Stats
Actions
Tags
From ttrpg-editorial
Audit AI-generated long-form fiction (anthologies, collections, multi-chapter novels) for recurring stylistic tics — structural patterns, near-identical phrases, syntactic constructions, and inter-story duplicates — then generate surgical alternatives per occurrence. Use whenever a writer suspects their AI agent is repeating itself, when editing multi-story collections before publication, when reviewing a draft for variation and freshness, or when asked to "check for tics", "find repetitions", "audit the style", "make the stories feel more distinct", or "prevent the AI from writing the same thing twice". Also use proactively when *generating* new content that must not repeat patterns already identified in a project. Two modes: AUDIT (analyze existing text → tic inventory + alternatives) and GUARD (generation time → actively avoid known tic patterns when writing new content).
How this skill is triggered — by the user, by Claude, or both
Slash command
/ttrpg-editorial:editorial-tic-auditorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Prevents AI-generated fiction from developing the verbal, structural, and syntactic habits
Prevents AI-generated fiction from developing the verbal, structural, and syntactic habits that flatten an anthology into a single voice. Works in two modes:
Seven categories. Every tic belongs to at least one. For a full catalog of known AI defaults
per category, load references/tic-library.md during audit setup.
Recurring scene shapes or narrative gestures: how stories begin, resolve, or end. Examples: protagonist always climaxes into private reflection; stories always close on a direct address to the reader; a specific vignette type (masturbation, prayer, letter) appears in every piece.
Detection signal: Step back from prose level. Ask: if I mapped only the scene shape — what happens, in what order — would multiple stories share the same skeleton?
Near-identical phrases or sentence fragments repeated across stories, often with minor
cosmetic variation. The content changes; the phrasing doesn't.
Examples: "Your hip. The angle. Let me—" / "Not yet" as a standalone spoken line.
Detection signal: Collect every line of dialogue or action that feels functional (advances the scene mechanically). Run them against each other. If two belong to different stories but could swap without the reader noticing: phrasal tic.
Recurring grammatical constructions that the AI defaults to under pressure: corrections,
contrasts, intensifiers. The specific words change; the sentence shape doesn't.
Examples: "Not X — Y." / "Not X, exactly, but Y." / "Less X than Y." / the em-dash
pivot as default for emotional revelation.
Detection signal: Strip content words and look at the skeleton. "Not pleasure, exactly, but release." and "Not arousal. Recognition." share a skeleton even if they mean different things.
Two stories share a specific motif, detail, or sensory image at the same level of particularity. Distinguishable from thematic resonance by specificity: two characters having ink-stained fingers is a doublon; two characters being defined by their craft is theme.
Detection signal: List every concrete sensory detail (taste, smell, texture, color, sound). Look for identical entries across stories. The more specific the detail, the more certain the doublon.
Recurring metaphor clusters — the AI defaults to a small pool of base images to signal emotion. Each individual metaphor may be original; the problem is that they all come from the same drawer. Examples: heat/fire/warmth as the default arousal cluster; gravity/pull/magnetic as the default attraction cluster; drowning/flooding/wave as the default overwhelm cluster.
Detection signal: Extract every metaphor or image used for emotion. Group by vehicle (what the thing is compared to). If the same vehicle dominates — heat in 6 of 8 arousal descriptions — it's a metaphoric tic, even if no two sentences are identical.
Body parts acting with autonomous agency: they find, trace, catch, tighten,
part — detached from the character's intention. The AI uses anatomical autonomy as a
shortcut to show involuntary emotion.
Examples: "her hands found his face" / "something caught in her throat" /
"her lips parted" as default arousal signal / "his eyes found hers across the room".
Detection signal: Scan every sentence where a body part is the grammatical subject. Ask: is the body part doing something the character chose to do, or is the body part acting on its own? Autonomous body parts are the tic. Count per story; flag if > 4.
Sentence-level rhythm defaults: the AI reaches for the same beat patterns when it needs to signal intensity, transition, or climax. Examples: the one-word paragraph used as a hammer; the three-noun fragment sequence; the slow-motion passage where every micro-movement is adverb-weighted.
Detection signal: Read aloud. If the same rhythm pattern appears at every emotional peak, the rhythm is doing the work that the content should be doing. The pattern itself becomes predictable — the reader anticipates the beat before the meaning arrives.
