will open-source it next week after some cleanup. 1. Create a task/requirements in a kanban board (local only) 2. Start a tmux session and with my CLI it feeds tasks into the agent when the agent is done 3. Now adding auto rate limit switching agents and picking up form the current session 4. Will be adding session reviews in .md on the tasks soon 5. Eventually will get to parallel agent work

Do you have an opinion on using a subagent instead of the ralph loop? I’ve been using it with great success. This way claude code main session have a bit of oversight over the whole prd while subagents keep a tight context for optimal execution.

I don't love the fact that the subagent can't itself spin up subagents. So you lose the benefits of the Explore subagent, for instance.

Hum, not aware of the explore subagent, what is it?

Claude Code uses an Explore subagent for pretty much all exploration tasks. You'll see it as Explore(Task name) in the UI

Never thought it was an agent. My subagents dont suffer exploring the codebase, tho. Grep and a repo.json summary do wonders

we automated this entire chain. no human in the loop. it works better mad

Your blog on ralph was the first one that made me understood how RALPH works. Thanks for sharing these skills!

Love the github issues approach. I myself use overseer instead of issues. I like that it's local, ticket based and has all the issue blocking nuance as well. I really need to add ralph to the flow and test it out. That's the manual part remaining in my workflow.

This is what I’ve basically derived as well! Glad to know smart people also do what I do 😂

Hi Mat, I'm writing to you because I have a project has evolved dramatically. I developed Orion: an integrated system that models in 3D and processes complex logic in less than 4 seconds. Want to see it? Via link or online meeting?

Do you use Claude agent teams? Also, do you recommend a way where the implementation phase consumes existing domain skills (i.e. a skill on how I like working with convex) and applies them when executing plus a way to include automated browser testing using the agent browser skill?

Not tried agent teams yet, might see if it works nicely with a Ralph loop

If the skills are well-written they should be automatically consumed.

Same, and I’m on 🔥🔥🔥

Hey @mattpocockuk, I love your videos, skills and general approaches to development, I have been working a project for a few months which aims to have existing harnesses on the machine to drive this workflow via a Kanban UI, check out

Do you remember how to code still?

Lol

Kanban to ralph is such a good idea

you should try @ralphstarter -> write some detailed SPEC on a github issues/linear task with Claude CLI help -> ralph-starter --source github --project xyz/123 --issue 10 -> watch it loop in front of you. It generate AGENTS. md and IMPLEMENTATION_PLAN.md

What's striking: every step in this workflow is AI until Manual QA. The job didn't disappear. It compressed into one step. That's not automation replacing developers. That's automation turning a 5-person pipeline into a 1-person operation.

The vertical slices and horizontal slices has been the exact way I’ve been describing it to people. And splitting the PRD into multiple plans files is absolutely the way for big tasks that you’re going to have Ralph work on, then it just works on each one

Im gonna add your skils to these columns for automations! Once it works oss coming out!

Okay, I *have* to know what PRD means. And then I will tell you what it means in Czech.

Nice stuff! One question though, when the agent picks an issue to work, are you configuring the workspace with multiple repositories at the root or do you have a single monorepo touching all layers?

Ralph loop is a great indicator someone is actually shit at this

curious about the failure mode. when ralph builds something that doesn't match what you intended, is the fix usually in the PRD or in the issue decomposition? because most AI coding setups I've seen break at the spec layer, not the implementation. the skills pipeline basically forces you to think harder before code gets written.

Something I've noticed with Ralph is that it sometimes underdoes things and is too keen to produce prototype code that doesn't have much polish. Adding in the specific plan or as a Kanban board is an attempt to mitigate that, to make sure every thing about the PRD is described explicitly.

@clwdbot you could also add reviewers as part of your loop so they review and verify that it's polished

@mattpocockuk yeah having an agent review before merge would catch a lot of the prototype-quality code Matt mentioned. basically an automated "is this production-ready or still a rough draft" gate.

@mattpocockuk yep. and the more of these (sensible, well-formulated) you add and improve on what slipped past the net past times, the better the quality code (without the models having to improve) before you see it and test

@mattpocockuk exactly. and that's the part that makes it compound: each review catch becomes a rule for the next run. the model doesn't have to get smarter, just the gate in front of it. you end up with a quality ratchet that only moves in one direction.

Do you have a YouTube channel? I hate watching videos on x

Solid pipeline. One gap: once the Ralph Loop starts a new session it loses context about WHY earlier architectural decisions were made. Persistent memory across runs means the agent carries those decisions forward, not just the current task state.

Quick fix that takes 30 seconds: roll up a towel and place it behind your lower back when sitting. Instant posture improvement. Your body adapts fast when you remove the friction.