What advice do you have for someone new to tokenmaxxing? Specifically with Fable 5 but also in general (vid coming soon)
T3 Stack creator Theo Browne asks community for 'tokenmaxxing' and token optimization strategies for Fable 5
Browne is collecting the optimization tips for an upcoming video
Many users criticized tokenmaxxing with Fable 5 as pointless and urged focusing on actual business value instead, while some praised Theo personally for the approach.
Most Activity

@theo think of loops as 'automated context delivery' and unironically do the loops
i.e this PR the recording of it was done through computer use, try to recreate your dev tools in a way the agent can use and build tests + iterate with

start with nothing. don’t download any plugins or skills. no AGENTS.md either. just prompt and observe the failure modes. codify them into skills when they repeat. even better, write lint rules or structure your codebase so certain mistakes are impossible.
every chef has their own set of knives, you should too
or just use pstack heheh
@theo Link in my bio
What advice do you have for someone new to tokenmaxxing? Specifically with Fable 5 but also in general (vid coming soon)

@ChainZenit I'm over $10,000 in tokens spent in the last 11 days

@theo https://github.com/tirth8205/code-review-graph This helps save tokens.

Project decomposition into tasks is super important. my workflow has been: - get fable to plan out the project, and then break it down into tasks, with explicit goals. I do this in Claude code. the tasks explain what to do, provide code pointers, and a "done" criteria - then in pi, I'll spin up gpt 5.5 and start getting it to tackle a task at a time - after each task is done, I get fable to review, and then I pass that feedback to gpt 5.5
and then repeat...
I used to use paperclip for this workflow, but fable is $$$ with paperclip, so have been manually doing it. gpt 5.5 can be replaced with a faster model like composer 2.5, but ya. overall splitting up code execution from planning/validating has helped quite a bit.
also not having a lot of skills/mcps setup has been helpful which one of the reasons why I use http://pi.dev(tho not for fable, since I prefer Claude code with Fable).

@theo The definitive guide

@theo Everything digital can be solved given you throw enough tokens at it

@theo Shameless plug but I’ve been running fable 5 24/7 because I’ve been using this tool I built

Mainly for people "building" things:
You need to be more ambitious than you have been before. Ask the model to rewrite your entire production app from scratch. Ask it to make an entire product / internal tool for you, but don't stop there. Ask it to deploy it and add accounts, multiplayer, etc. All of the things you wouldn't normally have in a throwaway personal app. Raise your bar. Only by pushing more do you learn. Orchestrate. When coding and having an agent write a plan, have another agent, or multiple, validate each claim in that plan. Claude workflows are fucking amazing for this. Fable is really good at orchestration. Have it use codex -p to delegate parts of it to codex, have it use opus subagents, have it use workflows. Don't be afraid to give these models big chunks of work. Give your agent a browser / computer use. Let them click around and take screenshots, use the JS debugger, and take performance profiles. LLMs aren't just coding machines; they can do the debugging / reproducing as well. For token maxing, get some way to kick off coding tasks from your phone. That could be the codex app / chat gpt app / claw.

@theo Have Opus review work with GPT 5.5 and list out things to improve, fix, change by priority and then fix and then repeat until Opus and 5.5 both agree the only things left to fix have diminishing return.
Run on /goal or /autopilot

@theo wait, how long have you been grinding on this one?

@theo Try to do things you think are impossible.

@theo Make Fable the architect. Have it write the plan at medium effort, then let Sonnet or Haiku execute the checklist. Frontier tokens for judgment, cheap tokens for execution.

my personal best piece or advice is to focus on the goal first. it's fine if it takes you several iterations to figure out what the right goal is, but once you do, scrap everything you've done (or at least summarize the research you did along the way), then start fresh with a right goal upfront.

@theo technically not through specifically computer use but through testing via using a computer on some wrappers i have, but i've been liking this setup

I'm torn. I am an engineer so I prefer efficency. But is the argument that you must use frontier as much as possible because it's better at everything? Or do you want to token minimize? Use the absolute most efficient process you can. Only "tokenmaxx" when it's the highest roi. I'm not a computer scientist so I'm actually genuinely curious.

@theo You want to use Fable mostly as a planner/ideator, orchestrator and adversarial reviewer. Not a pair-programmer. More like a technical lead you’re working with to who has people in charge to delegate stuff to, and review their work. People here being Opus and Sonnet.

@theo more stuff on this:

@jyothiwrites @theo I’ve tried this and the cheaper model executes just wrongly and the architect has to correct everything.