Forward Future AI founder Matthew Berman asks how to use leftover AI tokens, drawing a joke proposal to rewrite codebases from scratch

@MatthewBerman Have Claude Code translate your entire codebase to COBOL, then have Codex translate it to Latin, then ask Cursor to figure out what the app does. Loser gets deleted

@MatthewBerman - Sentry auto bug fixing. - analyse code produced during the day and fix bugs

@MatthewBerman happy to help send some my way !!

alternatively you could run Warden audits with gpt https://github.com/getsentry/warden

or adapt @every's visual prompt to run through all your UI interactions with opus/fable

@MatthewBerman @shadcn just shipped https://github.com/shadcn/improve and I have been having a nice time making using it. You could probably automate it in a way that uses up more idle tokens too

@_StanGirard Good idea!

@MatthewBerman You can go really deep into testing. Linting, have it fix those bugs, code complexity review and have it make the code less complex. All sorts of static analysis.

@MatthewBerman Nuke it and rebuild every Friday.

@MatthewBerman Depends on ur project but here are things I do personally 🫡

@smithstephen 🤣I want it to be actually useful!

@MatthewBerman the cursor thermo review skill

@MatthewBerman You can tokenmaxx with Cursor CLI or Codex OAuth by adding them to @nounsresearch Hermes.

You can still code yourself, but agents can handle recurring codebase tasks too. For example: an agent checks Sentry every hour via MCP, finds ongoing bugs, and opens fixes autonomously.

@MatthewBerman i have the same problem with just claude code...

@MatthewBerman Realistically if the code base is free of errors then you should have the models do daily checks for new CVE vulnerabilities. Maybe also check for new versions of dependencies and continuously move your code base forward.

@MatthewBerman Run scientific studies. Try not to create the AGENTS.md and the skills. Instead let the AI create a harness for itself and measure objectively (time, num tokens, $) which harness works better.

Same harness but 20 repos with 10 tasks in each. Which performs best overall?

I have a max plan on all

Codex automations are great. Run these daily for cleaning up codebases @steipete talks about his a lot and has good skills on his page.

The sweet thing is that Cursor has a great Automations feature too - and they can all reuse the same skill and trigger each other.

Make a fall back that can cascade to using Claude/cursor/codex since they all have CLIs.

@MatthewBerman Ship of thesus

Rewrite the entire codebase for each new model release

@MatthewBerman Have each build alternative versions of whatever you build, each based on an x-lens. Then have each one get the best ideas from the other two. Deploy all three, and pick then one that makes more money.

@MatthewBerman do you use superset or all these separately?

@MatthewBerman add loops to everything

@MatthewBerman weirdly - right now have codex looping backend tasks on at least xhigh - frontend use composer 2.5 and bring in fable / opus for heaver front end design tasks - lastly use a workflow with fable to do code review bugfixing as a planner agent with subagent opus

everything in use!