@arbiterforge/ca-pi 0.6.0

package/plugins/ca-pi/includes/anti-slop-design/core.md

@@ -0,0 +1,244 @@
+# anti-slop-design · core
+
+The always-loaded decision layer. Load this for any design-bearing deliverable, then add the craft
+and medium leaves the Design Read calls for.
+
+> Section numbering is shared across the whole bundle, so this file skips some numbers on purpose:
+> §4 typography, §5 color, §6 layout, §7 medium playbooks, and §9 images live in their own leaves.
+> core owns §0-§3 (decision layer), §8 (universal tells), and §10 (core pre-flight).
+
+## 0. Why slop happens (the one idea that explains every rule)
+
+A language model produces output by regressing toward the statistical center of its training
+distribution. For design, that center is the average of everything published, and that average is
+generic, templated, and forgettable. "AI-slop" is not a style. It is the absence of a decision: the
+model defaulting to the mean instead of choosing for *this* brief, *this* audience, *this* artifact.
+
+Every rule in this bundle forces a deliberate move off the center. When you reach for the first thing
+that "looks designed," you are reaching for the center. Excellence lives in the specific, the
+considered, and the slightly uncomfortable choice a generic generator would never make.
+
+Two corollaries that govern everything else:
+
+- **A default chosen without a reason is a tell.** Inter, a centered hero, three equal cards, a purple
+  gradient, skill bars on a resume, a pie chart: none is wrong in isolation. They are wrong as
+  *reflexes*. If you cannot say in one sentence why a choice fits this brief, it is slop.
+- **Consistency is cheap credibility; novelty is expensive credibility.** Locking one accent color,
+  one corner radius, one type system costs nothing and reads as intentional. A striking layout costs
+  judgment and can fail. Spend the cheap credibility everywhere; spend the expensive kind only where
+  the brief earns it.
+
+## 1. The Prime Directive: read the brief before you generate
+
+Before producing anything, establish a one-line **Design Read**. Do it every time, internally.
+Surface it only when it helps the user follow your choices; in a mid-task or CLI context a silent read
+usually beats a printed preamble.
+
+> "Reading this as a `<medium>` for `<audience>`, in a `<register>` register, leaning toward a
+> `<aesthetic family>`."
+
+Read the signals in order: **medium** (doc / chart / slide / resume / web / report), **audience**
+(who decides, who scans, who parses it with software), **register** (institutional vs. expressive),
+**existing assets** (brand color, logo, type, prior templates), **quiet constraints** (accessibility,
+regulated industry, ATS, projection, print). Constraints override taste.
+
+If the read genuinely diverges, ask **one** question, never a multi-question dump. If you can infer
+confidently, declare the read and proceed.
+
+## 2. The Dials
+
+After the read, set four dials. Every density, layout, and emphasis decision in the bundle is gated by
+them. Use these names; do not invent aliases.
+
+- **`STRUCTURE`** (1-10): 1 = rigid symmetry and grid, 10 = asymmetric and expressive.
+- **`DENSITY`** (1-10): 1 = airy/gallery, 10 = packed/cockpit.
+- **`REGISTER`** (1-10): 1 = austere/institutional/legal, 10 = playful/consumer/brand.
+- **`MOTION`** (1-10): *interactive media only*. 1 = static, 10 = cinematic. **N/A and ignored for any
+  printed or static artifact** (PDF, docx, slide export, resume).
+
+### 2.A Inference from the read
+
+| Brief reads as | STRUCTURE | DENSITY | REGISTER | MOTION |
+|---|---|---|---|---|
+| Legal / compliance / regulated / public-sector | 2-3 | 4-6 | 1-2 | 1-2 |
+| Technical review / due diligence / whitepaper | 3-5 | 6-8 | 2-4 | 2-3 |
+| Executive brief / board deck | 4-6 | 3-5 | 3-5 | 3-5 |
+| Resume / CV | 3-5 | 5-7 | 2-4 | n/a |
+| Dashboard / analytics | 3-5 | 7-9 | 3-5 | 3-6 |
+| Editorial / report-as-narrative | 5-7 | 3-5 | 4-6 | 3-5 |
+| Consumer / product landing | 7-9 | 3-5 | 6-8 | 6-8 |
+| Agency / portfolio / experimental | 8-10 | 3-4 | 7-9 | 7-10 |
+
+The values are a starting point, not a script. A regulated brief with a strong brand can still earn
+REGISTER 4. The point is to choose, and to be able to defend the choice.
+
+### 2.B What each dial actually changes
+
+A dial is useless if it does not change an output. Concretely:
+
+- **`STRUCTURE`** drives composition: low → centered, symmetric, single grid; high → asymmetric, split
+  composition, deliberate off-grid moments. Above 4, drop the centered-everything default (see layout
+  §6.B).
+- **`DENSITY`** drives whitespace and elements-per-screen: low → generous margins, few items, card
+  chrome; high → tight rows, more per screen, drop card chrome and let data breathe (layout §6.C).
+- **`REGISTER`** drives tone of type, color, and copy: low → restrained faces, single muted accent,
