Fable went viral, but half the internet says it's useless

Fable 5, the talk of the town!

People think it's either the most incredible thing ever built, or a complete waste of time.

So what's really the deal?

Here's my experience & recommendations after a few days:

#1 - It actually uses less tokens than Opus

This might be a little bit of word-smithing things, but in my experience, Fable uses LESS tokens to get something done than Opus.

It seems to think for much longer, but the output is better with fewer tokens used.

The only caveat there is obviously that the tokens cost 2X as much.

So I'd probably only put it on tasks where you think you'd be 2X more effective, or if there are problems Opus (or Codex or whatever you use) can't actually solve effectively.

#2 - Exceptionally good at one-shotting from specs

For stuff that I care about being good, I only use spec driven tools like OpenSpec.

And if you use Fable to create your specs and implement them… it's really good.

That said, given #1 above, I'd only use it for harder stuff.

#3 - Takes a very long time to complete tasks

A thing I've noticed is that if you have a clear spec in an existing project, it's relatively quick to get things done.

But if you have something net-new without existing patterns, it takes a very long time to complete.

That said, it's usually very accurate once completed.

For example, I kicked it off on tearing down and rebuilding a skill library that I use for finding trending topics without having to actually be on social platforms

And it took 5 hours to move through all the specs

#4 - Really great at cross-checking & integrating complicated asks

A lot of my workflows look roughly like:

• Research and gather a bunch of context
• Review through and find out what's good or not
• Integrate those findings into an “explore what this feature could be” type of workflow
• Make sure it's all cohesive with my existing project
• Gather more context & repeat until it's good to go

And it's really good at understanding what the main “thread” is, what your goal REALLY is, and building a solid plan from there.

For example, I kicked it off recently on forking off a popular repo, and integrating a huge dump of research I had for improvements I wanted to make.

I had it “grill me” exhaustively about my goals and intent based on all of that context, and build a plan from there.

And the output was on the money across all of that convoluted context, a huge repo to sift through, and most importantly all of the questions I answered

#5 - Really great at surfacing contradictions in “lesser” agent tasks

If you wanna save on tokens a bit, you can use Fable as just the plan maker and orchestrator

And in that case, it's very good at finding issues with sub agent decisions and intervening.

So time for the million dollar question:

What should you do given your time with it (it goes away from standard plans June 22nd)?

There's two very concrete things I'm doing right now, and you should too:

#1 - auditing & improving existing skills

Probably one of the biggest immediate value-adds is improving all of your important skills to make them even better.

I've kicked off a lot of “deep research” surrounding the domain of my skills, and then used Fable to integrate those into the skills/plugins to make them better.

And it's been incredibly good at that.

#2 - refactoring complicated pieces of your projects

If you have some critical lynchpin of a project that you know might not be 100% optimal (or even downright shitty), this is the time to refactor the living daylights out of it.

For example, in the topic sourcing project I mentioned earlier, I had it refactor and improve the entire ranking algorithm.

First it fanned out research for best practices on such algorithms, then grilled me, and then implemented the fixes (alongside a bunch of other improvements).

I wouldn't be wasting my time on design exercises or other random one-shotting nonsense.

Smoke em while we got em!

NEWSWORTHY

AGENTS GONNA START BUYING STUFF FOR US

A company called WorkOS (provides auth and other stuff for projects) recently created a system for allowing agents to register for things on your behalf. You can read about it here if you're interested.

You guys know how when you're coding up a storm, those pesky coding agents have the AUDACITY to give us manual checklists of things, like signing up and getting API keys for stuff?

Well, with this type of system, it doesn't have to - it can sign up and create API keys all on its own.

Firecrawl integrated this into their system, “betting on the next 1B+ users being agents” https://x.com/firecrawl/status/2064387335794434281

From there, it's a hop-skip-and-a-jump to being able to buy on your behalf, too.

LOOP ENGINEERING GOING VIRAL

All it takes is 2-3 accounts to make a tweet these days and the AI world goes to battle.

And that's exactly what happened with “Loop Engineering”

I did a video of it on my YouTube that you can check out if you're interested.

One of the best write ups I've seen on it comes from the director of Google Cloud

Worth a read if you're interested in practically understanding more about these Fabled loops (see what I did there?)

If not, here's the TLDR:

Instead of prompting an agent precisely what to do, you build a loop that prompts the agent.

You give a goal, and the loop runs on that goal until it's actually complete.

So if the old way was:

You type something, read the output, then type the next thing

Then the new way is:

You build a loop that finds the work to be done, hands it off to sub agents, reviews it, and then moves on to the next thing.

Worth nothing: this was possible before with stuff like Ralph Loops, but now they're becoming native parts of our coding agents (e.g., /goal from Codex and Claude Code, /loop, etc)

Fair warning though, if you don't already run ~3-5 terminal windows at the same time now doing stuff, you should probably just ignore most of the loops conversation.

WORTH CHECKING OUT

How To Sell Apps People Will Actually Buy: a lot of people can vibe code stuff now that would theoretically sell. But, most people fail as early as the landing page. No amount of viral UGC hacks or cold outbound will matter if you don't follow the principles in this free 90 minute video I posted in the community: https://www.skool.com/tech-snack/how-to-sell-products-people-will-actually-buy

Sub Agents Can Be Nested Now: If you do not show fear in the face of potential runaway token consumption... good news! Your sub agents can now create their own sub agents. Instead of one orchestrator kicking off 3 sub agents, each of those 3 subagents can now kick off up to 5 of their own sub agents. See the news here.

Claude Workflows Are Great: On the topic of loops... something that got swept under the doormat with all the Loop Engineering and Fable 5 noise is Claude Workflows. Did a full video on it here, but it's actually a huge unlock that, to borrow language from my YouTube colleagues "I'm surprised no one is talking about right now". But for real this time! It moves the orchestration of agents out of the agents themselves, meaning your agent isn't allowed to check it's own homework.

Riffrec: One of the creators of my favorite vibe engineering plugin (compound engineering), created this sweet tool. It allows you to record your screen when testing (or meeting with customers), but it actually captures what's going on inside of the app, what you're saying on screen, console logs, etc. He uses this to automagically feed feature requests and bugs into his coding agents. Very relevant for the loop conversation above! Pass the loop the video recording and let it do its thing.

Mainframe: as a vibe coder trying to get better, learn actual Computer Science, and generally try to build stuff that's actually good, one big problem I run into is not being able to visualize how systems connect or what a PR is REALLY touching. This tool helps with that by creating an explainer video of any PR! Pretty cool

Firecrawl Workflows: Firecrawl is one of the best scraping tools out there, and they recently released something called "workflows", which are "outcome focused skills" for completing tasks end to end. Check it out.

NVIDIA'S SkillSpector: Skills are sketchy, an easy way for people to attack your machine! This new skill is a security scanner for AI agent skills. It detects vulnerabilities, malicious patterns, and security risks.

That's it for this week, catch ya next time!

-Sean