Fable 5 Rules Anthropic Doesn’t Want You to know

AAI LABS
Computing/SoftwareManagementInternet Technology

Transcript

00:00:00you already know that soon enough fable 5 is going to be taken off our plans but you shouldn't be
00:00:04worried about that anthropic already posted that they plan to bring it back so it's certainly not
00:00:09going away forever now when fable came out anthropic put out a prompting guide and you
00:00:13might think that those are the only rules for using the model but they're not that guide is
00:00:17just for using the model in an app it doesn't tell you how to actually use it in your own workflow
00:00:22you already know that we're a software company and due to that we've been running this model for
00:00:27days now both in our own products and in the ones that we've been building for our clients and from
00:00:32that we've gathered certain rules that you can implement for example what settings should you
00:00:36actually use to get the most out of the model then there's a free tool that you can use to reduce your
00:00:41fable usage but in order to use it you need to make sure that your coding project is set up in the right
00:00:46way we do know that we're a bit late but by the end of this video you'll clearly understand how you need
00:00:51to start using this model so the first thing you need to change is the effort setting and this one
00:00:56alone will cut your usage in half without decreasing the quality at all now you already know effort is
00:01:02just how hard the model thinks before it answers you but what you probably haven't thought about
00:01:06is that on fable 5 turning it up is actually a huge trap here's what you probably do you see the max in
00:01:11the menu and you think that it's the smartest model so it should work as hard as it can and you crank
00:01:16it all the way up but this changes nothing for the output that the model gives you it gives out the
00:01:20same output on x high and max as it does on high this was honestly surprising to us and it's not
00:01:25just us other people have experienced this as well if you push it to higher settings it's just going
00:01:30to get brutal on your usage and ultra code is just a clawed code skill that has been attached on top of
00:01:35the effort level which we honestly think hasn't really produced anything worthwhile ever it's just
00:01:40one of the ways that anthropic gets you to spend more on their product so keep the effort between low and
00:01:45high and never go past it and honestly you'd be surprised at how well the low and medium effort
00:01:50levels perform on fable now what happens whenever a huge model and a smaller but smarter model come
00:01:55out one of the most common patterns that you see throughout the ai space is that people start
00:01:59saying you should use the larger model for planning and the smaller model for carrying out the actual
00:02:04work and this happens every single time when sonnet 5 came out the new trending lesson became that you
00:02:10should use sonnet 5 for all the smaller tasks that are not worth the opus model because
00:02:15fable hadn't come out at that time still it's not bad advice for example if you've built a second
00:02:20brain you already know that if you want to ingest documents or read huge pdfs you shouldn't be using
00:02:25large models for that this is what we do in the ai lab second brain when we want to ingest competitor
00:02:30content for research and even though you might have heard it a lot of times we've still found some
00:02:35tools that might help you out in using multiple models in your workflow if we talk about the workflows
00:02:40that you can use then there's the most simple one in which you plan out with one larger model
00:02:44and then ask the smaller model to implement it for example in our community project you can see that
00:02:49we have the docs folder and inside that we have a subfolder which we always use while implementing a
00:02:54new feature and whenever we come up with a new feature it always has a prd and this prd is based
00:03:00on this template whenever we create a new prd we use fable or other larger models and implement it with
00:03:05smaller models if the new feature isn't too big now you might be thinking can you copy the same template as
00:03:11us you can't really copy it because it's specifically tailored to this project where whenever we make a
00:03:16new feature it always asks how it's going to affect different parts of the project for example are there
00:03:21going to be any new users created in the project and then the data model and migration means do we
00:03:26actually need to change the database behind the app or not also this custom community is where you'll find
00:03:31all the skills design systems and things that we build and show you inside our videos so if you found
00:03:37value in our videos and want to support the channel this is the best way to do it the link is going
00:03:42to be in the description there's another important part of working with agents which is spinning up
00:03:47multiple sub-agents and for that theo here actually gave a relevant section of his claude.md which makes
00:03:53a lot of sense for these sub-agents he has said that for running a task via sub-agents the main
00:03:58model orchestrating the session should have a model graph like this where they should be rated according
00:04:03to their cost intelligence and taste although we haven't really implemented this in our own
00:04:07workflows using the claude.md but we do manually route through these different models and the
00:04:12ranking that theo has done is pretty much accurate here so if the main session wants to use claude
00:04:18models it can obviously spin up sub-agents using that model but how does it use the codex models theo
00:04:24has written that the codex models are only reachable through the cli but for this specific
00:04:28purpose open ai has actually made a plugin for claude code which you can install with this
00:04:33claude code slash command and although it does require having the codex cli it does give you a
00:04:39handful of commands that you can go through if you like these codex models we ourselves have started
00:04:44using codex a lot as well but we use them for review work because we think that the codex models are way
00:04:49better at reviewing what other models have done but we haven't really compared them to fable 5 yet
00:04:54now even though we just told you that you should be using the larger models for planning there's a
00:04:59particular thing we've seen while working with fable 5 this model is way better at finding creative
00:05:04solutions to hard problems so occasionally we've also enjoyed using it directly in our coding tasks
00:05:09as well but that still doesn't solve the original problem which is that it costs too much to solve
00:05:14this you can actually use ponytail which is a set of rules that basically makes your ai coding agent
00:05:20write less code and it's not doing anything new it's just a list of old coding rules that makes your
00:05:25models write less code but if you just go ahead and use it to write less code chances are that your
00:05:30app which already has users might break so you need a certain structure inside your app so that you
00:05:35can start using this tool now you might already know the concept of tdd which means test driven
00:05:40development where tests are written first and then the agent writes code as you already saw we make prds
00:05:46for every new feature and the tests are written based on those prds but before we move to how we
00:05:51use it let's have a word by our sponsor veed if you've ever tried creating videos with ai you know
00:05:56the biggest problem is that every model lives on a different platform and you end up juggling five
00:06:01subscriptions just to test what works veed is an all-in-one ai-powered video creation platform that puts
00:06:07every major ai video generator in one place their gen ai studio lets you type a simple text prompt and
00:06:13get a full social media video instantly script visuals voiceover subtitles all done they even
00:06:18have their own model called veed fabric 1.0 that turns a single image into a realistic talking video
00:06:25up to five minutes long and if you want to compare different ai video generators side by side their ai