+  literal copy; high → expressive display type, a bolder accent, more voice.
+- **`MOTION`** drives animation budget on interactive media only: low → state-change feedback only;
+  high → composed reveals and transitions. Ignored for static artifacts.
+
+### 2.C When dials conflict
+
+Resolve in this order: **quiet constraints (accessibility, ATS, print, regulated) > REGISTER > the
+rest.** A legal brief with a strong brand (high REGISTER pull, low REGISTER constraint) resolves
+toward the constraint: restrained, with the brand expressed in the one accent and the type, not in
+layout fireworks.
+
+## 3. Universal anti-slop laws (every medium, no exceptions)
+
+These apply to a resume, a chart, a slide, and a landing page equally.
+
+### 3.A The em-dash rule
+
+Banned: the em-dash (`—`) and the en-dash (`–`) used as a **sentence-level separator in generated
+prose** — headlines, body, captions, labels, button text, slide titles, marketing copy. That specific
+use is the single highest-signal tell of machine-generated text in current testing, and it is the
+first thing a reader's pattern-matcher flags. Restructure instead: split into two sentences with a
+period, use a comma, parentheses, or a colon.
+
+This is a scoped ban, not a blanket glyph ban. The following are **exempt** and never findings:
+
+- **Quoted content** — a quote, citation, or excerpt from a user, source, or document that itself uses
+  the dash.
+- **Code and string literals** — anything inside code blocks, identifiers, or literal strings.
+- **Math** — minus sign, and en-dash where it is correct typography.
+- **File paths, URLs, and identifiers** that contain the character.
+- **Proper nouns** — a brand, product, or publication title that legitimately contains a dash.
+- **Numeric and date ranges** where the en-dash is correct typography (`pp. 12–18`, `2019–2024`).
+
+The reviewer BLOCKs only on the banned case (a dash as a prose sentence-separator); an exempt
+occurrence is not a finding. A stray separator dash in otherwise-fine copy is a fix, not a
+catastrophe — restructure it and move on.
+
+### 3.B Copy self-audit (read every visible string before shipping)
+
+Flag and rewrite any string that is:
+
+- **A filler verb or empty intensifier, in metaphorical or marketing use.** Flagged: *elevate,
+  unleash, seamless, seamlessly, leverage, revolutionize, empower, streamline, robust, cutting-edge,
+  next-gen, game-changing, supercharge, unlock, harness, navigate, dive into, delve, tapestry, realm,
+  landscape, in today's fast-paced world.* Replace with concrete verbs that say what happens.
+  **Literal and technical uses pass and are never findings:** "robust error handling," "navigate to
+  `/home`" or `router.navigate`, a test "harness," "leverage" in its finance sense, "unlock" a screen
+  or feature, "the deployment landscape" as a literal map of environments. The tell is the empty
+  marketing gesture, not the word.
+- **A rhetorical AI cadence.** The "it's not just X, it's Y" construction; "whether you're A, B, or C";
+  opening on a rhetorical question; the rule-of-three on every list. One is fine. The *pattern*
+  repeating is the tell.
+- **Monotone rhythm.** AI prose tends toward sentences of even, medium length. Vary it. A short
+  sentence after two long ones reads human. Uniform rhythm reads generated.
+- **A fabricated-precise number** (see 3.D).
+- **Grammatically broken or hallucinated cleverness.** Wordplay that does not parse, forced metaphors,
+  fake-humble craftsman labels ("field notes," "on our bench," "loose plates"). When unsure whether a
+  phrase earns its place, use the plain functional version. Boring-but-correct beats clever-but-wrong.
+
+### 3.C Generic-placeholder ban
+
+- **Names:** never "John Doe," "Jane Doe," "Sarah Chen," "Jack Su." Use realistic, locale-appropriate
+  names. A deliberately-sample identity is fine when **marked** the same way a sample number is (see
+  3.D): an explicit `example`/`sample` tag or a `<!-- sample -->` comment. The tell is a generic
+  placeholder shipping as if it were real; a marked sample is not.
+- **Brands:** never "Acme," "Nexus," "SmartFlow," "Cloudly," "TechCorp." Invent names that sound real
+  for the sector, or use real ones where appropriate. Mark sample brands as above.
+- **Avatars/logos:** no egg-silhouette avatars, no generic user-icon glyphs, no plain text wordmarks
+  where a real logo or a designed monogram belongs.
+- **Stock clichés:** no handshake photo, no diverse-team-at-whiteboard, no glowing-brain-AI image, no
+  arrow-hitting-target.
+
+### 3.D Data and number integrity (this is correctness, not taste)
+
+Fabricated precision is a credibility failure, and in a technical review or report it can be a factual
+one.
+
+- Numbers like `92%`, `4.1x`, `47.2%`, `5.8mm` are acceptable **only** if they come from real data, or
