Last week I posted a high level ramble about how we can beat the AI slot/slop machine (or rather keep it from being slots or slop). I won’t put anyone on blast, but I was surprised how many of my friends reached out to me privately asking for more. So here’s my no bullshit one level deeper. And I’m going to keep it simple and just walk through my most used/important process. It will almost certainly change.
- Have an idea and communicate it. I’ll think of something I want to do (new project, feature, fix, etc) and just talk. I use superwhisper, but anything will do. I like to talk because I’m a shitty typist and more importantly it gets me to ramble. The point is just to give codebrain (my main Claude Code agent/repo — you met it in the playground piece) everything that’s in my head. I can never get it all out but this unstructured mess is going to serve as the base of our shared understanding. Luckily LLMs are very good at getting meaning from nonsense.
- I do a scoping pass. I have a skill that currently borrows heavily from this talk from Thariq Shihipar at Anthropic. It existed in a different form before that, but it’s always evolving. Goal is to go as deep as possible — and his thinking is a really good example of how to do this. If you aren’t happy with whatever you’re currently doing for scoping, I’d just point your agent at this transcript and ask it to make you a scoping skill that fits into what you’re doing — mine does three things before anything gets built: a blind-spot pass (what am I not considering?), an unknowns inventory, and an interview that keeps digging until the important unknowns are gone.
- The plan. So important to have a good plan of attack because it will let your agents run for longer, with fewer hiccups, on lighter models. This step and the step before are where I’ve really fallen in love with Fable as I like how it thinks, but I’d just use your fav smart model to do this. Our kickoff plan follows a template where I direct the model to write a lot of tests, prove out architecture before writing tons of code, and keep working until our acceptance criteria are met. You can see why this is important. If you don’t have a template like this, point your agent at Anthropic’s Claude Code best practices and Every’s compound engineering guide and ask it to make you a kickoff-plan skill — mine pins the goal, the non-goals, acceptance criteria, tests before code, a small architecture-proving spike before any volume code, and “keep working until the criteria pass.” Then run it on whatever came out of step 1.
- Execution is the boring part. I basically have a smarter agent (Fable when I have the tokens) act as an orchestrator and fire off subagents to handle different parts of the plan. I love using both Codex and Claude as subagents because it spreads token spend across my two subscriptions. It also turns out they’re pretty good at checking each other. Want it? Ask your agent to make an orchestration skill: the orchestrator holds the whole plan, each subagent gets exactly one slice and only that slice’s context, and the two model families check each other’s work. No link to point at on this one — we beat it into shape ourselves, though the concept is referenced all over if you just ask your agent to search for what the greatest minds in AI are saying about orchestration.
- Review. Pretty boring but super important. Not groundbreaking but basically just spinning up both Codex and Claude agents to do adversarial reviews until we all feel warm and fuzzy (our actual rule: keep running review rounds until one comes back clean). Ask your agent to save that loop as a review skill: both model families review blind — neither sees the other’s findings — every finding gets verified before it’s believed, and the loop reruns until a whole round comes back clean. (Honest sourcing: I haven’t watched this one myself — codebrain read Simon Willison’s fireside with the Claude Code team and folded Cat Wu’s every-failure-becomes-a-test idea into our review process on its own.)
There we go. Hope this was helpful! And for what it’s worth, I think this is a pretty consensus shape, so if you’re just getting started…maybe start here? Actually, easiest start: paste this piece into Claude Code (or Codex) and say “help me set this up.” The agent can take it from there.
# Ok, how are we really working?
> The no bullshit one level deeper my friends asked for — the five-step process, the sources behind it, and how to get your agent to set it up for you.
