Walter MCP：進階用戶的 5 個提示詞

Most people using Walter Writes AI's MCP server never get past the basic version: paste a paragraph, ask Claude to humanize it, done. That works, but it leaves most of the integration unused.

Once Walter is connected as an MCP server, Claude has three tools it can call, and it can call them in sequence, conditionally, and across many inputs. In Claude Code, it can also read the drafts from your repo and write the results back. That turns a one-off paraphrase into something closer to a content pipeline: detect first, humanize only what needs it, lock the terms that must survive, and report per file.

This post covers the setup in two minutes, then five prompts that use it properly. Each prompt is in a code block so you can copy it straight into Claude or Claude Code.

Table of Contents

1. What You're Working With: Tools vs Skills
2. Setup in Two Minutes
3. Prompt 1: The Detect-and-Humanize Chain
4. Prompt 2: The Keyword-Locked Humanize
5. Prompt 3: Agency QC Over a Folder of Drafts
6. Prompt 4: The Local SEO Multiplier
7. Prompt 5: Programmatic SEO From a CSV
8. Bonus: The Repurposing Cascade
9. Stacking Prompts and Watching the Tool Calls
10. FAQs

1. What You're Working With: Tools vs Skills

Two separate things do the work here, and mixing them up is the most common reason a prompt doesn't behave as expected.

The MCP server provides the tools. Walter's server exposes three:

• walter_humanize rewrites AI-generated text to read naturally, with adjustable rewrite strength. It preserves keywords, named entities, URLs, and document structure, and returns a preservation report.
• walter_detect returns a 0 to 100 AI score, a verdict (likely human, unclear, or likely AI), and the high-risk spans behind the score. Explanations can be turned on for detailed writing advice.
• walter_batch_humanize processes several texts in one call with shared defaults and per-item overrides. If one item fails, the others still return, with per-item reports and a batch summary.

Skills provide the process. Walter publishes 12 free, MIT-licensed skill files in github.com/walterwritesai/walter-skills. Each one is a SKILL.md that tells Claude how to run a specific workflow: when to humanize, which mode to use (light, balanced, or aggressive), which keywords to lock, and what the summary should look like. The four used in this post are walter-agency-qc, walter-local-seo, walter-programmatic-seo, and walter-content-repurposer.

Without the server, a skill gets followed but can't actually humanize or detect anything. Without the skills, the server works but you re-explain your process every session.

2. Setup in Two Minutes

Claude Code. Add the server with HTTP transport, scoped to your user so it's available in every repo:

claude mcp add walter-writes https://mcp-server.walterwrites.ai/mcp -t http --scope user

Then run /mcp inside Claude Code, select walter-writes, choose Authenticate, and sign in to your Walter account in the browser. Walter uses OAuth, so there's no API key to paste.

Claude on the web or desktop. Add Walter as a custom connector using the same URL, then sign in when prompted. Walter's docs have step-by-step guides for Claude web, desktop, mobile, ChatGPT, Gemini CLI, and Open WebUI.

Skills. Install all 12 with one command:

npx skills add walterwritesai/walter-skills

Or just one:

npx skills add walterwritesai/walter-skills --skill walter-agency-qc

For Claude on the web, open the skill's folder in the repo, copy its SKILL.md, and paste it into your project instructions.

If you get stuck on setup, waltermcp.com is a community-run guide with walkthroughs. It isn't an official Walter Writes AI resource, so treat Walter's own docs and GitHub repo as the source of truth if the two ever disagree.

Quick check: ask Claude to run walter_detect on any paragraph. If you get a score and a verdict back, the tools are live. If you get a rewrite with no score, Claude is paraphrasing on its own and Walter isn't connected yet.

3. Prompt 1: The Detect-and-Humanize Chain

The biggest inefficiency in how people use a humanizer is running the whole document through it, including the parts that were never going to get flagged. This prompt makes detection the first step and humanization the second, applied only where it's needed.