+  are explicitly marked illustrative (`example`, `sample`, `<!-- mock -->`).
+- Do not invent benchmark figures, spec values, or metrics to make a page "feel precise." If you do
+  not have the number, say so or omit it.
+- Real data is messy. **When simulating** realistic data, avoid suspiciously round figures (`50%`,
+  `99.99%`, `1,000,000`); real distributions are lumpy. This heuristic applies only to invented or
+  illustrative data. A real round number (a genuine 50/50 split, a real 1,000,000 row count, an SLA
+  literally specified as 99.99%) is fine and is never a finding.
+- Label estimates as estimates. A reader who later finds an invented number stops trusting the whole
+  document.
+
+### 3.E The four consistency locks
+
+Pick once, apply everywhere, audit before shipping. These are the **default**; a high-`STRUCTURE` or
+high-`REGISTER` brief (portfolio, editorial, experimental) may override a lock where the expressive
+move is the point — a deliberate second accent, a composed theme switch — provided the override is a
+documented rule applied consistently, not a one-off accident. State the override as a visible
+artifact-level note (a design-intent comment or a stated convention) so a reviewer can verify it, the
+same discipline used to mark a sample value in 3.C/3.D. The lock is the floor; breaking it costs the
+expensive credibility from §0, so spend it only where the brief earns it.
+
+1. **Color lock.** One accent color across the whole artifact. A warm-grey report does not grow a blue
+   callout box on page 7. A chart palette does not change between figures.
+2. **Shape lock.** One corner-radius / border treatment system. Sharp, soft, or pill, chosen once.
+   Mixed only under a documented rule applied everywhere.
+3. **Type lock.** One display family, one body family, optionally one mono. Emphasis comes from weight
+   and italic of the *same* family, never from a foreign font for one word.
+4. **Theme lock.** One light/dark/print theme for the whole artifact. No section inverts mid-scroll
+   unless it is a single, deliberate, composed switch.
+
+### 3.F Cut ruthlessly
+
+Slop is often just **too much**. Every section earns its place or is deleted. A 20-row table, a
+12-bullet slide, a 6-paragraph hero, a 30-item award list: these are layout failures, not content. The
+fix is a different component (group, summarize, link to detail), not a longer list.
+
+## 8. Universal AI-tells (cross-medium quick-scan)
+
+Each is banned *as a default*; the brief can override any with a reason. Medium-specific tells live in
+the medium leaves.
+
+**Punctuation and copy**
+- Em-dash / en-dash as a prose sentence-separator (3.A): banned, with the 3.A exemptions.
+- Filler verbs in marketing use and "fast-paced world" cadence (3.B).
+- "It's not just X, it's Y"; "whether you're A or B"; rhetorical-question openers; rule-of-three on
+  everything.
+- Generic names and brands: John/Jane Doe, Acme, Nexus, TechCorp (3.C).
+- Fake-precise numbers (3.D); suspiciously round numbers.
+- Generic step labels: "Step 1 / Phase 1 / Stage 1." The step content is the label.
+- Performative-craftsman section labels: "field notes," "on the bench," "loose plates," "quietly
+  trusted by."
+
+**Visual (cross-medium)**
+- AI-purple / blue glow gradient; neon outer glows.
+- Centered-everything composition.
+- Pure `#000` / `#fff` large fills.
+- Generic stock clichés (handshake, whiteboard team, glowing brain).
+- Egg-silhouette avatars and generic user-icon glyphs.
+- Rainbow categorical palettes.
+
+**Structural**
+- Eyebrow label above every section.
+- Every section the same layout family.
+- Decorative status dots before every list item / nav link.
+- Locale/time/weather atmospheric strips with no real function.
+- "Scroll" cues and version stamps (`v0.6`, `BETA`) used as decoration.
+- Section-number eyebrows (`00 / INDEX`, `001 · Capabilities`).
+
+## 10. Core pre-flight (run before delivering anything)
+
+If a box cannot be honestly ticked, it is not done. Each medium leaf adds its own slice.
+
+- [ ] Design Read established (medium, audience, register, aesthetic family); internal if not surfaced.
+- [ ] Dial values chosen with a reason, not silently defaulted.
+- [ ] **No em-dash / en-dash used as a prose sentence-separator** (the 3.A exemptions are fine).
+- [ ] Copy self-audit done: no marketing-filler verbs, no AI cadence, no monotone rhythm, no
+  hallucinated cleverness.
+- [ ] No generic placeholder names, brands, avatars, or stock clichés.
+- [ ] Every number is real, marked illustrative, or omitted. No fabricated precision.
+- [ ] Color lock, shape lock, type lock, theme lock all hold across the whole artifact.
+- [ ] First impression delivers the point in the space available.
+- [ ] Content cut ruthlessly; no data-dump where a different component belongs.
+- [ ] No universal tell (§8) ships without a brief-driven reason.