00:06:30playground gives you access to google vo cling c dance and more without leaving the platform plus
00:06:35there's a full video editor with dynamic subtitles background removal clean audio and auto resizing built in
00:06:42sign up with our code in the description for 30 off on your first month for example we have the
00:06:47community dashboard here and if i go back you can see that alongside the files having the dashboard
00:06:52code we also have tests for the dashboard as well so whenever anything is changed in the app we run
00:06:57these tests these tests make sure that whatever the dashboard is supposed to do it keeps doing
00:07:02that for example if i click on the settings icon it is supposed to open up this menu a test file tests
00:07:08that this still happens if the agent writes some code that breaks the button and doesn't show you the
00:07:13menu that means the original code has been broken so this is why these tests which are called unit tests
00:07:18exist because they test a single unit which in this case is the dashboard so just like the dashboard the
00:07:24whole app has tests written for it so whenever we are going to make a new feature new tests are written
00:07:29first and those tests are not written by the same agent those tests are written by the tdd test author now it is
00:07:35important that another sub-agent writes the tests because the existing agent will often write tests
00:07:40that are supposed to pass its own faulty code so because our repo has been structured in this way
00:07:45and we always write tests first no matter how the model writes code we always know that it will work
00:07:50also if you want to use this tdd test author agent in your own projects you can get it inside our ai
00:07:56labs pro community as well since ponytail is a plugin in claude code it comes with skills and other stuff as
00:08:02well but you only need the prompt in its skill.md and then you can put that in your main claude.md or
00:08:08paste it as a prompt before you start using fable 5 for coding now you might be thinking how do fable
00:08:135 and ponytail come together so if you are not familiar with the concept of refactoring it basically means
00:08:18that you modify the code so that there is less of it and ponytail is really good in refactoring code
00:08:23since we actually have these tests written whatever it writes using ponytail i always trust that it will work
00:08:29because those tests are passing and in this way you can build apps and constantly keep changing code
00:08:34before we continue to the other rules there's one thing we wanted to clear up there's this one
00:08:38benchmark going around that basically says that fable got dumber and that it can't really code anymore
00:08:44because it will route to opus when it is asked to code and honestly it's just not true it's only this
00:08:49one benchmark and we honestly don't trust it because it ranks sonnet 5 as the number one model in
00:08:54reasoning which is just not true we've been using this model all this time and we haven't seen any
00:08:59rerouting encoding tasks other than when we ask it to explain its reasoning so you shouldn't really
00:09:04get paranoid about this at all we've also seen what other people have been doing with this model and one
00:09:10of the most important things they've forgotten is the reason anthropic released it in the first place
00:09:15they said mythos was too powerful and based on the testing it could find security risks that no other
00:09:20model had which means it could jailbreak those as well so take out your code review agents and your
00:09:25security review agents and specifically focus on security so for example we're currently working
00:09:31on adding new features in the community and a major redesign but in here we basically use some sub
00:09:36agents as part of our verification loop while we're actually building the platform so for example one of
00:09:41them is the code review expert which is basically a sub agent built on top of the skill from cursor
00:09:46called the thermonuclear code quality review and it's a really great skill that actually checks the
00:09:51architecture of the code and make sure that the code base is going in the right direction when we're
00:09:56constantly adding these new features using our agents and you can see that it went ahead and came
00:10:01back with a small change that we need to implement in the app and two other minor ones i'm not going to
00:10:06show these changes to you but we have patched them in and then another sub agent that we run is the
00:10:10security reviewer sub agent which is actually based on this github action called the clawed code security
00:10:16review a github action is an autonomous review that runs when you upload your code to github and you can
00:10:22actually give the link of this repo to clawed code and just ask it to convert it into an agent and it's
00:10:27going to do that now this was a security review of the new feature that we added and this time it said
00:10:32that it came back clean and that there was nothing that needed to be changed but when we actually ran it
00:10:37on the whole platform that was set up it spun up six of these security reviewer agents all based on
00:10:42fable 5 and they were running in parallel across the whole platform which was divided into multiple
00:10:47parts and this proved much more fruitful it came out with a lot of bugs on the platform which we have
00:10:52now successfully patched so you should pretty much focus on the security and code base reviews which is
00:10:57what anthropic had actually warned and advertised it for so recently tarik who is from anthropic also
00:11:03published a field guide to fable in which he described his own experience while working with fable 5 and
00:11:09what he thinks about it in my opinion you only need to focus on some of the things here he has actually
00:11:14made some example artifacts which explain things much better than the article in the artifact he has
00:11:19divided his process into multiple steps i think that you are already implementing these processes with
00:11:25the old models so you do not really need to dive deep into it but one thing which you should start
00:11:29doing if you are not already doing it is first making mockups of the features that you are going
00:11:34to implement so if we look inside our repo again you can see that we have this design folder and inside
00:11:39that design folder we have a mocks folder now whenever we implement a new feature that feature is
00:11:44first added on to this html mock-up this html mock-up is a one-to-one prototype clone of the whole
00:11:50website and we basically add the new feature on top of this before we actually start to work because
00:11:55it is important that you visualize your changes but other than that one thing that i liked in this
00:12:00article was quizzing before merging so you already know that while working on coding you basically
00:12:05create a git branch make the changes in that and then you merge now this entirely depends on the
00:12:10person working on the ai agents but some people do not really look at what the agent has implemented
00:12:15and i am not saying reading every line of the code but just reading a small summary of how the agent did
00:12:20what it did and that is actually really important so the prompt that he has given make sure that before
00:12:25you merge fable actually comes to you and quizzes you on these changes and this practice basically makes
00:12:30sure that you know what is going on in your project and i highly recommend this over time we have worked
00:12:35with a lot of projects including our custom community and custom automations to make our own process
00:12:40better and multiple other apps for clients as well and we have seen that if you do not really have
00:12:45context on what is happening in your code base you will not be able to steer your agent no matter what
00:12:50you do whether you use a goal or you use a loop it will have to end and you will have to add new
00:12:56features or make changes to your project and not knowing what is in your project will come back
00:13:00to bite you at that time that brings us to the end of this video if you'd like to support the channel
00:13:05and help us keep making videos like this you can do so by using the super thanks button below as always
00:13:11Thank you for watching and I'll see you in the next one.