Short phrases (2–5 words) or individual lexical items that appear at high absolute frequency across the corpus. These are invisible to qualitative reading — no single occurrence is wrong, but the cumulative count creates a rhythm the reader unconsciously registers as repetition.
Two sub-types:
"for a heartbeat" / "her voice was" /
"as if she" / "felt the""witness" / "confession" / "jasmine" / "ancient"Detection signal: This category cannot be reliably detected by qualitative reading. It requires a word-frequency count. When using a script or tool is possible, run one. When working manually, extract every occurrence of the P-scan targets plus any word that "feels" frequent, count them, and build a frequency table before judging any individual instance. Do not skip this scan. F-tics are the category most likely to survive all other scans and still flag in editorial review.
Thresholds (prose fiction, 10-story anthology ~200 pages):
Critical distinction — concentration vs. distribution: A word appearing 15× evenly across 10 stories is different from 15× concentrated in 2 stories. Flag both patterns, but treat them differently:
Recurring grammatical patterns in the first sentence of each story. These operate at the table-of-contents level — a reader who glances at story openings detects the pattern before they've read a word of narrative.
Common AI defaults for story openings:
"I had [past perfect for duration]" — "I had worn my cassock for twenty-three years""She had [past perfect for state]" — "She had my wife's face""I recognized X before I recognized Y" — contrast opening"She was [adjective]" — physical description openingThreshold: Flag if > 2 stories share the same opening grammatical structure. Target: Fewer than 2 stories with the same first-sentence skeleton. Detection signal: List only the first sentence of each story. Strip content words. Compare the grammatical skeletons. If two are identical, that is an O-tic.
Recurring game design defaults: the AI applies the same mechanical expressions, stat conventions, encounter structures, and NPC templates throughout a supplement. Unlike prose tics, mechanical tics have functional consequences: the GM stops reading carefully because every section is structurally predictable; players stop being surprised because every challenge follows the same calibration logic. Examples: every moderate check is DC 15; every major NPC has exactly one secret; every faction ends with exactly three hooks in the same format.
Detection signal: Strip the fiction. Look at the numbers and structure alone. If the mechanical skeleton of Section 4 is indistinguishable from Section 9 — same DCs, same action economy, same hook count, same template order — it's a mechanical tic.
Two sub-types:
Receive the corpus: a list of stories/chapters. If they are not in context, request them. Note for each: title, character/POV, setting, core thematic arc (1 sentence).
If the user has already done partial tic identification (as with the What the House Keeps document), import their findings directly and move to Step 3.
For each tic category, work through the full corpus:
S-scan: Map scene shapes. Note: opening gesture / escalation method / climax type /
closing gesture. Flag any skeleton appearing ≥ 2 times.
Cross-check against S-DEFAULT entries in references/tic-library.md.
P-scan: Extract all:
references/tic-library.md.V-scan: Extract all sentences using:
Not X — Y / Not X, but Y / Not X, exactlyor not only, or rather, or perhaps)truly, deeply, profoundly, utterly)something that might have been X / what felt like X)she found herself doing X / she heard herself say)
Count per story. Flag any structure appearing > 2× in a single story or > 3× across corpus.
Cross-check against V-DEFAULT entries in references/tic-library.md.D-scan: Build a sensory inventory per story. Cross-reference. Flag identical specifics.
M-scan: Extract every metaphor used to signal emotion. Group by vehicle (the image
used, not the tenor). Flag any vehicle cluster dominating > 3 occurrences across the corpus.
Cross-check against M-DEFAULT clusters in references/tic-library.md.
A-scan: Find every sentence where a body part is the grammatical subject. Tally per
story. Flag stories with > 4 autonomous body part constructions.
Cross-check against A-DEFAULT entries in references/tic-library.md.
R-scan: Identify peak-intensity passages in each story. Map their rhythm. Note: one-word paragraphs / fragment sequences / slow-motion micro-movement passages / "And then." paragraph breaks. Flag recurring peak-beat patterns across stories.