package/plugins/ca-pi/includes/anti-slop-design/images.md

@@ -0,0 +1,32 @@
+# anti-slop-design · images
+
+Load when the artifact carries visual assets.
+
+Scope: this leaf governs **visually-composed artifacts** (web/marketing pages, slides, and reports
+that are designed rather than text-first). It does NOT mandate images for text-first deliverables — a
+whitepaper, a legal memo, a resume (which often wants no photo at all), CLI output, a PR description,
+or a CHANGELOG are complete without imagery. Load it only when the INDEX map routes your medium here.
+
+## 9. Images and visual assets
+
+- **Generate or source real images first.** If an image-generation tool is available, use it for
+  section-specific assets at the right aspect ratio. Otherwise use a real photo source (a placeholder
+  service, licensed stock, brand assets). Check provenance and license before using a sourced image:
+  prefer assets whose license permits the use, consistent with the project's dependency-provenance
+  posture. Do not embed an image of unknown origin.
+- **A visually-composed artifact needs real images.** For a designed page or deck, a pure-text result
+  is not minimalism; it is incomplete. A restrained piece still wants at least one or two real images.
+  (This does not apply to the text-first deliverables named in the scope note above.)
+- **Real logos** for credibility (icon libraries or generated monograms), rendered correctly in both
+  light and dark, logos only with no category labels.
+- **Last resort:** if no image is possible, leave a clearly labeled placeholder slot and tell the user
+  exactly what is needed and at what size. Do not fill the gap with fake-div screenshots or hand-drawn
+  decorative SVG.
+
+## Tells (images)
+
+Generic stock clichés and egg-silhouette avatars live in core §8. Image-specific tells:
+
+- Fake-div screenshots; hand-rolled decorative SVG as a substitute for real imagery.
+- Plain text wordmarks where a real logo or designed monogram belongs.
+- A sourced image with no checked license or provenance.

package/plugins/ca-pi/includes/anti-slop-design/layout.md

@@ -0,0 +1,45 @@
+# anti-slop-design · layout
+
+Load it for any medium whose row in the INDEX load map names it. Medium-neutral; the medium leaves add
+specifics.
+
+## 6.A The first-impression law
+
+Every artifact has a first screen that decides whether the reader continues: the first page of a
+report, the top of a resume, the first slide, the hero of a page, the headline chart of a dashboard.
+**That first impression must deliver the core value in the space available without forcing a scroll or
+a page-turn to find the point.** If the content does not fit, cut copy or reduce scale; do not let it
+overflow and bury the point.
+
+## 6.B Structure and rhythm
+
+- **Grid over ad-hoc math.** Use a real grid system; do not hand-tune percentage widths until things
+  "look about right."
+- **Anti-center bias when `STRUCTURE > 4`.** Centered-everything is the safe default and reads generic.
+  Use asymmetry, split composition, deliberate white space. Centered is fine for a manifesto or a
+  single-message moment where the message *is* the design.
+- **No three-equal-cards reflex.** The identical-three-feature-cards row is the canonical slop layout.
+  Vary cell sizes, alternate composition, or use a different family.
+- **Layout-repetition cap.** A layout family appears at most a couple of times per artifact. Eight
+  sections that all look the same is templated. Aim for at least four distinct families across a long
+  page or report.
+- **Eyebrow restraint.** The small uppercase wide-tracked label above every section header is a
+  templated rhythm tell. Maximum one per three sections. Usually the headline alone is enough; the
+  section's position already categorizes it.
+- **No floating corner sub-text** and no "big headline left / tiny explainer paragraph right"
+  split-header as a reflex. If you need a headline and an explainer, stack them.
+
+## 6.C Cards and elevation
+
+Use cards (boxes with elevation) only when elevation communicates real hierarchy. Otherwise group with
+a hairline, a divider, or white space. Tint shadows to the background hue; no pure-black drop shadows on
+light fills. At high `DENSITY`, drop card chrome entirely and let data breathe in plain layout.
+
+## Tells (layout)
+
+Centered-everything, same-layout-every-section, and eyebrow-above-every-section live in core §8.
+Layout-specific tells:
+
+- Three identical feature cards.
+- Glassmorphism on everything.
+- Split-header ("big headline left / tiny explainer right") as a reflex.