Use the walter-writes MCP server.
1. Run walter_detect on the draft below with explanations on.
2. Humanize only the high-risk spans it returns, using balanced mode.
   Leave every other sentence exactly as written.
3. Run walter_detect again on the full result.
Return: the final text, the before and after scores, and a list of
which sentences changed.

[paste draft]

Why it works: walter_detect returns the specific spans behind the score, not just a number, so Claude knows exactly which sentences to send to walter_humanize. Your own phrasing survives in the sections that were already fine, you use fewer words from your plan, and you get a clean changelog instead of a full rewrite you have to diff yourself.

In Claude Code, point it at a file instead of pasting:

Run the detect-and-humanize chain above on @drafts/launch-post.md.
Write the result back to the same file and print the changed sentences.

4. Prompt 2: The Keyword-Locked Humanize

Humanizers are designed to change wording, which is a problem when your draft contains a product name, a statistic, or a technical term that has to survive word for word. Walter preserves keywords, entities, and URLs by default, but listing the critical ones explicitly removes any guesswork.

Use walter_humanize on the product description below in balanced mode.
Preserve these exact terms with keyword preservation:
- "Acme Sync Pro"
- "47 percent faster"
- "ISO 27001 certified"
- every URL
Rewrite everything else freely for tone and sentence rhythm.
After humanizing, show me the preservation report and confirm each
locked term appears unchanged.

[paste description]

Why it works: Be specific. "Don't touch the numbers" is weaker than listing the exact figures, because a vague instruction leaves room for interpretation. Asking Claude to confirm each locked term against the preservation report turns "I hope it kept my keywords" into a check you can see.

5. Prompt 3: Agency QC Over a Folder of Drafts

If you QC client work regularly, the walter-agency-qc skill gives Claude a fixed checklist: run detection, humanize in balanced mode if the score is above 30, re-check, verify keywords, flag thin content (under 300 words) or missing H2 structure, and output a summary with a publish recommendation.

In Claude Code, that checklist can run across a whole folder:

Using the walter-agency-qc skill, QC every .md file in @content/client-drafts/.
Target keyword for each file is in its front matter under `keyword:`.
For each file:
- run the full QC checklist
- write humanized versions to content/client-drafts/qc/ with the same filename
Then print one summary table: file, original score, final score,
keyword status, word count, flags, recommendation.

Why it works: The skill keeps the process identical for file one and file twenty, and the output format stays consistent, which matters when the summary goes to a client or a junior editor. Writing results to a separate qc/ folder keeps your originals untouched, so you can diff before you accept anything.

6. Prompt 4: The Local SEO Multiplier

Multi-location service pages are one of the fastest ways to produce near-duplicate content, because the usual approach is a template with the city name swapped. The walter-local-seo skill is built to avoid that: it requires local context in each page, preserves the "[service] in [city]" phrase in the right places, humanizes in balanced mode, and flags pages that are too similar to each other.

Using the walter-local-seo skill, adapt the base service page in
@pages/emergency-plumbing.md for these cities:
Austin, Denver, Raleigh, Tucson, Columbus.
Each page needs at least one real neighborhood or landmark and must read
like it was written locally, not templated.
Run walter_detect on each page individually.
Save each page to pages/locations/<city>.md and print the summary table:
city, word count, detection score, keyword status, similarity flags.

Why it works: "Individually" is the key word. It's easy to optimize for an average across five pages while one still reads as obviously templated. A per-page score and the skill's similarity flags catch that before you publish.

7. Prompt 5: Programmatic SEO From a CSV

This is where Claude Code really pays off. The walter-programmatic-seo skill generates one page per CSV row using a fixed template (an H1 from the row, an intro, three H2 sections, and three FAQs), humanizes each page in balanced mode, and locks the H1 phrase with keyword preservation. It also rotates sentence openings and H2 order to avoid duplicate patterns, and it flags bad rows instead of stopping the batch.