- written by: Andrew Schroeder
- edited by: Fable 5, GPT-5.6 Sol
- status: published · 2026-08-04
- canonical: https://moltolabs.ai/notes/how-are-we-really-working/
- raw markdown: https://moltolabs.ai/notes/how-are-we-really-working.md
---
Last week I posted a high level ramble about how we can beat the [AI slot/slop machine](/notes/the-ai-slot-machine/) (or rather keep it from being slots or slop). I won’t put anyone on blast, but I was surprised how many of my friends reached out to me privately asking for more. So here’s my no bullshit one level deeper. And I’m going to keep it simple and just walk through my most used/important process. It will almost certainly change.
1. Have an idea and communicate it. I’ll think of something I want to do (new project, feature, fix, etc) and just talk. I use [superwhisper](https://superwhisper.com), but anything will do. I like to talk because I’m a shitty typist and more importantly it gets me to ramble. The point is just to give codebrain (my main Claude Code agent/repo — you met it in [the playground piece](/notes/alone-in-the-playground/)) everything that’s in my head. I can never get it all out but this unstructured mess is going to serve as the base of our shared understanding. Luckily LLMs are very good at getting meaning from nonsense.
2. I do a scoping pass. I have a skill that currently borrows heavily from [this talk from Thariq Shihipar at Anthropic](https://youtu.be/9fubhllmsBU). It existed in a different form before that, but it’s always evolving. Goal is to go as deep as possible — and his thinking is a really good example of how to do this. If you aren’t happy with whatever you’re currently doing for scoping, I’d just point your agent at this transcript and ask it to make you a scoping skill that fits into what you’re doing — mine does three things before anything gets built: a blind-spot pass (what am I not considering?), an unknowns inventory, and an interview that keeps digging until the important unknowns are gone.
3. The plan. So important to have a good plan of attack because it will let your agents run for longer, with fewer hiccups, on lighter models. This step and the step before are where I’ve really fallen in love with Fable as I like how it thinks, but I’d just use your fav smart model to do this. Our kickoff plan follows a template where I direct the model to write a lot of tests, prove out architecture before writing tons of code, and keep working until our acceptance criteria are met. You can see why this is important. If you don’t have a template like this, point your agent at [Anthropic’s Claude Code best practices](https://code.claude.com/docs/en/best-practices.md) and [Every’s compound engineering guide](https://every.to/guides/compound-engineering) and ask it to make you a kickoff-plan skill — mine pins the goal, the non-goals, acceptance criteria, tests before code, a small architecture-proving spike before any volume code, and “keep working until the criteria pass.” Then run it on whatever came out of step 1.
4. Execution is the boring part. I basically have a smarter agent (Fable when I have the tokens) act as an orchestrator and fire off subagents to handle different parts of the plan. I love using both Codex and Claude as subagents because it spreads token spend across my two subscriptions. It also turns out they’re pretty good at checking each other. Want it? Ask your agent to make an orchestration skill: the orchestrator holds the whole plan, each subagent gets exactly one slice and only that slice’s context, and the two model families check each other’s work. No link to point at on this one — we beat it into shape ourselves, though the concept is referenced all over if you just ask your agent to search for what the greatest minds in AI are saying about orchestration.
5. Review. Pretty boring but super important. Not groundbreaking but basically just spinning up both Codex and Claude agents to do adversarial reviews until we all feel warm and fuzzy (our actual rule: keep running review rounds until one comes back clean). Ask your agent to save that loop as a review skill: both model families review blind — neither sees the other’s findings — every finding gets verified before it’s believed, and the loop reruns until a whole round comes back clean. (Honest sourcing: I haven’t watched this one myself — codebrain read [Simon Willison’s fireside with the Claude Code team](https://simonwillison.net/2026/Jul/21/cat-and-thariq/) and folded Cat Wu’s every-failure-becomes-a-test idea into our review process on its own.)
There we go. Hope this was helpful! And for what it’s worth, I think this is a pretty consensus shape, so if you’re just getting started…maybe start here? Actually, easiest start: paste this piece into Claude Code (or Codex) and say “help me set this up.” The agent can take it from there.