package/plugins/ca-pi/includes/anti-slop-design/medium-cli.md

@@ -0,0 +1,39 @@
+# anti-slop-design · medium: CLI / terminal output
+
+Load for output rendered in a terminal: status lines, progress, reports printed to stdout, log lines,
+TUI panels. Pair with `core`. Type/color/layout leaves mostly do not apply (the terminal owns the
+font; color is a constrained palette).
+
+## 7.F CLI / terminal output
+
+- **The terminal is a constrained medium; respect it.** Variable width, a 16/256/truecolor palette
+  that the user's theme controls, monospace only, and a reader who often pipes or greps the output.
+- **Degrade without color.** Never let color be the only carrier of meaning (core 3.B/§5 spirit): a
+  red number must also read as bad when color is stripped. Always honor `NO_COLOR`. For a normal CLI,
+  also drop ANSI on a non-TTY pipe so redirected output stays parseable. The exception is an
+  *intentionally-piped colored UI* (e.g. a statusline the host always pipes and renders in color):
+  there, keep color and gate only on `NO_COLOR`, never on `isatty` — an isatty test would strip color
+  in normal use.
+- **Width is unknown.** Fit to the reported width with a margin; never assume 80. A line that wraps in
+  a narrow terminal corrupts a box or table. Clamp and truncate deterministically.
+- **Glyph width is real.** CJK and many emoji are two columns; combining marks are zero. Count visible
+  columns, not characters, or box-drawing and alignment break.
+- **No decoration tax.** Emoji sprinkled per line, gratuitous box-drawing, and rainbow ANSI are the
+  terminal equivalent of clip art. One accent color, aligned columns, and whitespace carry it.
+- **Copy laws still apply.** core §3.A/§3.B hold for any prose in help text, errors, and summaries.
+  Error messages say what happened and what to do, not "an error occurred."
+
+## Tells (CLI)
+
+- Color as the sole signal; ignoring `NO_COLOR` (or, for a normal CLI, no non-TTY fallback).
+- Hardcoded 80-column assumptions; lines that wrap and corrupt a box.
+- Character-count math that misaligns on CJK/emoji width.
+- Emoji or box-drawing as decoration rather than structure.
+- Vague error strings ("something went wrong").
+
+## Pre-flight slice (CLI)
+
+- [ ] Renders correctly at narrow and wide widths; no wrap corruption.
+- [ ] Honors `NO_COLOR`; a normal CLI emits plain output when piped; color is never the only signal.
+- [ ] Column math counts visible width (CJK/emoji safe).
+- [ ] Decoration earns its place; errors are actionable.

package/plugins/ca-pi/includes/anti-slop-design/medium-dataviz.md

@@ -0,0 +1,43 @@
+# anti-slop-design · medium: data visualization
+
+Load for dashboards, charts, and technical-review figures. Pair with `core`; add `typography`, `color`,
+and `layout` (axis labels, titles, and the annotated-insight sentence are type choices).
+
+The governing principle is **maximize the share of ink that carries information** (Tufte's data-ink
+ratio). Slop in charts is decoration mistaken for design.
+
+## 7.B Data visualization
+
+- **Remove chrome:** no 3D, no drop shadows on bars, no gradient fills on series, no heavy gridlines,
+  no chartjunk. Light or absent gridlines; let the data shapes carry the figure.
+- **Right chart for the question:** bar for category comparison, line for trend over time, scatter for
+  correlation, small-multiples for many series. **Avoid pie charts for more than two or three slices**
+  (humans read angle poorly). Avoid dual-axis charts that imply false correlation.
+- **Honest axes:** bar charts start at zero. Do not truncate an axis to exaggerate a difference. If you
+  must zoom a line chart, say so.
+- **Direct labeling** beats a legend when it fits; the reader should not bounce between a key and the
+  data.
+- **Color encodes meaning:** sequential scale for ordered magnitude, diverging for a meaningful
+  midpoint, categorical for unordered groups, capped at roughly six hues. Colorblind-safe palettes.
+  Color is not decoration here.
+- **Sort by value**, not alphabetically, unless the category order is itself meaningful.
+- **Annotate the insight.** A good figure tells the reader what to notice. One sentence of takeaway near
+  the relevant mark.
+- **Data integrity** (`core` 3.D) is non-negotiable in a technical review. A fabricated benchmark in a
+  CTO-facing document is a fireable mistake, not a styling choice.
+
+## Tells (data visualization)
+
+- Pie charts with many slices; 3D charts; gradient-filled bars; chartjunk.
+- Truncated axes that exaggerate a difference; dual-axis false correlation.
+- A legend where direct labels would fit.
+- Alphabetical sort where value-sort carries the insight.
+- A figure with no annotated takeaway.
+
+## Pre-flight slice (data visualization)
+
+- [ ] Right chart for the question; no pie-with-many-slices, no 3D, no chartjunk.
+- [ ] Axes honest (bars from zero, no deceptive truncation).
+- [ ] Color encodes meaning; colorblind-safe; direct labels where they fit.
+- [ ] The insight is annotated, not left for the reader to hunt.
+- [ ] Every figure's data is real and sourced (integrity is correctness here).

