sisiphamus/shakespeare

shakespeare は何をしますか？

Target: a suspicious human reader, not a detector.

Two modes. Default is generic human (category samples + rules). "Their voice" only if they give real samples. No samples → stay generic. Don't invent a personal voice.

Process

1. Discover (don't require answers)

Tell them: answer as much or as little as you want. More of their specifics = less generic.

Ask what's open (skip what they already said; topic-only is fine):

1. What (email, essay, post, paper, story…)
2. Who for — email: who→whom; essay: what the receiver expects at this stage
3. Topic + their actual take (what a generic piece would miss)
4. Outline if they have one
5. Notes / bullets / half-thoughts
6. Real examples, stories, numbers, quotes they want in (private grit — not invent-specific)
7. Anything they don't want

Core idea: temperature = their ideas. Private grit beats eloquence. Invented "vivid" still reads AI.

2. Sample category

Closest under samples/:

• academic — papers, formal argument
• narrative — essays, stories, long personal
• conversational — chatty notes, informal
• fun — humor / playful
• email — emails, short pro asks

Mixed → primary + maybe one adjacent file.

Read samples/INDEX.md, then 2–4 full files (email: several short ones).

3. Rules

Read all of: rules/craft.md, rules/ai-tells.md, rules/reader-psych.md (esp short external).

4. Personal voice (optional)

Ask if they want their voice. If no → generic.

If yes: where are samples? local path, paste, Drive/Docs MCP, Gmail sent. Pull a lot (5–20+ if you can). Prefer same register. Nothing available → say so, fall back to generic. Match rhythm/stance, don't copy phrases.

5. Outline then section write

Short outline, light OK (or just write if they said so). One section at a time. After each, one question for uniqueness (angle? real example? too formal? cut what they'd never say?). Use their lines. Editor of their thought, not replacement author.

Obey rules + match samples (and their voice if twin).

6. Review (required — no skip)

Full draft exists → re-check before return.

Prefer a fresh subagent with intake + samples used + personal samples + rules/* + full draft. No subagent: self-review only after re-reading rules/intake from disk, not draft memory alone.

Reviewer: kill AI signs, voice match if twin, reader psych for short external, quote offenders fix them, return clean draft. Offer another pass if they want.

Hard rules while writing

1. Frame first (samples + their ideas). "Sound human" is not a method.
2. Their specifics > your eloquence. Private grit only.
3. No assistant costume: no Great question, no prompt restates, no menus mid-draft.
4. No performed humanness: no I'll be blunt, fake lowercase, punchy fragment stacks. No default short declarative openers (bare 4–8 word S–V–O para hooks).
5. No interest-bridges. No "the part/section/bit X". No definitional punches ("That joining is fusion."). No fake-plain gloss (shows up as / falls out as heat/light).
6. No sealed product (stock vignette kit, no-deep-lesson bow, multi-beat coach memo when short works).
7. Vary sentence length; never three similar in a row.
8. Take a side when needed; leave loops open when maxim isn't needed.
9. Em dashes: avoid in external prose.
10. Don't invent their opinions/bio/facts.
11. Not a detector-evasion product.

Quick path (they dumped everything)

Infer category → 2–4 samples + rules → ask once personal voice? → draft (sections if long; short email can be one shot + review) → always review.

Files

SKILL.md (here), AGENTS.md (entry + must-check), README.md, LICENSE, rules/, samples/