Given a CSV like this:

h1,city,service
Roof Repair in Austin,Austin,roof repair
Roof Repair in Denver,Denver,roof repair
Gutter Cleaning in Austin,Austin,gutter cleaning

Run:

Using the walter-programmatic-seo skill, generate a page for every row
in @data/pages.csv.
Lock each row's h1 exactly with keyword preservation.
Save each page as pages/generated/<slugified-h1>.md.
When finished, print the summary table: row, word count, detection score,
H1 preserved (yes/no), flags. Don't stop for a bad row; list it at the end.

Why it works: The skill handles structure and anti-duplication, Walter handles the humanizing and the per-page score, and Claude Code handles the files. Detection risk resets per page, so a per-row score is the only number that matters.

8. Bonus: The Repurposing Cascade

One long post can become a LinkedIn post, an X thread, a newsletter, and an Instagram caption. The walter-content-repurposer skill handles each format's length and tone, humanizes each one in light mode so it reads casually and natively, and runs detection on every output.

Using the walter-content-repurposer skill, turn @blog/pillar-post.md into
a LinkedIn post, an X thread, an email newsletter, and an Instagram caption.
Save them to blog/repurposed/ and print the summary table:
format, length, detection score.

Each derivative gets its own score because, again, detection risk is per piece, not per source article.

9. Stacking Prompts and Watching the Tool Calls

Once these feel natural, combine them. A detect-and-humanize chain with locked keywords is a perfectly good single prompt:

Run the detect-and-humanize chain on @drafts/pricing-page.md, but keep
"Acme Sync Pro", "$49/month", and "14-day free trial" exactly as written.

A few habits that make these prompts reliable:

• Name the server and the tool. "Use the walter-writes MCP server to run walter_detect" removes any chance of Claude paraphrasing on its own.
• Watch the tool calls the first few times. In Claude Code you can see each call to walter_detect and walter_humanize as it happens. If a step you expected doesn't appear, tighten the prompt.
• Write outputs to a separate folder. It keeps originals safe and makes git diff your review step.
• Detect before you humanize. It's cheaper on your plan's limits and keeps more of your original voice.

Usage runs under your Walter workspace's plan, which sets limits like daily word volume and batch size. If a long batch stops partway, that's usually the reason; split it and continue.

10. FAQs

Do I need Walter's API for these prompts?
No. Everything here runs through the MCP server and works as normal messages to Claude or Claude Code. The REST API is only for calling Walter from your own code.

What's the difference between the MCP server and Walter Skills?
The server gives Claude the tools: walter_humanize, walter_detect, and walter_batch_humanize. Skills are instruction files that tell Claude how to use those tools for a specific workflow, like agency QC or local SEO. You can use the server without skills, but the batch prompts work much better with the relevant skill installed.

How do I install a single Walter skill?
Run npx skills add walterwritesai/walter-skills --skill walter-agency-qc, swapping in the skill you need, such as walter-local-seo, walter-programmatic-seo, or walter-content-repurposer.

Is waltermcp.com official?
No. It's a community-run guide to Walter's MCP setup and skills. It's a useful second reference, but Walter's own docs and the GitHub repo are the source of truth.

Do these prompts work in Claude Code?
Yes, and the file-based versions only work there, since Claude Code can read files with @path references and write results back to your repo. Add the server with claude mcp add walter-writes https://mcp-server.walterwrites.ai/mcp -t http --scope user and authenticate with /mcp.

How do I make sure locked keywords survive humanization?
List the exact terms rather than describing them, and ask Claude to check each one against the preservation report walter_humanize returns. That gives you a visible confirmation instead of an assumption.

Which humanization mode should I use?
The Walter skills use balanced mode for most content, light mode for casual social formats, and aggressive mode only when a passage still scores high after a balanced pass. Start with balanced and only go stronger if detection says you need to.

Does this work in languages other than English?
Yes. Walter's humanizer supports more than 80 languages, and the prompts work the same way since they describe process, not language.