package/plugins/ca-pi/includes/anti-slop-design/medium-diagram.md

@@ -0,0 +1,35 @@
+# anti-slop-design · medium: diagrams
+
+Load for architecture, flow, sequence, and entity diagrams (Mermaid, Graphviz, ASCII, drawn). Pair
+with `core`; the INDEX map also loads `color` and `layout` for this medium.
+
+## 7.G Diagrams
+
+- **A diagram exists to make one relationship legible.** If it does not answer a specific question
+  ("how does a request flow?", "what depends on what?"), it is decoration. Name the question first.
+- **Cap the node count.** Past roughly 12-15 nodes a single diagram becomes a hairball. Split by
+  subsystem, collapse detail behind a labeled group, or show the slice that matters.
+- **Edges carry meaning; minimize crossings.** Pick a consistent direction (top-down for hierarchy,
+  left-right for a pipeline) and stick to it. Label edges when the relationship is not obvious; do not
+  label every edge if direction already says it.
+- **Color encodes, never decorates** (the same color-encodes-meaning discipline used for charts): use
+  it for one categorical axis (layer, ownership, status), colorblind-safe, capped at a few hues. A
+  rainbow of boxes is noise.
+- **A legend when the encoding is not self-evident**, and consistent shapes (one shape per node kind).
+- **Honest abstraction.** Do not invent components or connections to make the picture symmetric. The
+  diagram is a claim about the system; a wrong box is a wrong claim (core 3.D spirit).
+
+## Tells (diagrams)
+
+- A hairball: too many nodes, crossing edges, no clear direction.
+- Decorative multi-color boxes with no encoding meaning.
+- Mixed directions / inconsistent node shapes within one diagram.
+- Every edge labeled, or none labeled where the relationship is ambiguous.
+- Invented boxes/links that do not exist in the real system.
+
+## Pre-flight slice (diagrams)
+
+- [ ] The diagram answers one named question; node count is bounded.
+- [ ] Consistent direction; crossings minimized; shapes consistent.
+- [ ] Color encodes a real axis (colorblind-safe) or is absent; legend where needed.
+- [ ] Every box and edge maps to something real in the system.

package/plugins/ca-pi/includes/anti-slop-design/medium-documents.md

@@ -0,0 +1,70 @@
+# anti-slop-design · medium: documents
+
+Load for reports, memos, whitepapers, resumes, PR descriptions, CHANGELOG sections, tables, and spec
+sheets. For a laid-out document pair with `core` + `typography` + `images`. For PR/CHANGELOG Markdown
+prose (§7.A.1), `core` + this leaf is enough — those are copy-law deliverables, not composed pages.
+
+## 7.A Documents (reports, memos, whitepapers, resumes)
+
+- **Hierarchy through type and space**, not heavy rules and shaded boxes on everything. A horizontal
+  hairline under a section title beats a filled banner.
+- **Serif or humanist-sans body** for sustained reading (see `typography` 4.A). Body size 10-12pt for
+  print, 16-18px for screen.
+- **No clip art, no rainbow section colors, no decorative icons** scattered for "visual interest."
+
+### Resume specifics (the dominant tells in this medium)
+
+- No skill-proficiency bars ("Python: 85%"). They are meaningless, unverifiable, and read as template.
+  List skills plainly or, better, demonstrate them through quantified accomplishments.
+- Reverse-chronological, consistent date format throughout, one page unless senior or academic.
+- Quantify every bullet that can be quantified, with real numbers (`core` 3.D applies hard here).
+  "Reduced deploy time 40%" beats "improved deployment efficiency."
+- No photo unless the regional hiring norm expects one. No two-column layout if an ATS will parse it,
+  unless you know the target system handles it.
+- Active voice, concrete verbs, no filler (`core` 3.B). Cut the "Objective" statement; cut "References
+  available on request."
+
+## 7.A.1 PR descriptions and CHANGELOG (Markdown render target)
+
+These are documents rendered as Markdown on a platform (GitHub, a release page), not laid-out pages,
+so layout/typography/color do not apply. What matters:
+
+- **Write for the render, not the source.** Use real Markdown structure (headings, lists, fenced code)
+  that renders cleanly; do not hand-align columns that the renderer will collapse.
+- **The copy laws carry the weight.** core §3.A (no prose separator dashes) and §3.B (no marketing
+  filler) are the whole game here. A PR body says what changed, why, and how it was verified, in plain
+  declarative prose.
+- **Lead with the point.** First line / summary states the change; details follow. A reviewer reads the
+  first two lines and the test plan.
+- **Link, do not paste.** Reference issues, ADRs, and commits by link/id; do not inline large blobs.
+- **CHANGELOG:** keep entries terse, grouped (Added / Fixed / Performance), user-facing, and in the
+  project's existing format. Real version numbers only (core 3.D); no invented metrics.
+
+## 7.E Tables and spec sheets
+
+Long structured lists fail the same way everywhere: a default `<ul>` or a row-per-line table with a
+hairline under each row.
+
+- **Minimal rules.** Horizontal hairlines between logical groups, not under every row. No vertical
+  lines unless the data demands them.
+- **Align numbers.** Right-align or decimal-align numeric columns.
+- **No zebra striping** unless the table is genuinely dense.
+- For more than five items, reach for a different component: grouped 2-3 clusters with sparse dividers,
+  a card-per-item grid, tabs or an accordion if categorizable, or a featured-few-plus-collapsed-rest
+  disclosure. The list itself is rarely the answer to a long list.
+
+## Tells (documents)
+
+- Skill-proficiency bars on resumes.
+- Hairline under every table row; row-per-line spec sheets.
+- Office-default body face (Calibri, Times New Roman) chosen by reflex.
+- Decorative icons / rainbow section colors scattered for "interest."
+- Filled section banners where a hairline would do.
+
+## Pre-flight slice (documents)
+
+- [ ] Body face deliberate (serif/humanist for long reading), not an Office default by reflex.
+- [ ] Hierarchy from type and space, not boxes and rules on everything.
+- [ ] Tables: minimal rules, aligned numbers, grouped not row-striped.
+- [ ] Resume: no skill bars, quantified bullets with real numbers, consistent dates, ATS-safe if
+  needed.

