Sam Tucker

Helping people not make the mistakes I have running an e-commerce business across 3 continents.

Father of 3 children under 4 😴

[Sam Tucker @scanningwithsam]

This safety check approach to check what you are doing from open ai when using GPT-5.6 Sol seems a more pragmatic approach than Anthropic with Fable 5.

Well done.

[Sam Tucker @scanningwithsam]

Learning the "business of business" is something I've never quite fully grasped.

I am very early into my entrepreneurship journey (under 5 years). So take this with a pinch of salt if you are beyond me in experience.

I look at my 'stage' as if I have started reception at primary school.

Below is one thing I have learned recently.

Maybe its helpful to people on the their entrepreneurship journey...

--

'Finding the Superstars'

When you employ someone you never know how effective and valuable they will be.

You need to give them space to find that out.

A few things to try with new starters (or current):

[1] - Ask them to complete tasks from as many different parts of the business

[2] - Assess how they perform and (this is the important one) how THEY like this part of the business.

[3] - Allow them autonomy to make decisions and take ownership of tasks.

[4] - Provide them space to hit failures/roadblocks and get things wrong.

[5] - Encourage mistakes, its where we learn the most.

When you provide this framework to people, I have found you will see certain people grow and develop at a astonishing rate.

You will quickly sense that "ok this area is where they flourish in".

That is your queue to let them take ownership of this part of the business.

Your role at as leader is to establish where they fit best and give them even MORE autonomy in that area.

Then, sometimes, magic happens.

They excel so far in that part of the business that they become what I would call a 'superstar'.
(stolen term from @AlexHormozi).

These superstars are the people you want to hold onto as long as you can.

Which, yes, means evaluating pay, benefits etc to match the value they bring.

The percieved shift in your head can become a misinterpretation of a power imbalance (due to ego).

You presume they have a better understanding of that part of the business than you, this may understandably worry you.

That because they probably do.

Remind yourself this is ok. Embrace it.

Power as an entrepreneur is allowing people take ownership give them space to try and become superstars.

Find the superstars.
Have fun with it.
Try and let go of pride.

- A lesson to myself.

[Sam Tucker @scanningwithsam]

Been working hard with Fable 5 on Muxary, my little control centre workspace for my Al agents.

The right click > open html in a new pane is now working and honestly - super useful.

sorry for the notepad in the screenshot, trying to cover up my email address.

Hope to get this out to people open source super soon.

:)

[Sam Tucker @scanningwithsam]

Wild.

[Vadims @vadi_ms]

This is me talking to my computer without making a sound.

After just a month of collecting data, our model is already approaching dictation in accuracy. We were surprised to see that it generalizes to unseen participants as well!

(1/n)

[Sam Tucker @scanningwithsam]

I'm late to really using local ai models

Currently using it for a "gated" project where data is sensitive and cant be shared to cloud servers.

There is something quite different about the chat back and forth, knowing that the data is not leaving your machine.

Surprising how good a local model can be. I'm using Qwen3.6-35B-A3B and it is very very good.

Will be doing more of this, for sure.

[Sam Tucker @scanningwithsam]

@iluminatibot Passive income

[Sam Tucker @scanningwithsam]

Did that to me too, halfway through.

However I’ve not found it often still completes the task.

Could be due to subtask delegation - I use Fable to plan and specifically ask it to delegate to other Claude/codex models.

Perhaps the delegation had already began and the reclassification didn’t affect the process, as the model running that task (opus 4.8) wouldn’t be affected by the reclassification.

If that makes any sense?

Worth trying that (unless you already do)

[Sam Tucker @scanningwithsam]

Fable 5 is unbelievable.

I threw an internal problem at it, I thought too hard to automate. Without hyping into boring detail it included: Reverse engineering internal APIs for a process without an api + organising a huge amount of unstructured invoices from suppliers into a coherent process.

It nailed it.

And honestly I think I didn’t even touch the sides of what it is truly capable of.

It was probably just strolling and getting a coffee on the way.

What caused me to try?

After watching @natebjones video - https://t.co/Eazl4YXnoL.

Give it a watch, it’s only 4 mins and will get you reframing what is possible with Fable.

Bottom line of what’s possible - pretty much anything legal.

[Sam Tucker @scanningwithsam]

Decent Fable 5 token streamlining approach seems to be:

1. Fable - plan/ research and output a detailed plan.

2. Codex 5.5 (select reasoning based on task) to implement plan. Or /goal <paste your implementation plan>

If you are particularly paranoid - you can do a final step to ask Fable 5 to review the Codex work/code.

[Sam Tucker @scanningwithsam]

After testing in the wild. This works.

This model routing approach makes token use far more economical + you still get the “Fable Brain” for planning, research etc.

Doesn’t burn your tokens anyway near as fast as just staying on Fable 5 for everything.

Worth a try if you can.

[Sam Tucker @scanningwithsam]

This is a my "control centre" as an ongoing attempt to manage agents across mutiple workspaces/repos.

- Custom web brower (agents can drive claude design via custom cli), plus you can just use the web brower and stay in the same workspace.

