lexisim:design

What this skill does

# Lexisim — Writing Style Designer

Design a writing voice through deep, structured questioning. Either build one from scratch through guided preference exploration, or start from reference samples and refine from there. The output is a ready-to-use writer skill and writing assistant agent tuned to the designed voice.

## Mode Check

If plan mode is active, exit it now using ExitPlanMode before starting this workflow. All lexisim commands run in execute mode.

## Why This Process Works

Shallow instructions like "write casually" or "be professional" collapse into generic output. Stylometry research shows that writing voice lives in the interplay of independent axes — sentence length variance, vocabulary register, punctuation fingerprints, information density, emotional temperature, and transition patterns. By probing each axis independently and then testing for coherence, this questionnaire produces voice profiles that are richer and more distinctive than any single-paragraph style description could achieve.

## Reference Files

All reference files live within this skill's directory. `${CLAUDE_SKILL_DIR}` resolves to the directory containing this SKILL.md. Use these exact paths — do not search or glob.

- **Question bank**: `${CLAUDE_SKILL_DIR}/references/question-bank.md`
- **Reading level calibration**: `${CLAUDE_SKILL_DIR}/references/reading-level-calibration.md`
- **AI tells catalog**: `${CLAUDE_SKILL_DIR}/references/ai-tells.md`
- **Anti-voice catalog**: `${CLAUDE_SKILL_DIR}/references/anti-voice-catalog.md`
- **Writer skill template**: `${CLAUDE_SKILL_DIR}/assets/writer-skill-template.md`
- **Writer agent template**: `${CLAUDE_SKILL_DIR}/assets/writer-agent-template.md`

## Workflow Overview

```
Phase 1:  Entry Path & Setup            (~2 min)   → Choose path, name profile, set scope
Phase 2:  Reference Analysis             (~3 min)   → Emulate path only — analyze pasted samples
Phase 3:  Audience Persona               (~3 min)   → Who reads this, what they need
Phase 4:  Reading Level Calibration      (~4 min)   → 8 show-don't-tell comparison pairs
Phase 5:  Voice Character                (~4 min)   → Emotional temperature, pacing, structure
Phase 6:  Anti-Voice Definition          (~3 min)   → What this voice is NOT
Phase 7:  Implicit Influence Probe       (~1 min)   → Design path only — name influences
Phase 8:  Uncanny Valley Detection       (~3 min)   → Test paragraphs, collect reactions
Phase 9:  Data Extraction                (auto)     → Build structured handoff document
Phase 10: Analysis & Generation          (~1 min)   → Sub-agent analyzes + generates files
```

Total: ~20-25 minutes for a comprehensive voice profile.

## Isolation Rule

Work ONLY with data collected through this questionnaire. Do NOT:
- Search for writing samples on the user's machine
- Read other skills, profiles, or configuration files
- Look for documents, blog posts, or archives
- Reference any external material not collected in this session

Everything needed to build the voice profile comes from the user's responses to the prompts below.

---

## Phase 1: Entry Path & Setup

Present questions Q1-Q3 from the question bank (`references/question-bank.md` → Phase 1).

- **Q1** determines the entire flow: "Design from scratch" skips Phase 2 and includes Phase 7. "Emulate a reference" includes Phase 2 and skips Phase 7.
- **Q2** (free text): Store as `PROFILE_NAME`. Convert spaces to lowercase kebab-case.
- **Q3**: Store as `SCOPE`. Resolve output paths — project scope writes to `.claude/skills/` and `.claude/agents/` in the current working directory; user scope writes to `~/.claude/skills/` and `~/.claude/agents/`.

---

## Phase 2: Reference Analysis (Emulate Path Only)

_Skip this phase entirely if `ENTRY_PATH` is "Design from scratch". Proceed directly to Phase 3._

Present questions Q4-Q6 from the question bank (`references/question-bank.md` → Phase 2).

When the user provides samples (Q4), perform lightweight stylometric analysis on each:
- Count sentences and calculate mean/stdev of sentence length
- Note vocabulary register (formal/informal/technical/conversational)
- Identify punctuation patterns (em-dash frequency, parenthetical rate, semicolons, exclamation density)
- Measure sentence length variance (burstiness)
- Note contraction rate, first-person frequency, opening patterns

Store both the verbatim samples and the analysis results.

Q5 reveals which dimensions of the reference the user is drawn to versus which are incidental. Use it to weight the Q4 analysis — amplify the dimensions the user calls out, treat uncalled dimensions as secondary.

If the user selects "Add more influences" in Q6, loop back to Q4 and Q5 for additional references. Cap at 3 total reference sets to avoid dilution.

---

## Phase 3: Audience Persona

Present questions Q7-Q10 from the question bank (`references/question-bank.md` → Phase 3).

After collecting all answers, synthesize a brief audience persona summary and confirm
with the user: "Here's who we're writing for: [summary]. Sound right?"
Use AskUserQuestion with options: "Yes, that's right" / "Close, but let me adjust".

Cross-reference Q8 (knowledge level) with Q9 (reader frustrations) — frustrations
often reveal implicit knowledge level assumptions. If the user wants to adjust, collect the correction as free text and update the synthesized persona before proceeding.

---

## Phase 4: Reading Level Calibration

Read the reference file `references/reading-level-calibration.md` for the 8 example pairs.

Present Q11-Q18 from the question bank (`references/question-bank.md` → Phase 4). For each pair, show both Version A and Version B in full from the calibration file, then present the AskUserQuestion.

Each dimension is tested twice. Average the two responses per dimension:
- Both "Version A" = strong preference for the simpler end
- Both "Version B" = strong preference for the complex end
- Split or "Somewhere between" = moderate/middle preference

Build a 4-axis reading level profile:
- **Vocabulary**: simple <-> technical
- **Syntax**: short/direct <-> nested/complex
- **Density**: few ideas per passage <-> many ideas per passage
- **Scaffolding**: high (explains why) <-> low (just the facts)

### Cross-Reference with Audience Persona

After all 8 pairs, cross-reference the reading level profile with the audience persona from Phase 3. Flag contradictions and surface them to the user. Common tensions:

- **Expert audience + high scaffolding preference**: "You're writing for experts but prefer writing that explains the 'why' behind things. Should we assume expertise on the domain but still provide reasoning for decisions and recommendations?"
- **Beginner audience + low scaffolding preference**: "You're writing for beginners but prefer writing that gets straight to the facts. Should we define terms but skip the hand-holding on concepts?"
- **Scanning readers + high density preference**: "Your readers skim, but you prefer information-dense writing. Should we use dense prose but with strong structural signposts (headings, bold key terms, front-loaded paragraphs)?"

For the **Emulate path**: also compare these preferences against the empirical analysis of reference samples from Q4. If the reference samples sit at a different reading level than the stated preferences, note this — it indicates the user wants to adjust the reference rather than replicate it exactly.

---

## Phase 5: Voice Character

Transitional message:

> Now let's get into the personality of this voice — how it sounds, how it moves, how it opens.

Present questions Q19-Q24 from the question bank (`references/question-bank.md` → Phase 5).

For Q21 (Sentence Structure), this is a show-don't-tell question — present the three preview paragraphs from the question bank before the AskUserQuestion options.

After collecting all voice character answers:

> Great, that gives me a clear picture of the voice character. Now let's define what this voice is NOT — someti