package/plugins/ca-pi/includes/anti-slop-design/medium-slides.md

@@ -0,0 +1,30 @@
+# anti-slop-design · medium: slides
+
+Load for presentations and slide decks. Pair with `core`; add `typography`, `color`, `layout`, `images`.
+
+## 7.C Presentations / slide decks
+
+- **One idea per slide.**
+- **The slide title is the takeaway, not the topic.** "Latency dropped 60% after the cache rewrite,"
+  not "Performance." The audience should grasp the point from the title alone.
+- **Big type, few words.** No paragraphs, no seven-bullet avalanche. If a slide needs a paragraph, it
+  is a document, not a slide.
+- **Charts redrawn for projection:** huge labels, few series, high contrast. A chart pasted from a
+  spreadsheet at desk-reading scale is unreadable from row five.
+- **Consistent master and grid.** Do not center every element by default; use the same margins and
+  baseline across slides.
+- **No clip art, no word art, no transition zoo.** If motion is used, it is motivated (a reveal
+  sequence that matches the argument), not decorative.
+
+## Tells (slides)
+
+- Topic-label titles instead of takeaway titles.
+- Seven-bullet avalanche; paragraphs on a slide.
+- Spreadsheet charts pasted at desk-reading scale.
+- Clip art, word art, transition zoo.
+
+## Pre-flight slice (slides)
+
+- [ ] One idea per slide; title is the takeaway, not the topic.
+- [ ] Big type, few words; charts redrawn for projection.
+- [ ] Consistent master; motion (if any) motivated.

package/plugins/ca-pi/includes/anti-slop-design/medium-web.md

@@ -0,0 +1,39 @@
+# anti-slop-design · medium: web and interactive
+
+Load for landing pages, portfolios, and component UIs. Pair with `core`; add `typography`, `color`,
+`layout`, `images`. This is the leaf `frontend-author` loads.
+
+## 7.D Web and interactive interfaces
+
+Front ends carry every universal law in `core` §3 plus these web-specific essentials:
+
+- **Real images, not fake-div screenshots.** A product UI built from styled `<div>` rectangles to
+  simulate a screenshot is the number-one web tell. Use a real screenshot, a generated image, a real
+  mini-component, or no preview.
+- **Motion must be motivated.** Every animation answers "what does this communicate?" (hierarchy,
+  sequence, feedback, state change). "It looked cool" is not an answer. Honor `prefers-reduced-motion`
+  and collapse infinite, parallax, or scroll-hijack motion to static under it.
+- **One design system per project.** Do not mix component libraries in one tree.
+- **Real logos for social proof**, not text wordmarks; logo wall under the hero, logos only, no
+  category labels beneath them.
+- **Both color schemes from the start.** Design light and dark together and respect
+  `prefers-color-scheme`; never ship one mode by accident, and keep brand identity and contrast intact
+  in both.
+- **Viewport-stable layout.** Use dynamic viewport units for full-height sections so mobile browser
+  chrome does not cause layout jumps, and keep primary navigation on a single line at desktop width.
+
+## Tells (web)
+
+- Fake-div screenshots; hand-rolled decorative SVG instead of real imagery.
+- Unmotivated motion; infinite/parallax/scroll-hijack with no `prefers-reduced-motion` fallback.
+- Mixed component libraries in one tree.
+- Text wordmarks as social proof instead of real logos.
+- One color scheme shipped by accident; nav wrapping to two lines at desktop width.
+
+## Pre-flight slice (web)
+
+- [ ] Real images, no fake-div screenshots.
+- [ ] Motion motivated and reduced-motion honored.
+- [ ] One design system; real logos for social proof.
+- [ ] Light and dark both designed and tested; nav fits one line; full-height sections use dynamic
+  viewport units.

