Antislop Docs

Anti-Slop Docs — so no document reads or looks AI-generated

Double goal: remove the "AI tells" of a generated document in its TWO layers. An AI-made document gives itself away by how it is WRITTEN (an intro that restates the prompt, everything in bold, "Furthermore/Moreover", bullets for everything, an "In conclusion" that summarizes) and by how it LOOKS (Calibri 11 single-column, headings with no hierarchy, box-bordered tables, zero air). This skill forces deterministic structure, voice and editorial design BEFORE you write, and verifies them at the end.

Same philosophy as antislop-web: bias correction + signature element, here over the word and the page. Zero MCP, zero external service.

When to use it

• At the start of any document (proposal, report, memo, one-pager, whitepaper, case study): generate the DOC.md (Part A) before writing.
• While writing or refactoring sections: apply the voice directives (Part B1) and the editorial design directives (Part B2).
• In your quality gate before delivery: run the pre-flight (Part C).

When NOT to use it

• It does not replace substance. A well-laid-out document with no real content is still empty.
• It is not for text that is PRESENTED (lots of titles, little prose, projected): that is a deck, use antislop-decks.
• Legal/regulated documents: the VOICE anti-slop applies, but the structure is governed by the norm/legal template; do not impose BLUF on a contract. Coordinate with the relevant legal process.

Engine integration (how it runs)

• docx from scratch -> docx-js/OOXML or python-docx: build the .docx with real styles.
• docx over a template / editing someone else's -> redlining/tracked changes. For legal/business documents use the redlining workflow.
• Google Docs -> a Docs automation/API when the client collaborates in Workspace.
• Final PDF -> a PDF engine for the read-only deliverable.
• Voice (text layer): lean on a plain-writing / de-slop / proofreading skill as the "how" of the prose; this skill gives the discipline (what to prune, what to force).
• Base discipline: inherit from antislop-web the palette, the emoji ban, credible data and re-tokenization to the client brand.

PART A — DOC.md generator (per-document source of truth)

Generate a DOC.md in the project folder. Single source of truth for structure, voice and design.

A.1 Dials (translate "make it look professional" into rules)

Set three values 1-10. Suggested baseline: FORMALITY 6, DENSITY 5, ORNAMENT 2.

• FORMALITY (conversational 1 <-> institutional 10): governs voice. An internal memo lives at 4; a regulatory whitepaper at 8. Do not confuse formal with stiff.
• DENSITY (airy one-pager 1 <-> dense technical report 10): governs prose per page and figure count. >7 forces an executive summary and navigation (table of contents, headers).
• ORNAMENT (sober editorial 1 <-> magazine with devices 10): almost always <= 3 in a business document. High ornament without justification is a tell (a Canva template).

A.2 Structure (the document's #1 lever)

The biggest tell is the model-essay shape: an intro that reformulates the task, N sections with a bold lead-in, a "Conclusion" that repeats. Kill it with architecture.

• BLUF / pyramid (Minto), by default: the conclusion and recommendation GO FIRST. The busy reader gets the what and the "so what" in the first paragraph; the detail supports afterward. Do not make the thesis wait until the end.
• BAN the intro-that-restates-the-prompt: no "In today's world...", "This document aims to...". First sentence = the single most important claim.
• BAN the conclusion-summary: the close delivers implication or next action, it does not repackage what was said.
• Titles that inform: a title states the section's point ("The migration pays back in 7 months"), not a label ("Cost analysis"). Sentence case, no all-Title-Case, no ALL CAPS.
• Each section earns its place: if a section does not change the reader's decision, cut it.

A.3 Voice (writing anti-slop) — the tell you READ

• BAN bold-bombing: bold only for real, sparse signposting (a key term the first time, a critical figure). If half the page is bold, nothing is bold.
• BAN AI connector cliches: "Furthermore", "Moreover", "Additionally", "It is important to note", "In today's fast-paced world", "Delve", "Leverage", "Robust", "Seamless", "Elevate". Concrete subject-verb-object sentences.
• Prose as default, list as exception (A.4).
• Concrete over vague: real names, figures, dates and examples. Zero filler ("several", "various benefits", "numerous advantages") without the data behind it.
• Varied sentence rhythm: not every sentence the same length and mold. Syntactic monotony is a tell.
• No em-dash as a crutch (overuse), no emojis, no decorative unicode. Quotes and dashes consistent with the document config.

A.4 Prose vs list (discipline, the structural tell)

• Prose is the default. An argument, an explanation or a narrative go in paragraphs, not bullets. AI chops everything into lists because it is easy; it reads like notes.
• A list earns its place only when items are genuinely parallel and scannable (steps, requirements, criteria). 3-7 items; if there are 12, it is a table or two groups.
• BAN one-line bullets chained as a chopped paragraph. If the bullets form a running sentence, it was a paragraph.
• Lists parallel in form: all items the same grammatical structure (all start with a verb, or all with a noun). Mixing is a tell.

A.5 Editorial design (the tell you SEE)

• Real typography, not default. A serif/sans pair with character (inherit from the client DESIGN.md if it exists). BAN Calibri 11 / Times New Roman 12 with no decision. Body with generous leading (1.4-1.5), measure 60-75 characters (not full page width).
• Hierarchy by scale AND weight, consistent. H1/H2/H3 with clear jumps and more space above than below (the heading belongs to the text that follows). No hierarchy by underline or ALL CAPS.
• Margins and air: generous margins; text glued to the edge is a default-template tell. Line spacing and paragraph spacing that breathe.
• Designed tables and figures: no box-borders per cell; use horizontal rules (border-t) and alignment (numbers right, in mono if they are data). Every figure with a number, an assertive caption and a source. Never a pixelated screenshot.
• Cover/header with the client's brand, not the vendor's default cover. Footer with pagination and a discreet metadatum.
• No emojis, no clipart, no decorative bullet icons (inherited).

