Conciseness

The body is dense expert material — tables, exact commands, and failure signatures — with no explanation of concepts Claude already knows. Minor trims remain: "Never hand-edit anything under generated/" is restated in the purpose, the Phase 5 hard rule, Common Pitfalls, and the reference header, and the "A text-only dense GPU model therefore reads this file plus the examples…" prose could be tightened. Anchor 4 (efficient, minor over-explanation), not 5 (those repetitions don't earn their place) and not 3 (nothing is padding or generic).

Actionability

Fully executable throughout: copy-paste commands ("patchgen veomni.models.transformers.<m>.<m>_gpu_patch_gen_config -o veomni/models/transformers/<m>/generated --diff", "uv sync --frozen --extra gpu --group dev"), code skeletons for every wiring pattern, the patch-primitives API table, and exact failure signatures ("F821 / F811", "IndentationError: unexpected indent", "0 selected / N deselected") with the concrete commands to recover. Matches anchor 5.

Workflow Clarity

Phases 0–8 are sequenced with a suggested plan, per-phase **Validation** checkpoints, Phase 7 acceptance criteria, and explicit feedback loops — "If not `5.16.1`, re-sync the default env", the "patchgen --check --fix" resync loop instead of ping-ponging with ruff, and "do NOT assume the test doesn't exist; re-check with --collect-only". Clear sequence with explicit validation and error recovery, matching anchor 5.

Progressive Disclosure

The spine/reference split is explicit and quantified: the "What to read for your model" routing table points to references/model-examples.md, references/moe.md, and references/multimodal.md — all real files, one level deep, loaded per category — and states a text-only dense model "skips about 380 lines of MoE and multimodal material". Category-specific detail is correctly offloaded rather than inlined; matches anchor 5.

Specificity

Enumerates multiple concrete actions — "Author or refresh a VeOmni model's patchgen-generated modeling under generated/ — GPU and/or NPU config, dense or MoE, text / VLM / Omni. Covers the patchgen decorators, sharing patches across sibling models via name_map, MoE fused-expert weight loading, Ulysses SP in multimodal forwards, __init__.py registration, running codegen, and the test cases" — comprehensive coverage with no vague filler. Clearly anchor 5, not 4 (no gaps in the action list) and not 3 (far more than 1–2 actions).

Completeness

Explicitly answers both: what ("Author or refresh… patchgen-generated modeling… covers the patchgen decorators… name_map… MoE fused-expert weight loading… __init__.py registration, running codegen, and the test cases") and when (the "Trigger:" clause with concrete phrases), plus a scope boundary ("This is the modeling step of adding a new model, not only of refreshing an existing one"). Matches the anchor-5 example structure exactly.

Trigger Term Quality

An explicit, natural trigger list — "Trigger: 'add patchgen for a model', 'write a patch_gen_config', 'regenerate the generated modeling', 'add NPU patchgen', 'port a model to patchgen', 'transformers v5 migration'" — covering synonyms and the migration refresh case a user would actually say. Not 4: no common variation is missing for this niche.

Distinctiveness Conflict Risk

Clear niche with unmistakable triggers — "VeOmni", "patchgen", "'write a patch_gen_config'", "'add NPU patchgen'" — that no sibling skill would claim. Only "transformers v5 migration" is even plausibly shared, and it legitimately belongs here per the pin-bump survey phase; conflict risk is minimal, fitting anchor 5 rather than anchor 4's "minor overlap risk".