The body is lean and assumes Claude's competence — it explains only non-obvious specifics (the load-bearing `cd`, the two-checkpoint limitation, the crystallisable-vs-soluble caveat) and never defines ProteinMPNN or inverse folding, matching the level-3 anchor where every token earns its place.

Provides a complete, copy-paste-ready bash invocation with exact flags plus concrete follow-ups (the `--model_name` error, the LigandMPNN `--model_type soluble_mpnn` path, output at `out/seqs/<stem>.fa`), meeting the level-3 fully-executable anchor.

The single running action is unambiguous with a called-out load-bearing step and explicit downstream validation ("check that the resulting designs still fold via boltz or esmfold2"), so under the simple-skill scoring note the clear sequence with a verification checkpoint scores 3.

No bundle files are present and the body is under 50 lines with well-organized section headers and only one-level-deep peer-skill pointers (proteinmpnn, ligandmpnn, boltz, esmfold2), satisfying the simple-skill exception where well-organized sections alone score 3.

Names multiple concrete actions — "Inverse-fold a backbone with SolubleMPNN" for "sequences biased toward cytosolic expression and reduced aggregation" — matching the level-3 anchor listing specific concrete actions rather than the level-2 partial coverage.

Explicitly answers both what (soluble-biased inverse folding) and when ("Reach for this skill when designs from vanilla ProteinMPNN are aggregating or going to inclusion bodies..."), satisfying the level-3 anchor with explicit triggers.

Includes natural terms a protein engineer would actually say — "aggregating," "inclusion bodies," "soluble expression," "E. coli expression screen" — giving good coverage rather than only some relevant keywords.

Occupies a clear niche (soluble-biased inverse folding) with distinct triggers unlikely to fire for the wrong skill, distinguishing it from vanilla ProteinMPNN and LigandMPNN.