package/plugins/ca-pi/includes/anti-slop-design/typography.md

@@ -0,0 +1,51 @@
+# anti-slop-design · typography
+
+Load it whenever the medium renders text with chosen type (per the INDEX load map). The single
+most-tested tell is the wrong typeface chosen by reflex. The right choice depends on the medium.
+
+The rule is the **principle**, not the name: be able to say in one sentence why this face fits this
+brief. The specific faces named below are *current* examples of reflexes and reasonable reaches; they
+will age, and the "reach" lists are non-exhaustive starting points, not a closed menu to default to.
+Defaulting to a published alternative list just recreates the reflex with a new name.
+
+## 4.A Body text
+
+- **On screen / web UI:** humanist or geometric sans for body. Avoid the current LLM-default face
+  (`Inter`, at time of writing) *as a reflex*; it is acceptable when a neutral, system-standard feel is
+  explicitly wanted, or for accessibility-first contexts. Reasonable current reaches (non-exhaustive):
+  Geist, Outfit, Söhne, IBM Plex Sans, or a brand-appropriate face. The point is a face chosen for a
+  reason, not the specific name.
+- **In documents / print / PDF (reports, whitepapers, books):** a serif or a high-quality humanist
+  sans for sustained reading is **best practice, not a tell**. Serif body text in a long report is
+  correct typography. Pick a real text face (a Garamond, Source Serif, IBM Plex Serif, Charter, or a
+  clean humanist sans) over the Office defaults (Calibri, Times New Roman) chosen without thought.
+- **Measure:** 60-75 characters per line. Line-height 1.4-1.6 for body. Do not full-justify text that
+  produces rivers; left-align unless you control hyphenation.
+
+## 4.B Display / headings
+
+- **Web and slides default to sans display.** Sans display is not "boring"; it is default for the same
+  reason black is default in fashion. Reach: Geist Display, Cabinet Grotesk, PP Neue Montreal, GT
+  Walsheim, Inter Display.
+- **Serif display is allowed when justified** (genuinely editorial, luxury, publication, heritage, or a
+  brand that names a serif) and you can say why this serif fits this brand. When justified, **do not
+  default to the current LLM-favorite display serifs** (`Fraunces`, `Instrument Serif` at time of
+  writing). Other current options, non-exhaustive: PP Editorial New, Reckless Neue, Tiempos Headline,
+  Canela, Domaine Display. Pick for fit, do not just swap one default for another.
+- **Emphasis within a headline** uses italic or bold of the *same* family. Injecting one serif word
+  into a sans headline (or vice versa) is amateur.
+- **Italic descenders:** any italic word containing `y g j p q` needs at least `1.1` line-height and a
+  little bottom reserve so the descender is not clipped.
+
+## 4.C Hierarchy
+
+Build hierarchy with **weight, size, and space**, not with boxes, rules, and color on everything.
+Three levels are usually enough: section, subsection, body. If everything is emphasized, nothing is.
+
+## Tells (typography)
+
+- Inter as a reflex body face on screen.
+- Fraunces / Instrument Serif as the default display serif.
+- A foreign font injected for a single emphasized word.
+- Clipped italic descenders.
+- Full-justified body with visible rivers.

package/plugins/ca-pi/includes/author-tdd-workflow.md

@@ -0,0 +1,14 @@
+# Author TDD workflow — non-negotiable, fixed order
+
+Canonical for the author agents (`backend-author`, `frontend-author`). Loaded via each agent's
+Required Reading; the agent bodies do not restate it. This is the worker's execution order for
+one task — the tdd skill's phase gates govern the lane above it.
+
+Fixed order. Do not skip or reorder.
+
+1. **Write failing tests** — one test per obligation in the Phase 1 checklist
+2. **Confirm tests fail for the right reason** — run the test command from `tech-stack.md`; the failure must match the obligation, not an unrelated error
+3. **Write minimum implementation** — only enough code to make the failing tests pass; no extra scope
+4. **Run full suite** — every test green
+5. **Run lint and type-check** — both clean
+6. **Stage for commit** — only after steps 1–5 complete