When to Use

Test, not a list: any generated artifact meant to be read, run, or referenced outside the chat turn itself gets these standards applied. This explicitly includes code and script files, not only prose documents. Reports, essays, letters, resumes, cover letters, blog posts, README files, runbooks, and generated scripts all pass this test.

Out of scope: conversational chat responses, covered by working-preferences' distilled anti-slop reference. Document-type-specific presets and vocabulary for network admin, DevOps, and DevSecOps audiences are covered by the technical-documents skill, which extends this one rather than replacing it.

Making a new skill is covered by skill-creation, and deciding whether a need justifies one, or locating an existing skill, by skill-discovery.

Dials

FORMALITY (1-10): 1 is casual and conversational, 10 is formal or ceremonial. DENSITY (1-10): 1 is explanatory, spelled-out prose, 10 is terse reference material. AUDIENCE_LEVEL (1-10): 1 is written for a reader who needs the why, 10 is written for a reader who only needs the what.

Vibe-Word to Dial Mapping

Document-Type Presets (General)

Technical document types (runbooks, ADRs, incident postmortems, on-call handoff notes, change-management docs, SOPs) and their audience-specific presets live in technical-documents, which layers on top of these dials rather than duplicating them.

Preset priority: when a document could plausibly match a preset in both this file and technical-documents, the technical-documents preset wins whenever its trigger conditions (audience, document type) are met, since it represents the more specific match. This file's own presets are the fallback for everything else.

Content Read

Before generating, state the read: "Reading this as: [document type] for [audience], dials at [F/D/A]." This is a read-back, not a question; proceed unless corrected.

Keep the read itself to that one line. Where the dials depart from the preset, or blend two document types, add one sentence under it naming what moved and why. That sentence belongs to the read rather than breaking it: a dial value differing from the table is a judgment the reader needs to be able to check, and an unexplained departure is indistinguishable from an error.

This is the document-specific instantiation of working-preferences' Ambiguity Handling principle. Document type and target audience are the variables that actually vary per task, so the dial-setting step lives here rather than in the universal file.

working-preferences' distilled anti-slop list is a floor that applies everywhere. The full reference-file check below is the required ceiling for any document that triggers this skill; the distilled list never substitutes for it here.

If the descriptive half of the read names more than one document type or register, the stated dial values must reflect an actual blend or an explicit, stated choice between them, rather than the values for only one of the named types. Check the two halves against each other before proceeding, and put the reasoning in the sentence described above.

When Not to Generate Yet

The ban on fake-precise numbers below has a consequence worth naming. When a request supplies almost nothing verifiable, honoring that ban produces a document made largely of bracketed placeholders. That is the correct output for the rules as written, and it is usually the wrong thing to hand someone: a form asserting the shape of an answer reads as a finished document to anyone who skims it.

Before generating, weigh what the document's substance rests on against how much of it the request actually supplied. Where placeholders would carry most of that load, the move is working-preferences' dialog-before-large-builds gate: ask the calibrating questions first, then generate once against real values. A document explicitly requested as a template is the exception, since placeholders are its content rather than a gap in it.

Either way, say which path you took. "Generating against placeholders for [X] and [Y]" is a read the reader can correct before acting on it. Shipping the same document without that line leaves them to discover the gap themselves, usually later than they would have chosen.

Pre-flight Checklist

Run before shipping any document:

• Zero em-dashes, zero en-dash-as-separator. Grep-verify, don't eyeball it.
• No banned vocabulary or phrases from references/banned-words.md.
• No structural patterns from references/structural-patterns.md.
• Fake-precise numbers only if real or explicitly labeled as example.
• One copy register maintained throughout; no mixing technical, editorial, and marketing tone in the same document.
• Every claim traces to something real, stated, or explicitly marked as illustrative.
• Scoring rubric run (below), threshold met.

State the completion at delivery as a short block, one entry per item above, each marked done or explicitly marked skipped with a reason. A block rather than a single line, because seven items compressed into one sentence become a claim that the checks happened, which is the very thing the completion statement exists to evidence. A mechanical check that already left a visible trace, a grep that ran in the transcript, can cite that trace instead of restating it. Don't leave checklist execution to be inferred from what happens to be visible in the response.

Scoring Rubric

Rate 1-10 on each dimension:

Below 35/50: revise before shipping.

Three-Tier Threshold Framework

Switch from the current two-tier structure (SKILL.md plus flat references/) to a three-tier structure (gateway plus category index plus individual skill files) when any of the following trip:

• Template count in references/templates/ exceeds 8 files.
• Combined reference-file token cost (banned-words plus structural-patterns plus examples loaded together) exceeds roughly 8K tokens.
• A second, genuinely distinct dial system emerges that this file's FORMALITY/DENSITY/AUDIENCE_LEVEL structure doesn't serve well.

This is a standing check, not a one-time decision. Re-evaluate whenever references/templates/ grows. technical-documents has its own templates folder and should apply this same threshold independently.

References

• references/banned-words.md: full categorized vocabulary and phrase lists, lifted near-verbatim from source material where exact phrasing matters for pattern matching.
• references/structural-patterns.md: full pattern catalog with detection signals.
• references/examples.md: before/after transformation pairs.
• references/templates/: general document-type templates, populated iteratively as real documents get produced. Empty at initial build.