F-scan — MANDATORY, run before all other scans:
This scan must produce a frequency table before the qualitative work begins. If a script is available, use it. If working manually, do this explicitly:
Step 1 — Phrase frequency: Extract every 2–5 word sequence that appears more than
once. Build a ranked list. Flag anything appearing > 5× across the corpus or > 2× in
one story. Priority targets (check these first — they are confirmed AI defaults):
"for a heartbeat" / "her voice was" / "as if she" / "felt the" /
"as if he" / "seemed to" / "something in her" / "for a moment"
Step 2 — Thematic word frequency: Count every word carrying thematic or aesthetic
weight. Flag anything > 15× corpus. Priority targets:
witness / confession / surrender / tender / ancient / sacred /
shatter / fracture / undone / ache — plus any word in the project title
or a character name that could migrate into prose
Step 3 — Atmospheric word frequency: Count sensory/setting defaults. Flag > 10×:
jasmine / candlelight / velvet / stone / smoke / shadow / silk /
silver / gold — note concentration by story
Step 4 — Framing verb frequency: These are structurally weak and accumulate
invisibly. Count per story, flag if > 5 per story:
"felt the" / "her voice was" / "as if she/he" / "seemed to" /
"looked like" / "appeared to"
Cross-check results against F-DEFAULT entries in references/tic-library.md.
O-scan: Extract the first sentence of each story/chapter. Strip content words.
Compare grammatical skeletons. Flag any skeleton appearing ≥ 2 times.
Cross-check against O-DEFAULT patterns in references/tic-library.md.
G-scan (RPG supplements only):
G-S structural scan: Extract the template skeleton of each entry type (NPC, location, faction, encounter, magic item). Compare across entries of the same type. Flag when ≥ 70% of entries share the same section order and section count.
G-N numerical scan: Build a DC distribution table, a CR distribution table, and a damage expression inventory. Flag:
references/tic-library.md.Produce a Tic Register in this format for each tic found:
## TIC [CATEGORY][NUMBER] — [Short Name]
**Type:** S / P / V / D
**Occurrences:** [n] across [list of stories]
**Effect (when it works):** [1 sentence — what is the tic achieving?]
**Problem:** [1 sentence — why repetition neutralizes it]
| Story | Current Instance | Verdict |
|-------|-----------------|---------|
| Title | Quoted passage | KEEP / REPLACE / REMOVE |
**Replacement Principles:**
[2–4 concrete strategies that achieve the same effect through different means]
**Distribution Rule:**
[Which story gets to keep the original; what the others must find instead]
After all tics are registered, produce a summary table:
| Tic | Action | Stories Touched |
|---|---|---|
| [ID] — [Name] | KEEP × n / REPLACE × n / REMOVE × n | [list] |
Close the report with 1–3 sentences articulating the editorial logic specific to this corpus — what the tics reveal about the AI's defaults, and what the revision is protecting.
Example: "Each device must feel discovered in this story — not imported from a template. Test: if a phrase functions identically in three different stories, it belongs to none."
AUDIT produces a diagnosis. REVISION executes the repairs.
Critical rule: The auditor alone cannot write the replacements correctly. Replacing a passage in literary erotica requires the voice of the story, the emotional register of the moment, and the character's psychology — not a generic editorial suggestion. Always load the appropriate writer skill before executing replacements.
Choose based on the project's readership and register:
Literary erotica for female/romantasy readership (What the House Keeps and similar):
→ Load erotica-fantasy-feminine — handles interiority, female desire as primary,
the romantasy register, and explicit content that serves emotional truth.
This skill contains the Tic Revision Mode with voice anchors for each courtesan.
General literary erotica (English, mixed readership):
→ Load erotica-fantasy-en + intimate-noir-prose for voice calibration
French erotica:
→ Load erotica-fantasy-fr
Non-erotica prose fiction: → Load the appropriate genre writer skill for the project
The auditor identifies and briefs. The writer skill executes.
