For long-form writing, Gemini and Claude need mostly the same prompt: material first, request last, clear delimiters, and examples of the style you want. The differences are in defaults and details, such as Gemini's concise default answers and Claude's formatting tips. This article compares what each vendor's own documentation says, and I should be clear that it's a documentation comparison, not a head-to-head writing test.
Facts about models change quickly. Everything below reflects the vendors' pages as I read them in November 2026; confirm against the current docs before you commit to a workflow.
What do both vendors agree on?
Five core practices appear in both sets of documentation.
| Practice | Google (Gemini docs) | Anthropic (Claude docs) |
|---|---|---|
| Context first, request last | "Supply all the context first. Place your specific instructions or questions at the very end" | Put long documents near the top; queries at the end can improve quality by up to 30 percent in tests, especially on complex multi-document inputs |
| Delimiters | XML-style tags or Markdown headings; choose one and be consistent | XML tags with consistent names for instructions, context, examples and inputs |
| Examples | "Always include few-shot examples"; too many may cause overfitting | Include 3-5 relevant, varied, structured examples |
| Be specific | "Be precise and direct"; define ambiguous terms | "Be specific about the output format and constraints you want" |
| Critical rules prominent | Put role, constraints and output format in the system instruction or at the very start of the prompt | Use system prompts; give reasons behind instructions |
That overlap is why one prompt layout travels well, as discussed in can I reuse the same prompt in ChatGPT, Claude and Gemini? For how to order the parts of a chapter prompt, see prompt structure for turning an outline into a draft.
Where do they differ?
In defaults and emphasis, according to the docs.
| Topic | Gemini docs say | Claude docs say | What to do |
|---|---|---|---|
| Default verbosity | Gemini 3 models give direct, efficient answers by default; ask explicitly for a conversational or detailed response | Newer models calibrate response length to the task; docs recommend being specific about length and format | State the word target and the level of detail either way |
| Formatting control | Show the format with consistent few-shot examples | Tell it what to do, use XML indicators, and match your prompt style to the output style | Write the prompt in the style you want back |
| Prefilling the answer | Documents a "completion strategy" for formatting | Prefilled responses are not supported on Claude 4.6 and later (API error); use instructions, tags or structured outputs | For chat, use instructions; for API, check which models allow prefill |
| Strong emphasis | Avoid "unnecessary or overly persuasive language" | Newer models can over-trigger on aggressive language; soften "CRITICAL: You MUST" | Plain wording, with reasons |
The second row matters most for authors who want strict structure; see how to make Claude follow a strict output format for that.
What about context window and output length?
Both vendors now list 1M-token context windows on their model pages, but a big window doesn't mean you should paste a whole book. Anthropic's models page says 1M tokens is roughly 555k words on the current tokenizer, and lists a 128K-token maximum output for the synchronous API on its current models. Google's models page lists a 1M-token context for its flagship model, with a maximum output of 65K tokens and, for long decodes, up to 1M. These numbers change often and depend on the model, tier and API. The practical point holds regardless: research on long inputs (Liu et al., 2023) found that models often use information at the start and end of a long input better than information in the middle. So give each request one chapter's worth of material, even when the window could hold more.
A rough budget helps. A 1,500-word section is about 2,000 tokens at the common rule of thumb of 0.75 words per token (an approximation; it varies by tokenizer, and Anthropic notes its current tokenizer fits fewer words per token than older ones). Thirty such sections are around 60,000 tokens, which is well inside either window, but the quality benefit comes from focus, not capacity.
How do I compare them fairly on my own book?
Run the same chapter brief in both with identical inputs, then score them against your checklist. Here is a simple protocol (my suggestion):
- Pick one section from your outline and write one brief, with notes and a voice sample.
- Run it once per tool with the same wording, in a fresh chat each.
- Score each on five yes/no items: did it answer in the first two sentences, hit the length range, use only your notes, match the voice sample, and follow the format?
- Repeat for a second, different section so one lucky result doesn't decide.
- Note the cost and time too, since they differ by plan.
Scoring against your own checklist beats general reputation. The same habit of testing small changes appears in how long should a prompt be. If you'd rather not depend on one assistant's quirks, a workflow tool can hold the brief steady: Interactive eBook Builder separates planning, drafting and polishing into Agent A, B and C with editable instructions, so the plan and rules stay put while you try different drafting options.
How do I keep a long piece consistent across several requests?
Carry a short "state block" from request to request instead of the whole earlier text. Neither vendor's documentation prescribes a single method here, so this is my suggested pattern, and it works the same in both tools. After each section, update a block of about 150 words containing:
| Field | Example content |
|---|---|
| Done so far | Sections 1-4: why pricing matters; three methods; worked example; common mistakes |
| Terms fixed | "starter rate", "floor rate" (defined in section 2; use unchanged) |
| Promises made | Section 6 will give the pushback script |
| Voice reminders | Second person; average sentence about 16 words; no rhetorical questions in openers |
Paste the state block with the next section's notes and the voice sample, with the request last. It keeps terminology stable and avoids repeating earlier content, without the cost of re-sending 10,000 words. This is the handoff idea from prompt chaining, applied to a whole manuscript. Treat the block as a checklist as well: if section 6 doesn't deliver what the block promised, you'll catch it immediately.
In Interactive eBook Builder, the Chapter Plan written by Agent A Spec Builder plays a similar role, because each section is drafted against a plan that already lists what the other chapters cover.
Before you hit send
- Is your context first and your request last?
- Are the delimiters consistent (tags or headings, not a mix)?
- Did you state the word target and level of detail explicitly?
- Are there 2-5 varied examples with identical formatting?
- Did you test the same brief in both tools before choosing?
The takeaway: write one solid, portable prompt, then tune the few places where the tools differ. Choose by testing on your own chapter, not by reputation.