A.6 Data, citations and credibility (inherited + doc)

• Organic, real figures (no 99.9% filler). Every factual claim with its source; no invented quotes or numbers. Contextual names and examples, not "Acme"/"John Doe".

A.7 Anti-patterns (the document's AI tells — encode as "NEVER" in DOC.md)

No intro that reformulates the task. No "In conclusion" that summarizes. No bold-bombing. No "Furthermore/Moreover/It is important to note". No bullets for everything. No chopped-paragraph bullets. No Title Case or ALL CAPS in headers. No Calibri/Times default. No box-bordered tables. No emojis or clipart. No vague filler without data. No sentences all the same length. No AI copy cliches. No fake round numbers. No vendor default cover.

A.8 Signature element (the one AFFIRMATIVE rule — mandatory)

Passing every ban makes a correct but forgettable document. Each document declares ONE signature element: the device it is remembered for. Just one.

Criteria (all three): memorable, owned (comes from the content/business, not a template) and sustained (coherent across the document). It can live in:

• Data/structure: an owned way of opening every document for the client (a key-figures panel at the top, a "status in one sentence") that nobody in the sector uses.
• Typographic system: a display face with character in the headers, a recurring margin note, an owned callout treatment.
• Voice: a radical clarity (short sentences, zero jargon) sustained as a house style.
• Figure: an owned chart or diagram type that repeats and becomes recognizable.

Guard: declared in DOC.md BEFORE writing and verified in the pre-flight. If you cannot name it in one sentence, it does not exist.

A.9 Output template (DOC.md)

# Doc: [Title] — Client [X] — [type: proposal/report/memo/whitepaper]

## 0. Dials
FORMALITY: n | DENSITY: n | ORNAMENT: n   (baseline 6/5/2)

## 1. Format and engine
Deliverable: docx / Google Docs / PDF | Engine: docx-js / Docs API / PDF
Inherits brand from: [client DESIGN.md | identity declared below]

## 2. Structure (BLUF)
[The thesis/recommendation in one sentence (goes first). Then the section skeleton, each with
its title-that-informs. Mark whether it carries an executive summary and a table of contents.]

## 3. Signature element (mandatory, exactly one)
**[Name in one sentence]** — [what it is, which layer, why it comes from this content.]
Sustained in: [where it recurs].

## 4. Voice
FORMALITY n; prose by default; connector/bold bans; list only for [cases].

## 5. Typography and layout
Display: [font] | Body: [font] serif/sans | Mono: [data]
Measure 60-75ch, leading 1.4-1.5, generous margins, tables with horizontal rules.

## 6. Figures and data
[Figure type, numbering, cited source, palette derived from the accent]

## 7. Anti-patterns (BANNED)
[List from A.7 adapted to this document]

PART B — Production directives

B1. Voice (while writing)

1. First sentence = the most important claim (BLUF). Never a windup.
2. Prose by default; a list only if items are parallel and scannable.
3. Rare, deliberate bold; zero bold-bombing.
4. Prune connector cliches and vague filler; every sentence adds or is deleted.
5. Vary sentence length and mold. Reading aloud catches the monotony.
6. Writing support: a plain-writing / de-slop / proofreading skill as the "how"; this skill fixes the "what".

B2. Layout (while building the file)

1. docx from scratch -> docx-js/python-docx: define REAL styles (heading 1/2/3, body, caption, quote) once and apply them; do not hand-format each paragraph. Font, leading, measure and margins per the DOC.md.
2. docx editing someone else's or legal/business -> redlining: tracked changes; do not rebuild the client's document, propose changes.
3. Google Docs -> Docs automation/API: native paragraph styles; font available in Google Fonts.
4. PDF -> a PDF engine for the final read-only deliverable.
5. Verify the font is available before locking it (inherited): if not, it falls back to a default and breaks the brand.
6. Re-tokenization: if a client DESIGN.md exists, inherit its typography, accent and signature element in document form. The client must not be able to trace their doc to a template.

PART C — Pre-flight (quality gate before delivery)

Any failure = not shippable until fixed.

• BLUF: the thesis/recommendation is in the first paragraph, not at the end (A.2).
• No intro-that-restates-the-task; no conclusion-summary.
• Titles that inform, in sentence case (no Title Case/ALL CAPS everywhere).
• Prose by default; lists only where they earn it and are parallel (A.4).
• No bold-bombing; bold sparse and deliberate (A.3).
• No AI connector cliches or vague filler without data.
• Varied sentence rhythm (read a random paragraph aloud).
• Real typography (no Calibri/Times default), measure 60-75ch, generous leading (A.5).
• Consistent header hierarchy; margins and air; tables with horizontal rules, not boxes.
• Figures numbered with an assertive caption and a source; data organic and cited.
• Signature element (A.8) present, single, sustained. Name it in one sentence.
• No emojis/clipart; cover with the client's brand, not the vendor default.

Limitations

• It gives structure, voice and layout; it does not guarantee the argument is correct: substance rules.
• In legal/regulated documents the structure is set by the norm; apply only the voice and layout layer that does not contradict the legal requirement.
• The defaults (dials, typography) are 2026; check them against the brand and a11y before locking them in.