Key Takeaway

Optimizing Fable 5 workflows requires capping effort levels at medium, automating security and code reviews with parallel sub-agents, and enforcing a Test-Driven Development (TDD) structure to ensure code stability.

Highlights

  • Setting Fable 5 effort levels to 'low' or 'medium' matches the output quality of 'high' or 'max' settings while significantly reducing usage costs.

  • Using a larger model for planning and a smaller model for execution improves workflow efficiency in software development.

  • The 'ponytail' toolset provides a structured approach for AI coding agents to refactor and write less code while maintaining application stability.

  • Implementing Test-Driven Development (TDD) with a dedicated sub-agent for writing tests prevents the primary coding agent from generating faulty code that passes its own biased tests.

  • Fable 5 performs best when tasked with security and code-base reviews, as it can identify complex security vulnerabilities that other models miss.

Timeline

Optimizing Model Settings and Effort

  • Setting Fable 5 effort levels above 'high' yields no improvement in output quality.
  • Using effort settings between 'low' and 'medium' significantly reduces usage costs without sacrificing performance.

Anthropic's default prompting guide is intended for app usage rather than custom workflows. Increasing the effort setting to 'max' is a common trap that increases costs while producing identical output to 'high' settings. 'Ultra code' features added to effort levels provide no measurable value.

Agentic Workflows and Planning

  • Plan tasks with larger models and execute them using smaller, faster models.
  • Use custom templates, such as Product Requirements Documents (PRDs), to maintain project context during feature implementation.

Standard AI workflow patterns involve using large models for high-level reasoning and smaller models for implementation. When building features, creating a PRD based on a project-specific template helps the agent understand the impact on data models and migrations. While large models are better at creative problem-solving, their high cost necessitates a balance in task distribution.

Coding Stability and TDD Implementation

  • Integrate 'ponytail' rules to force AI agents to write more concise, refactored code.
  • Use a separate sub-agent for Test-Driven Development (TDD) to prevent the coding agent from creating tests that validate its own faulty code.

Refactoring code to be more concise is risky without a safety net, so implementing unit tests is essential. These tests ensure that existing functionality remains intact when an agent introduces new features. Assigning a dedicated sub-agent to write tests first prevents the primary coding agent from writing biased tests that overlook errors.

Security, Review, and Verification

  • Utilize Fable 5's strength in identifying complex security vulnerabilities by deploying parallel security reviewer sub-agents.
  • Require agents to quiz developers on implementation summaries before merging code branches.

Fable 5 is uniquely capable of finding security risks that other models overlook. Running multiple security reviewer agents in parallel across a platform provides deeper insights than individual manual checks. Before merging, developers should have the agent provide a summary and quiz them on the changes to ensure thorough understanding of the codebase.

Community Posts

View all posts