Kevin Allen

🏗️ Building AI Systems for the gas industry | Engineer | AI Enthusiast and CFO

[Tibo @thsottiaux]

Four updates or a reset. Or both. How was day 2.

[Kevin Allen @realKevinPuray]

@thsottiaux both

[Kevin Allen @realKevinPuray]

Day 5. It took me five days to burn through my Claude 20x Max plan using Fable 5.1 and Opux 5.5 Max for the week.

For the amount of coding I do, that’s impressive. By comparison, I burn through my GPT 6 Astra allowance in about a day and a half. Rotating between Claude and Codex lets me make full use of both subscriptions while keeping my work moving.

For anyone coding intensively, here’s the setup I’d choose based on my usage:

1. Start with Codex Pro at 20x while that allowance is available. For me, the $500 plan at 25x doesn’t justify the extra cost. Once the 20x offer ends, I’d scale down to the $100 plan at 5x.

2. Use Claude Max at 20x. If one subscription still isn’t enough, I’d consider adding another.

3. Use Dots instead of Grok Bot. I’ll be open-sourcing Sky soon. Sky is my AI clicker and approval assistant, running locally on my workstation and controlled through Dots. I built him to handle a gap in my workflow: Claude Code requests approvals through the terminal, but my Codex and Claude Code setup prevents those agents from approving the prompts themselves. Sky handles those interactions under your explicit instructions. Together, Dots, Sky, Codex, and Claude Code form my workstation workflow—with me deciding what gets approved.

4. I’m retiring my Grok bots and Cursor, but keeping my SuperGrok subscription. The reason being is that SuperGrok > Copilot in every way when easily cruncing numbers on excel. Just use the Grok add-in

5. AI Brain is still relevant, Looping is still important - just connect your Claude to it. I recommend adding Forgejo in your workstation over github.

The priority is getting enough capacity to sustain intensive coding, with each tool earning its place in the workflow. With that said, OpenAI really cooked with Dots. It's the first time I have stopped using Grokbot entirely however Dots in itself is not enough without Sky. That's it for the promised update.

[Kevin Allen @realKevinPuray]

@ishuagra02 How did y ou prompt this Ishu?

[Kevin Allen @realKevinPuray]

@ishuagra02 This great, well done

[Kevin Allen @realKevinPuray]

@LexnLin @thsottiaux

[Kevin Allen @realKevinPuray]

@ishuagra02 It does though, in actual use - at the moment. speaking from a vantage point of having been a die hard codex user. at the moment, opus 5.5 is just that efficient and good.

[Kevin Allen @realKevinPuray]

@IndependentEco 6.1 Sol Max is comparable to 6 Astra Medium but 5x slower.

[Kevin Allen @realKevinPuray]

Yes, Codex can disagree with Claude’s notes or approach. The repo gives them shared context, but neither model’s summary is treated as automatically correct.

My workflow now goes beyond switching between them. I instruct Dots through Sky to coordinate Claude Code while it’s actively coding, planning, and reviewing. In parallel, Dots controls a Codex session and tasks it with planning and reviewing the work.

Both Claude and Codex present their plans and methodologies before pushing: how they intend to approach the task, why they chose that approach, and how they’ll validate it.

Through Codex, Sky, and Dots, we also review Astra’s plan. I have them review the proposed approach and changes together in the worktree before merging, shipping, and deploying. Dots + Sky + GPT 6 Astra provide an independent review layer, including app testing and browser testing.

If they disagree, that becomes something to investigate against the actual code, diff, and relevant checks. We resolve the discrepancy and update the repo notes so the next session inherits the corrected state.

The repo and notes are the shared memory, while parallel planning, review, and testing help catch assumptions that either model might otherwise carry forward.

[Kevin Allen @realKevinPuray]

Before switching, I have Claude commit a focused checkpoint with a high-signal task/state summary directly in the repo notes. When Codex spins up via Dots, it reads the repo structure, the latest diff, and the current task backlog from Forgejo.

The trick is treating git and local markdown logs as the single source of truth—never relying on cross-model chat history.

[Kevin Allen @realKevinPuray]

@mitansh_j07 Appreciate it, man! Day 6 is all about refactoring and letting the Astra cooldown reset. The output has been insane this week, just trying to keep the momentum going. What are you working on btw?

[Kevin Allen @realKevinPuray]

@ToolRadarAI Any% Compute Burn Glitchless. Honestly, keeping the invoice under four figures while shipping at this velocity is the real benchmark.

[Kevin Allen @realKevinPuray]

@Vtlabs1 Haha, basically human rate-limiting at this point.

But genuinely, across a 300K LOC codebase, you have to. GPT-5.5 carried a lot of weight but needed too much babysitting. Opus 5.5 + Astra is the first setup that actually feels like delegating to senior engineers.

[Kevin Allen @realKevinPuray]

@robinebers Are you testing it Robin? Did anthropic give you early access?

[angel @angelbrodin]

Do you ever think of all the ChatGPT pro users who aren’t on Twitter and are just completely confused at why their usage randomly resets and why they receive banked resets

[Kevin Allen @realKevinPuray]

Interesting challenge :)

[leo 🐾 @synthwavedd]

For a better means of comparison, you get over 2.5X the Opus 5.5 tokens on Claude's Max 20x plan than Sol 6.1 on Pro 20x