Boltz Skill

The progressive-disclosure recipe for Boltz-2, the open-weights diffusion co-folder that takes protein, DNA, RNA, and ligand chains and returns mmCIF plus confidences, with an optional small-molecule affinity head. The recipe tells the agent how to write the input YAML, drive the external boltz package, and read the structures and confidence files that come back. Boltz itself is not vendored here. Of the four co-fold Skills this one is the default for binder-validation campaigns: fully open MIT weights, and the fastest sampler of the set.

The package, the weights, a GPU, and any optional MSA service all have to be available separately. Boltz-2 is the one co-folder here that hands back a number in experimental units — affinity_pred_value is log10(IC50 in μM) — and that is exactly what makes it so easy to quote as though an assay had produced it. It orders candidates against each other; it does not measure them. A passing ipTM is the same kind of claim: the model placed the interface consistently. Neither number is evidence that the two chains ever meet.

Install

A Skill is a directory of files, so installing one is copying that directory somewhere an agent looks. With Node 18+ and git on PATH (npm resolves a github: spec through it), and nothing cloned:

npx github:PKU-YuanGroup/OpenAI4S install boltz --target claude

--target claude writes to ~/.claude/skills, claude-project to ./.claude/skills, openai4s to <data_dir>/user-skills, and --dir <path> to anywhere you name. The resolved absolute path is printed before anything is written there, and --dry-run stops at that plan. A reinstall refuses to overwrite a copy you have edited or one it did not install, and uninstall removes only the files it wrote. For a recipe included in the npm release, npx @pku-yuangroup/openai4s-skills install boltz --target claude installs the published copy. The npm catalog can differ from this repository.

Without Node, take the directory itself — and turn it into a .zip if an upload field wants one. The tarball is the whole repository (over 100 MB), and the pipe is a POSIX shell recipe (macOS, Linux, WSL) that extracts only this directory:

curl -L https://codeload.github.com/PKU-YuanGroup/OpenAI4S/tar.gz/refs/heads/main \
  | tar -xz --strip-components=2 OpenAI4S-main/skills/boltz
python3 -m zipfile -c boltz.zip boltz

The click-through form of that same download is the repository zip, which is also the route to take on Windows: extract it and copy skills/boltz/ out (unzip main.zip 'OpenAI4S-main/skills/boltz/*' unpacks only that directory). If you already run OpenAI4S there is nothing to install — the wheel ships every bundled Skill, and a bundled Skill takes precedence over a same-named copy in <data_dir>/user-skills. Targets, provenance, and what the installer refuses to do: tools/skills-installer/.

Files

File Responsibility

SKILL.md Opens on the YAML entity blocks and the boltz predict flags, then gives most of its length to the affinity head, which no sibling co-folder has: a properties: block names one ligand chain as the binder, affinity_pred_value comes back as log10(IC50 in μM), and affinity_probability_binary is the score to rank hits by. One affinity ligand per input, capped at 128 atoms on Boltz v2.2.x, and a FASTA input cannot request it at all. Two traps get their own sections: msa: empty buys accuracy loss and no VRAM back, because the MSA search is CPU-side, and the --no_kernels fallback runs the reference PyTorch path, which the SKILL puts at roughly 2× slower and treats as correct rather than degraded — an unblock for a one-off, not a campaign. Nothing in this repo benchmarks that path or checks it against the fused kernels, so read it as guidance rather than as a measured claim of identical output. Then the confidence JSON, an error table, and licensing.

File	Responsibility
`SKILL.md`	Opens on the YAML entity blocks and the `boltz predict` flags, then gives most of its length to the affinity head, which no sibling co-folder has: a `properties:` block names one ligand chain as the binder, `affinity_pred_value` comes back as log10(IC50 in μM), and `affinity_probability_binary` is the score to rank hits by. One affinity ligand per input, capped at 128 atoms on Boltz v2.2.x, and a FASTA input cannot request it at all. Two traps get their own sections: `msa: empty` buys accuracy loss and no VRAM back, because the MSA search is CPU-side, and the `--no_kernels` fallback runs the reference PyTorch path, which the SKILL puts at roughly 2× slower and treats as correct rather than degraded — an unblock for a one-off, not a campaign. Nothing in this repo benchmarks that path or checks it against the fused kernels, so read it as guidance rather than as a measured claim of identical output. Then the confidence JSON, an error table, and licensing.