bhgtiger

Use for ColabFold/colabfold_batch: configure hosts, plan or run AlphaFold2(-Multimer), choose MSA/privacy, inspect outputs, or troubleshoot.

Portable, config-first assistant for CryoAtom2 — automatic atomic model building for proteins, RNA, DNA and protein-nucleic-acid complexes from cryo-EM density maps. Carries no host facts: it probes the machine it is running on, writes a site config, and only then makes machine-specific claims. Use whenever the user wants to install, configure, port, understand, plan, or run CryoAtom/CryoAtom2 on any system: standing up a new install (container or native conda), staging and pinning the six model weights, writing a build command, choosing sequence vs no-sequence mode, picking -pf/-nf databases, reading the output mmCIF and its confidence field, comparing against ModelAngelo, or troubleshooting weights/CUDA/OOM/getp errors. It never installs, downloads, or runs anything without explicit per-action confirmation. Triggers: cryoatom, CryoAtom2, cryoatom build, install cryoatom on a new cluster, atomic model building from cryo-EM map, protein-nucleic acid model building, RUNet/CryoNet checkpoints, CryoAtom weight c

Read-only advisor for CryoAtom/CryoAtom2 readiness, static CLI, NOT-RUN plans, and safety limits; use for explicit CryoAtom questions.

Authoritative guide for invoking the `claude` CLI as a subprocess from a host agent — for code review, plan critique, executing approved plans, multi-turn refinement, background/parallel delegation, or any programmatic hand-off to Claude Code. Use this skill whenever you are about to write or modify a call to `claude` (e.g. `claude -p`, `subprocess.run(["claude", ...])`, `Popen`, a Bash pipeline, an n8n/LangGraph/AutoGen/CrewAI node that shells out to Claude, or a Codex/GPT/Gemini orchestrator driving Claude Code). Triggers on phrases like "shell out to claude", "call claude -p", "have Claude review/execute this plan", "hand off to claude code", "spawn claude subprocess", "claude headless", "claude --print", "non-interactive claude", "background claude agent", or whenever a non-Claude agent or orchestrator drives Claude Code. Covers headless invocation, output formats, structured output, session resumption, tool/permission gating, background agents, system-prompt injection, model + cost controls, the dangerou

Authoritative guide for driving the OpenAI Codex CLI (the `codex` binary and `codex exec`) as a subprocess from another agent — for repository analysis, focused implementation, independent code review, debugging, structured JSON extraction, and multi-turn delegation. Use this skill whenever you are about to run or script `codex` — codex exec, codex exec review, codex exec resume, or codex mcp-server — and whenever the user says "use Codex", "ask Codex", "get a second opinion from Codex", "have Codex review this", or "delegate this to Codex". Covers non-interactive invocation, the exact per-subcommand flag positions, the sandbox and project-trust model, JSONL and last-message and JSON-schema output, session resume, exit codes, and verifying Codex's work. Consult BEFORE writing the command because flag placement and sandbox choice change what Codex can do and whether the call parses. Do not invoke for trivial work the agent can do directly, and never launch the bare `codex` TUI from automation.

Read-only advisor, command-planner, and troubleshooter for Namdinator — the automated MDFF (molecular-dynamics flexible fitting) pipeline that fits an already-roughly-docked atomic model into a cryo-EM or crystallographic map (VMD + NAMD2 + Phenix; optional Rosetta), via the local Namdinator_Generic.sh CLI or the namdinator.au.dk web service. Use whenever the user names Namdinator or namdinator.au.dk; asks whether it suits a model/map; wants a Namdinator command or web-form plan; asks what its flags do (-p -m -r -x -l -g -s -i ...); is losing ligands/metals/waters/HETATM in fitting; hits its errors (Bad global bond count, AutoPSF fails, atoms moving too fast, VMD/NAMD2 not found); needs to read last_frame.pdb / CC / clashscore / Ramachandran outputs; or is weighing automated MDFF against ISOLDE/Coot/ Phenix. Also for "should I MDFF-fit this into my map" even when unnamed. It PLANS and EXPLAINS only — never runs Namdinator, never submits the web form, and is not validated on any live runtime.

Config-first, validated assistant for Boltz (jwohlwend/boltz) — biomolecular structure and binding-affinity predictor (Boltz-1/Boltz-2; CLI `boltz predict`). Validated against Boltz v2.2.1 on Linux+NVIDIA (2026-06-23). Use whenever the user wants to install, configure, understand, or run Boltz: writing YAML inputs (protein/DNA/RNA/ligand, MSA, templates, pocket/contact/bond constraints), generating `boltz predict` commands without hallucinating flags, choosing Boltz-2 vs Boltz-1, running structure or ligand-affinity prediction, interpreting outputs (confidence/PAE/pLDDT, affinity_pred_value vs affinity_probability_binary), MSA-server vs custom MSA, or troubleshooting install/CUDA/kernel/OOM/MSA errors. ALWAYS runs a read-only env probe first; on a validated host it emits concrete commands with real paths and, after explicit confirmation, MAY run real Boltz jobs — never installs or runs without confirmation. Triggers: boltz, boltz predict, boltz2, affinity prediction, ColabFold MSA, use_msa_server.

Install and set up ModelAngelo (3dem/model-angelo), the cryo-EM atomic model builder, on a Linux/NVIDIA target. Config-first: it probes the target, picks an install route (personal conda, shared-cluster TORCH_HOME, container, SBGrid, or an HPC module like Biowulf), runs the official install_script.sh with confirmation, plans the ~10 GB weight + ESM download and TORCH_HOME cache, and verifies the install. Use whenever the user wants to install, set up, or configure the ModelAngelo environment, asks whether a machine can run it, hits an install / conda / torch / CUDA / weights / TORCH_HOME / hhblits error, or needs ModelAngelo wired into RELION 5. It assumes nothing about the current machine; configuration is captured per target first. It installs and verifies but does NOT run production builds and is NOT a validation tool. Triggers: ModelAngelo, model_angelo, install ModelAngelo, setup_weights, TORCH_HOME, ModelAngelo GPU/CUDA error, on Biowulf/SBGrid/Singularity, RELION ModuleNotFoundError.

Design, audit, or critique a closed-loop LLM agent — the propose → check → measure → feedback → refine pattern. Use whenever someone wants to add a self-critique / reflection / verifier / best-of-N / debate / self-consistency step, asks whether a refine loop will actually help, wants to optimize or debug a loop that underperforms or overspends, or asks "is this an antipattern." Triggers on phrases like "should I add a critic", "build a feedback loop for my agent", "why isn't my self-refine improving", "audit my agent loop", "best- of-N / self-consistency / reflexion / debate design", "is looping worth it here", or "what's wrong with my agent loop." Entry point is the verifier-edge decision flow; routes to a design guide, an audit checklist, and an antipatterns catalog backed by 28 digested sources.