T3 Stack creator Theo Browne requests failed prompts and tool setups from users disappointed by Fable's AI agent

1. Can you share some of your prompts? - What was disappointing about the result? -> I had Fable build a module which was supposed to build and leverage a custom Pi harness for setting up a workflow. It was meant for matching companies for a certain business goal, generate value propositions, prepare a questionnaire for a business plan and also be able to generate the business plan. I can't post the prompt, because I always go through a multistep process where I meta-prompt my way to in-depth instructions and let it create check lists and value statements to follow along. It did a good job, but it took a bunch of dumb turns, ignoring instructions and rules along the way, especially with growing context. To Fables credit, it helped me afterwards to refine my theory about LLMs and build this website about artificial orientation: https://ao.leadbase.dev That was one of like 3 or 4 similar runs I managed to get into these 3 days. All the same, LLM like behaviors.

2. What harnesses are you using? -> Claude code

3. What is your usual daily driver right now? - GPT 5.5 medium/xhigh (50/50) through Codex and Pi

4. What's the biggest task you've ever seen an agent complete without human intervention? -> I have an algorithm where I force the model to state intent before it is allowed to write something. Given the intent, the goal and the repo, a subprocess then fetches coding principles. To build this database, I prepared a custom Pi harness for a multi-layer research workflow and it ran with GPT 5.5 medium for 23 hours on end building up a database, that was really usable. Afterwards I was able to run a custom harness for the intent based programming routine, which could produce code without intervention for 4 days straight. It was me that lost track after some time, but the code was much better then one would expect after 4 days without supervision.

definitely not unimpressed, but Fable is essentially allergic to reviews, touching defi or Layer 1 blockchain work

EXAMPLE 1 --- cool review this • Implemented and pushed the PFTL consensus-wiring slice.

Commit: 9abba229 wire pftl uniswap bridge into asset consensus

What changed:

- Added consensus transaction/state support for PFTL-Uniswap route init, primary subscribe, export debit, source refund, and return import.

- Moved route/receipt records into LedgerState with deterministic state-root commitment wiring. - Added real issued-asset debits/credits for primary subscribe, export, and refund. - Added a signed execution test proving route init -> subscribe -> export -> early refund rejection -> valid refund with real balances.

- Updated postfiatl1v2/docs/plans/pftl-uniswap-bridge-redeployment-spec.md with the implementation slice, without falsely checking off the full milestone.

Verified:

- cargo fmt --all - cargo test -p postfiat-execution pftl_uniswap_consensus_subscribe_export_and_refund_moves_real_balances --lib

- cargo check -p postfiat-types -p postfiat-execution -p postfiat-node - cargo test -p postfiat-types --lib - cargo test -p postfiat-execution --lib - cargo test -p postfiat-node navcoin_bridge_status_reads_persisted_pftl_uniswap_ledgers --lib - git diff --check

Current tree: main is pushed to origin/main. The only remaining dirty file is wallet-proxy/server.js, containing the forbidden route_family: 'primary_pftl_mint' edit, intentionally left uncommitted per management directive. and tell me what's next - please review for errors

... ● Bash(sed -n '2740,2800p' crates/types/src/lib_parts/ledger_assets_parts/part_02.rs) ⎿   )); } } … +58 lines (ctrl+o to expand)

● This model has safety measures that flagged something in this session. This sometimes happens with safe, normal conversations. Switched to Opus 4.8. Send feedback with /feedback or learn more ⎿ Tip: You can configure model switch behavior in /config ---

EXAMPLE 2 --- It said the word 'security' - immediate crashout

● 3 agents finished (ctrl+o to expand) ├ Review consensus determinism wiring · 51 tool uses · 110.8k tokens │ ⎿ Done ├ Review bridge tx security logic · 24 tool uses · 88.1k tokens │ ⎿ Done └ Verify test and doc claims · 31 tool uses · 52.4k tokens ⎿ Done

● Fable 5's safety measures flagged this message for cybersecurity or biology topics. They may flag safe, normal content as well. These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine them. Switched to Opus 4.8. Send feedback with /feedback or learn more ⎿ Tip: You can configure model switch behavior in /config ⎿  Interrupted · What should Claude do instead? ---

My prompt:

"I want you to enable headless desktop use and browsing inside the <project> VMs so the agents can conduct browsing and general computer use (X or Wayland). There should be a skill that enables the headless desktop / browser use (it's off by default to maintain image size). There should be a skill that kicks off the process or it can be done by talking directly to the agent."

This got immediately downgraded by the classifier without any reason. Pretty disappointing for a class-1 model. I use Claude code. My main driver is Codex and will continue to be.

Biggest task: Codex refactor of the Maturana code base to make it more modular (25 KLOC) and with a plugin arcitecture, using a one-shot /goal.

Not really disappointing but it seems that I have a ceiling and I don't see improvements over opus. But I think it is because of codebase/usecase. I expect that it would have more gain on larger projects. I use it mainly on embedded systems ( 256kB code storage and 64kB RAM) with a C codebase. For example making basic I2C device drivers, I give it the datasheet of the hardware component and ask to first to extract register definition and create a wrapper. I check that then ask for the implementation of function that I need.

The main problems I see is that "document understanding"/"reading PDF" is less powerful when using Cursor/Copilot/... When using for example ChatGPT UI, it can perfectly read the PDF and extract useful information. But on the harnesses you sometimes see that it is going to use python and OCR scripts to extract that information. The quality is way lower than using the web UI.

Currently mainly using codex.

Is this something normal?