For each passage to replace, the auditor produces a brief in this format:
## BRIEF [TIC-ID] — [Story Title] — [Approximate Location]
**Tic category:** [F-P / F-W / V / P / S / etc.]
**What the tic is:** [exact phrase or pattern — e.g., "for a heartbeat"]
**Exact passage to replace:**
"[verbatim text — complete sentence(s) including the tic]"
**Word budget:** ± [n] words — [expand / contract / same length]
**What this passage does narratively:**
[1 sentence: its function — e.g., "marks a micro-pause of hesitation before she speaks"]
**Emotional register at this moment:**
[1 sentence: character state — e.g., "Severine has just yielded control for the first
time; she is raw, surprised by herself, not yet integrated"]
**Character voice signature:**
[3-5 words: e.g., "military precision, controlled, newly permeable"]
**Surrounding context:**
BEFORE: "[sentence immediately before the passage]"
AFTER: "[sentence immediately after the passage]"
**The tic to avoid:** [exact phrase or construction — do not reproduce]
**Direction (not prescriptive):**
[concrete suggestion from the auditor — e.g., "replace the time marker with a physical
anchor specific to this moment — something in the dungeon, her own body, his breath"]
Use when writing new content for a project that already has a tic registry.
The registry can come from:
Before generating any scene, internalize the active constraints as a negative space:
FORBIDDEN STRUCTURES (this session):
- S: [list of forbidden scene shapes]
- P: [list of forbidden phrases, verbatim or near-verbatim]
- V: [list of forbidden syntactic skeletons]
- D: [sensory details already used; do not repeat]
- M: [metaphor vehicles already dominant; find new vehicles]
- A: [body-part autonomy budget: max 4 per story; track count]
- R: [rhythm patterns already used at climax; vary the beat]
- F-P: [phrases at or near corpus threshold; avoid entirely this session]
- F-W: [thematic words at saturation; substitute or remove]
- O: [opening skeleton already used ≥ 2 times; this story must open differently]
- G-S: [RPG only — entry templates that are locked; this entry must use a different structure]
- G-N: [RPG only — DC/CR/damage values already saturated; use different numbers]
ASSIGNED VARIANTS (if this story has been assigned a specific variant):
- [e.g., "Use Option A: slowdown without withdrawal" for delayed orgasm]
- [e.g., "Closing gesture must be object-based, not declarative"]
- [RPG e.g., "This NPC entry has no 'secret' field — replace with a public contradiction"]
- [RPG e.g., "No DC 15 checks in this location — calibrate at 12 or 18"]
After drafting any scene, story, or supplement section:
Not X constructions; count per story (max 2)If a violation is found: do not preserve it as a stylistic choice. Revise immediately using the replacement principles from the register.
These apply across all tic types when the register doesn't specify:
For structural tics:
For phrasal tics:
For syntactic tics:
Not X — Y is the default: try assertion (It was Y.) or question (Was it X?)or not only): try omitting the correction — let the first
statement carry ambiguitytruly felt): remove the intensifier entirely.
The bare verb is almost always stronger.For doublon tics:
For metaphoric tics:
For anatomical tics:
"she touched his face" replaces "her hands found his face""She was gentler than he expected" — no body part namedFor rhythmic tics:
A tic report is good when:
A generation pass is good when:
AI agents develop tics because:
"Not yet" achieved tension once, the model
reuses the shape that produced tensionsomething, somehow, almost, what might have been — the model produces these when it is uncertain how much to commit. They are
emotional equivocation, not literary ambiguity"Not X — Y" for contrast. It is the path of
least resistance, not a calibration decision"for a heartbeat"
on page 3 and again on page 47 without any awareness that it has done so. This is why
F-category tics survive every qualitative scan: the model cannot detect its own
frequency patterns without an external count. Always run the F-scan with a count,
not a reading."I had [duration/state]". It feels like a choice; it is a prior.The auditor's job is to map the schema the AI has fallen into, then break it — not by forbidding entire categories, but by requiring specificity that forces the model off the default track.
Load references/tic-library.md during AUDIT Step 2 for a full catalog of known AI
default tics, organized by category (S/P/V/D/M/A/R/F/O/G), with detection examples and
standard replacement strategies. This accelerates the scan phase significantly.
The library includes:
Script for F-scan (use when filesystem is available):
from collections import Counter
import re
with open('corpus.txt') as f:
text = f.read().lower()
# Phrase frequency (2-4 words)
words = re.findall(r'\b\w+\b', text)
for n in (2, 3, 4):
ngrams = [' '.join(words[i:i+n]) for i in range(len(words)-n+1)]
top = Counter(ngrams).most_common(40)
print(f"\n--- {n}-grams ---")
for phrase, count in top:
if count >= 4:
print(f" {count:3d} {phrase}")
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub camauger/ludomancien-skills --plugin ttrpg-editorial