Gemini Writer

Gemini Writer

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.

When to Use

• Book manuscript rewrites — when the author has annotations across multiple chapters and cross-chapter issues need to be resolved simultaneously
• Structural refactoring — moving sections between chapters, resolving redundancy, reordering content
• Voice-consistent rewrites — when you have a voice guide + source material + current draft and need a revision that respects all three
• Batch processing — rewriting multiple chapters one at a time, but with full manuscript context for each

When NOT to Use

• Quick edits to a single paragraph (just use Claude directly)
• Tasks requiring web search or real-time information (use deep-research instead)
• Tasks where the total context is under 100K tokens (Claude handles this fine)

Architecture

The script sends content to Gemini's generateContent API (not the Interactions/research API). Two modes:

Single-Pass Mode

Send everything, get one big output. Best for analysis tasks (redundancy audit, structural plan).

Batch Mode (Recommended for Rewrites)

Send full manuscript as context, but process ONE chapter at a time. This means:

• Gemini sees everything (all chapters, all annotations, all instructions)
• But it only outputs one chapter per call
• This keeps output quality high (no truncation) while maintaining full cross-chapter awareness

Workflow

1. Prepare the Prompt File

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.

2. Dry Run

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

3. Execute

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

4. Review

Compare outputs against originals. Gemini's rewrites should be treated as strong first drafts that need human editorial review.

Prompt Engineering Tips

1. 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."

2. Give voice samples. Include 2-3 paragraphs of the author's best writing and say "match this voice."

3. 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]."

4. Use the system instruction for invariants. Voice guide, source material rules, theological constraints — anything that applies to every chapter goes in --system.

5. Order matters. Put the most important instructions first in the prompt. Gemini, like all LLMs, attends more to the beginning.

Cost & Performance

• Model: gemini-3-pro-preview (falls back to gemini-2.5-pro-preview-06-05)
• Input cost: ~$1.25/1M input tokens, ~$10/1M output tokens (as of Feb 2026)
• Typical manuscript rewrite: 6 chapters × ~15K tokens each = ~90K input + ~60K output ≈ $0.75
• Time: 30-120 seconds per chapter depending on length
• Max output: 65,536 tokens per call (~50K words — more than enough for a chapter)

Script Location

.claude/skills/gemini-writer/scripts/gemini-writer.py

Requires: GEMINI_API_KEY env var (set in ~/.zshrc), httpx (auto-installed on first run)