メニュー
Delegate large-context writing tasks to Gemini's 1M token window. Send an entire manuscript + detailed instructions in a single pass. Use for book rewrites, structural refactoring, redundancy analysis, and any writing task where the AI needs to see everything at once.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Session-end memory discipline. Detects which projects were touched, writes updates to WORKBENCH.md (volatile state) + skills + persistent memory, and proposes CLAUDE.md edits. Two-tier — obvious low-blast-radius writes are applied automatically and reported; evergreen/destructive changes ask first. Use when the user says "/wrap", "wrap up", "end of session", "let's close out", or when the session is clearly winding down.
When the user is writing the story portion of a launch video or demo and needs the framing where the founder pretends to BE the ICP experiencing the problem, then uses the product on-camera. Also use when the user mentions "demo story", "show don't tell", "day-in-the-life demo", "roleplay demo", "founder as customer", or "first person launch video". This is the story frame, not the full script.
When the user needs a 3-5 word opening hook for a product launch video — the first second of on-screen text or voice that stops the scroll. Also use when the user mentions "hook for my video", "first second", "opening line", "tagline", "one-liner for launch", or "catchphrase". Produces 10+ hook candidates, not one answer.
When a user wants to grow an X (Twitter) account from zero before a product launch, or asks how to get first followers, warm up the algorithm, hit ~500-1,000 followers, or prepare an account to make a launch video land. Also use when the user mentions "new X account", "warm up my Twitter", "first 1000 followers", "building in public strategy", "X growth", or "engagement before launch".
When the user is planning, scripting, or editing a product launch video for X (Twitter) and needs the structure. Also use when the user mentions "launch video", "demo video", "product launch on X", "60 second demo", "how to structure a launch", or "my launch video isn't working". Produces a beat-by-beat timing sheet, not copy.
When the user has X (Twitter) content that performed well and wants to relay it to LinkedIn 1-2 weeks later with reframing. Also use when the user mentions "repost to LinkedIn", "LinkedIn version of my tweet", "X to LinkedIn", "delayed repost", "LinkedIn for non-tech audience", or "LinkedIn relay". Also use when the user's ICP is non-tech and X is secondary — LinkedIn is the primary channel and this skill produces the content.
| name | gemini-writer |
| description | Delegate large-context writing tasks to Gemini's 1M token window. Send an entire manuscript + detailed instructions in a single pass. Use for book rewrites, structural refactoring, redundancy analysis, and any writing task where the AI needs to see everything at once. |
Leverage Gemini 2.5 Pro's 1M token context window for writing tasks that exceed what Claude can hold in a single context. The key advantage: Gemini can read the ENTIRE manuscript plus all annotations, voice guides, and editorial instructions in a single pass, then produce a rewritten chapter that accounts for everything at once.
The script sends content to Gemini's generateContent API (not the Interactions/research API). Two modes:
Send everything, get one big output. Best for analysis tasks (redundancy audit, structural plan).
Send full manuscript as context, but process ONE chapter at a time. This means:
Create a markdown file with detailed instructions. Two levels:
System instruction (optional, --system): Voice guide, style rules, constraints that apply to every chapter.
Prompt (--prompt): Specific task instructions — what to fix, what to preserve, what to move, what annotations mean.
Always start with a dry run to verify token counts:
python3 .claude/skills/gemini-writer/scripts/gemini-writer.py \
--prompt prompt.md \
--context manuscript/revised/ \
--dry-run
Batch mode (one chapter at a time):
python3 .claude/skills/gemini-writer/scripts/gemini-writer.py \
--prompt prompt.md \
--system voice-guide.md \
--context manuscript/revised/ \
--output manuscript/v3/ \
--batch
Single chapter:
python3 .claude/skills/gemini-writer/scripts/gemini-writer.py \
--prompt prompt.md \
--context manuscript/revised/ \
--output manuscript/v3/ \
--batch --chapter 01-introduction.md
Compare outputs against originals. Gemini's rewrites should be treated as strong first drafts that need human editorial review.
Be explicit about annotations. Tell Gemini exactly what {} means: "Text in curly braces {} are the author's editorial comments. Execute these instructions, don't reproduce them."
Give voice samples. Include 2-3 paragraphs of the author's best writing and say "match this voice."
Specify what to preserve. Gemini will rewrite aggressively unless you tell it what's working. Say "The following sections are strong and should be preserved with minimal changes: [list]."
Use the system instruction for invariants. Voice guide, source material rules, theological constraints — anything that applies to every chapter goes in --system.
Order matters. Put the most important instructions first in the prompt. Gemini, like all LLMs, attends more to the beginning.
gemini-3-pro-preview (falls back to gemini-2.5-pro-preview-06-05).claude/skills/gemini-writer/scripts/gemini-writer.py
Requires: GEMINI_API_KEY env var (set in ~/.zshrc), httpx (auto-installed on first run)