- Compose box with quick prompt buttons. (see the pi coding terminal). Main aim was so you can write a prompt and use enter for new line and shift enter to send the prompt. The terminal box is so hard to edit prompts in properly.

- Readable CLI/Terminal mode (seemingly very hard/potential skill issue)

- API locker (so you can keep your api keys in one place and copy them into the .env at will)

- Explorer mode so you can see the repo files and edit .md files etc.

- 1Password CLI - to try and make log in on web brower easier.

- HTML rendering (for progress.html), text editor (for long prompt responses)

- "Bridge function" so agents can talk to each other across panes.

I PROMISE I will open source this soon once I get it stable and secure.

However....

If you have built something like this where you run coding agents in terminals/cli and are looking how to route models for the best use cases, Fable 5 unearthed some interesting things:

My goal was to have Fable 5 as the brains of the operation and route to:

- Sonnet 5
- Opus 4.8
- Codex 5.5
- GLM 5.2 (via pi)
(all with varying thinking levels)

In order to complete tasks based on task type difficulty and reduce Fable token burn, so I can use it for what it does best.

Bottom line: Fable 5 seemed to be quite impressed with GLM 5.2 + some other interesting.

See reply for full details.

[Sam Tucker @scanningwithsam]

Fable 5 - on model routing:

-CODEX-

Set Codex to Full Access via /permissions keystrokes" — never needed. Codex takes it as launch flags: codex -s danger-full-access -a never -codel_reasoning_effort="xhigh" pre-configures the interactive pane; the same config keys drive headless codex exec (the wrapper already did this).

-THIKING LEVELS-

Thinking levels are the same story for all three harnesses: pi --thinking off…xhigh, claude --effort low…max. Keystroke automation of TUIs is the brittle path; I've encoded "launch flags, never keystrokes" in the skill.

-GLM5.2 vs OPUS/SONNET on FRONTEND-
- "Opus/Sonnet for frontend" — the research disagrees. GLM 5.2 is #2 on frontend human-preference boards, ahead of Opus-4.8-thinking, at ~1/5 the cost. Sonnet 5 drifts to a "house style" on open briefs — it needs concrete visual specs.

So frontend routes to GLM-with-verification or Sonnet-with-specs, Opus only for taste-critical polish.

- SECURITY MODEL-

- "Codex for security, Fable audits" — half-confirmed and strengthened:

Sonnet 5's cyber capability is deliberately capped, so the skill now hard-forbids routing security review there.

GLM won one security benchmark but showed reward-hacking under weak oversight — it gets cheap first-pass scans only, and is banned from verification and long-horizon unattended runs.

- LETTING FABEL SELECT ITS OWN THINKING LEVEL-

- “The mode needs to be told to me” — solved without a round trip.

Skills can read the live session effort via ${CLAUDE_EFFORT}, so /fable-boss now opens with a Step-0 check: it classifies the task, and if your current /effort doesn’t fit, tells you in one line what to set (ultracode for audit/migration-scale work, which also authorizes it to use Workflow orchestration).

[Sam Tucker @scanningwithsam]

Anyone else’s Claude Code limits just reset?

A nice touch, especially as his highness Fable 5 is back in action.

12 hours to burn this bad boy down to 0%

[Sam Tucker @scanningwithsam]

@AlexFinn Also just bought a 5090 for running local models in house at my company🫡

I’m in the UK and we had to import DDR4 (DDR5 was nearly triple) RAM from the USA.

Out of curiosity was RAM hard to get hold of out there/mega expensive?

[Sam Tucker @scanningwithsam]

Just checked UK Apple store prices out of curiosity, to see how things compare.

- MacBook Pro 16” M5 Max 40 core, 128gb Ram, 4tb. Nano display.

I bought one back in early April 26 = £6248.

Same spec one now, late June 26 = £8248

£2k increase in just just over 3 months.

Wild.

Presumably due to RAM price increase.

Genuinely worrying to think about where the ceiling on compute will go.

[Sam Tucker @scanningwithsam]

@DearS_o_n Exercise - go on a run until you absolutely cannot run anymore.

Eat as healthy as possible.

Stop drinking (if you do).

Avoid introspective/depressing music.

[Sam Tucker @scanningwithsam]

@theo Stay strong 💪

[Sam Tucker @scanningwithsam]

now available on huggingface

https://t.co/1SjW3QNSf9

[Sam Tucker @scanningwithsam]

Highly rate GLM 5.2 by @Zai_org

It’s a good model and open source. I don’t believe you can yet download the model. Presumable won’t be long.

Not available via API yet, however been using on the coding plan and quite like it.

I tend to use it via OpenCode and it behaves well.

Good option if Claude/codex goes down for a period of time

Good back up to have in your locker

[Sam Tucker @scanningwithsam]

https://t.co/7kHg0v5YLG

https://t.co/nTQzuPYeav

[Sam Tucker @scanningwithsam]

Soma FM Drone Zone music + Broadcastify Top 50 (pick any)

Perfect for vibe coding work.

Trust me.