An Agent Skill that derives a structural checklist for any document genre by comparatively analyzing a set of strong exemplar documents — instead of relying on convention, memory, or a generic template, it builds the standard from what real documents in that genre actually do.
Give it a target document (a draft, or a plan for one) and a handful of strong exemplars in the same or an adjacent genre — papers, grant proposals, incident postmortems, product one-pagers, long-form posts, whatever the target is. It walks through a repeatable method:
- Selects and logs why each exemplar was chosen (nearest genre match, nearest topical neighbor, register/venue diversity, or added specifically to fill a data point the others miss).
- Reads every exemplar in full at the structure level — noting where each structural move happens and how, anchored to a specific passage, never to memory of "how these usually go."
- Builds a comparison table per structural dimension (opening architecture, evidence engagement, gap justification, differentiation from nearest neighbors, proportions, closing architecture, or whatever dimensions matter for the genre).
- Reconciles the table into governing-pattern findings — distinguishing conventions that hold universally from ones that track sub-genre or venue, and naming which is which.
- Produces a short, numbered, mechanically verifiable checklist — each item checkable by inspection alone, no re-derivation required — plus, optionally, a paragraph-by-paragraph template for drafting the target document itself.
- Applies the checklist to the target draft, if one exists, producing a ranked list of the biggest structural deviations and why each one matters.
The output is evidence-based and falsifiable: every claim in the governing-pattern report traces back to a specific passage in a specific exemplar, and every checklist item can be marked pass/fail by a reviewer holding only the target document.
This is an Agent Skill following the portable SKILL.md convention —
a folder with a SKILL.md file (YAML frontmatter + instructions) that any compatible agent runtime
can load.
- Claude Code: copy this folder into
.claude/skills/structural-checklist-from-exemplars/in your project (or the user-level skills directory for a global install). - Any other runtime that supports the
SKILL.mdconvention: copy this folder into that runtime's equivalent skills directory. No build step, no dependencies — it's instructions only.
your-project/.claude/skills/structural-checklist-from-exemplars/
├── SKILL.md
├── README.md
├── LICENSE
├── references/
│ └── comparison-table-template.md
└── examples/
└── synthetic-postmortem-example.md
examples/synthetic-postmortem-example.md walks through a small, fully synthetic run of the method
end to end: five invented incident-postmortem "exemplars," one structural dimension worked through
in full (how each states its root cause), and an excerpt of the resulting checklist. It shows the
shape of the output — the exemplar-selection table, the per-dimension comparison table, a
governing-pattern finding distinguishing a universal convention from a genre-conditional one, and the
checklist items that finding produces.
A typical real run is larger: 6–12 real exemplars and 5–8 structural dimensions, applied to an actual target draft, ending in a ranked deviations list against that draft.
This skill is instructions-only — there is no CLI tool to run. The method is meant to be executed by an agent (or a person) working directly with the exemplar texts and the target document; the value is in the close reading and reconciliation, which doesn't reduce to a script.
MIT — see LICENSE.