Jacob

Yep! It used OBS on Mac. For the setup, it took screenshots on its own for placement. You can get it to put together a fully edited vid as well, but that can take a while. For instance, it can come up with a script/workflow, carry out the screen recording, edit with ffmpeg, generate actions / labels with hyperframes/remotion, use elevenlabs API key for voice, music, and sound effects, and iteratively look at and improve the vid.

@JacobMolBio @thsottiaux Wow 😳 Even your screen recording was handled by gpt5.6sol ?

Or… what if we gave you $100 in Codex credits if you tell us what you love about GPT-5.6 Sol or why you switched? Tweet it, claim your gift, enjoy more usage. First 10k get the free tokens! switch-to-codex.openai.chatgpt.site

@AgentXCrisp Thanks! The Fable used the workflow style from Thariq's article here and rendered the animation via heygen's hyperframes.

@JacobMolBio Cool graphic animations how did you make them?

Since Fable 5 is fantastic at everything, but pricier than other models, a good setup is to have Fable act as a planner or conductor for a project, where other models take on the execution. Many of us also pay for OpenAI's Codex, like Matt, which unlocks GPT-5.5 workers that won't affect your Claude usage limits. Luckily it's straightforward for Claude to drive Codex workers through the Codex CLI, which runs headless and reports back to the Fable conductor. Below is a Claude Code skill that sets Fable up as a conductor of Codex workers, which in turn can spawn their own Codex subagents to tackle tasks. It also routes UI and design tasks to Opus workers, and those Opus leads can spawn their own Codex fleets too. Fable made this vid of example team designs, rendered with HyperFrames.

You can point your agent at this skill repo to have Fable jump into the conductor role faster, without having to set the campaign structure up yourself. If you only have a Claude Code subscription, the skill still works well, with Fable conducting Opus and Sonnet agents instead of Codex agents. This skill also nudges Fable to improve as a conductor on your specific projects by saving learnings as it works. github.com/jvogan/a-fable…

Example setups: Fable gathers the shape of the project -> writes the plan -> spawns parallel Codex workers (own branch each) -> workers report back -> Fable merges (or has a worker do the merging) and sends out the next wave Fable hands a sub-goal to a squad lead -> the lead spawns its own Codex workers -> merges their work -> returns to Fable The lead can be Opus or Codex, with Opus preferred for UI/UX tasks. You can also tell Fable to route work to different models, such as Sonnet 5, for example.

You can also add these skills to Claude Science from your local repos or GitHub (see attached snapshot). Cloud compute resources can help you make the most of these workflows. For example, on the Claude Science credentials page, you can connect services like Boltz, BioHub, Modal, AWS, and more.

These skills are curated on the BioSymphony GitHub page below. These skills also nudge your AI agent to save learnings so your workflows get 'smarter' over time. Some other useful repos there include: Neocloud bridge, for helping agents set up runs, launch cloud compute, and Proteus, for helping agents use structural biology tools like PyMOL and ChimeraX. github.com/BioSymphony

This StructureFactory demo is an agent using SwitchCraft to design a (theoretical) novel cGAMP biosensor. The agent starts from the natural human STING cGAMP switch, measures the apo-to-holo shape change, then asks SwitchCraft to design a new protein that changes shape when cGAMP binds.

Structure Factory is an agentic harness for structural biology workflows. It helps agents set up tasks that stack binder design, structure mapping, cofolding, model comparison, candidate ranking, cloud runs, figures, and presentations. This repo includes learnings for using tools from BioHub, like ESMFold2, as well as Boltz, Genie3, SwitchCraft, RFdiffusion, ProteinMPNN, ChimeraX, PyMOL, and more. github.com/BioSymphony/st…

This CryoCore demo shows the agent moving from cryo-EM accessions into map/model work. The agent checks the map and model for a protein (K10), runs model-building paths, compares the generated model to the deposited structure, and then renders scenes. Background on K10: a 950-residue de novo protein from Frank et al., Science 2024, “Scalable protein design using optimization in a relaxed sequence space,” PDB 9F0L.

CryoCore is an agentic harness that helps agents work through cryo-EM mapping and model tasks. It pairs well with neocloud compute resources like RunPod and Lambda, as cryo-EM tasks can take a lot of storage and compute resources. github.com/BioSymphony/cr…

Small Molecules is an agentic harness that helps design and characterize molecules with computational docking, ADMET, retrosyn routes, and more. github.com/BioSymphony/sm…

This BioProspector demo shows an AI agent piecing together a biological route for Huperzine A (which is of interest for human health due to its effects on acetylcholine).

Ferm DoE (fermentation design of experiment) is an agentic harness for helping with constrained experimental design by using tools like BoFire, BoTorch, ENTMOOT, OMLT, TabPFN, and standard DoE methods. This is useful for helping manage the complexity of capacity, costs, and materials in bioprocess and biomanufacturing projects. github.com/BioSymphony/fe…

BioProspector is an agentic harness for mixing and matching genes for biosynthetic routes to natural compounds. For instance, if there is a natural precursor to an essential medicine that we want a new biological supply chain for, this skill can help your agent find candidate genes for each step in a pathway from many different organisms. This can be useful to help explore new biological supply routes for difficult-to-source natural products that benefit human health. github.com/BioSymphony/bi…

This is a GeneCluster example of an AI agent mining genes involved in berberine biosynthesis. Known pathway enzymes become search anchors. Search hits become candidate genes. Candidate genes get checked against controls and nearby genome context, then curated.

GeneCluster is an agentic harness for genome mining. This helps your agent turn a pathway, target molecule, or a missing biosynthetic step into a genome mining campaign. For example, it can help your agent answer questions like: does this genome contain gene clusters for the pathway that makes compound X? Or, we know about genes A - J for making compound X, but we are missing gene K, so let’s check if there is a candidate gene K in any relevant gene clusters. github.com/BioSymphony/ge…

BioSymphony is a set of skills to help your AI agent tackle complex biological research and planning tasks. These include agentic harnesses for binder design, genome mining, cryo-EM mapping, small molecule characterization, Bayesian optimization of experiments, and more. The BioSymphony skill repos are useful for researchers comfortable with coding agents like Codex, Claude Code, Grok Build, etc. They can also be added to the new Claude Science. For longer-horizon big projects, BioSymphony pairs well with OpenAI’s Symphony, where AI agents can be massively autonomously deployed against Linear tickets. Here are the relevant repos to get started:

@arthur_spirling Claude Science comes with its own Reviewer 2

I got early access to this, and it’s going to be a game changer for science. In my case, it automatically desk rejected my manuscript for “no novel theory”. saved me a bunch of time