生成式引擎优化

Language / 语言: English · 中文

dsh-geo is a DeepSeek Harness bundle that gives the agent explainable SEO, GEO and AEO tools for Markdown knowledge bases.

Technical package ID: dsh-geo · DeepSeek Harness plugin for SEO, GEO and AEO

What it does

• Audit one Markdown note for SEO, GEO and AEO readiness.
• Scan a vault for metadata gaps, missing sources, broken knowledge structure, orphan notes and duplicate titles.
• Check the configured root before scanning and create a shareable project report.
• Build a content brief with query, intent, audience, outline, questions and source gaps.
• Run the complete workflow from a public website URL, public account URL, Markdown export or HTML snapshot.
• Build a qualitative keyword plan from source signals, optional seed terms and Harness web search results; it never invents search volume.
• Turn the diagnosis into a four-stage production plan: diagnose, map keywords, draft, and verify.
• Check source provenance and freshness fields.
• Preview or apply a complete Markdown replacement, including safe creation of a new note inside defaultRoot.

The plugin is local-first. Core analysis runs through the Harness filesystem service. Public URLs use the official anonymous ctx.web fetch seam; cookies and credentials are never used. Private or JavaScript-rendered pages should be exported as Markdown/HTML and analyzed as a local snapshot.

User and business pain points

Teams publishing content for both search engines and AI answer engines commonly face these problems:

• Traditional SEO checks focus on keywords and metadata, but do not explain whether an answer engine can understand, trust and cite the content.
• Enterprise knowledge bases accumulate missing sources, stale notes, duplicate titles and orphan pages, making content quality difficult to govern at scale.
• Writers, subject-matter experts and SEO teams use different review standards, so content decisions are slow and hard to reproduce.
• Content optimization is often disconnected from the original Markdown source, creating manual copy-paste work and a risk of overwriting newer edits.
• Teams may not be able to send private product, customer or internal knowledge-base content to an external GEO/SEO service.

This plugin turns those needs into explainable, local-first checks with evidence, scores, priorities and guarded content updates.

Typical application scenarios

• Pre-publication review: Check a new article, product page or technical note for SEO, GEO and AEO readiness before release.
• Knowledge-base governance: Scan a team or enterprise Markdown vault to find missing metadata, weak provenance, stale content, orphan notes and duplicate titles.
• Content planning: Convert a topic or existing note into a brief with audience, intent, outline, questions and source gaps.
• Product and technical documentation: Improve documentation discoverability while preserving source context and review history.
• Agency and multi-brand operations: Apply a consistent, auditable review framework across clients, sites or content teams without sharing the source vault with a third-party service.

Included tools

Install from npm

dsh plugin --profile default add dsh-geo

Install from GitHub

dsh plugin --profile default add github:winyh/dsh-geo

The package builds itself after a GitHub source install. For reproducible deployments, pin the command to a reviewed commit, for example github:winyh/dsh-geo#<commit>.

GitHub discoverability

Publish this directory as a public repository and add the repository topic dsh-plugin. The package metadata already includes dsh-plugin; the GitHub topic is what makes the repository appear in the Harness plugin topic listing.

Usage

1. Configure and start Harness

Set the knowledge-base root in the Bundle configuration, then start DeepSeek Harness:

dsh web

For local development, build first and install the local directory from its parent folder:

pnpm install
pnpm run build
dsh plugin --profile default add ./dsh-geo

2. Ask with natural language

The recommended entry point for a real SEO project is geo_workflow. It keeps diagnosis, keyword adjustment and content production in one result instead of making you guess which tool to call next.

Analyze a public website or public account page:

Run geo_workflow for https://example.com. Goal: increase qualified product education traffic. Audience: first-time evaluators. Return the SEO/GEO/AEO diagnosis, qualitative keyword plan and a complete Markdown production plan. Do not write files.

Analyze a public or private account page after export:

Run geo_workflow for snapshots/account-home.html. Treat it as a private-page snapshot. Goal: make the profile easier to discover and quote. Return the diagnosis, keyword mapping, content brief and verification checklist. Do not write files.

The source boundary is deliberate:

• Public http(s) URLs are fetched anonymously through Harness ctx.web.
• Public pages that require JavaScript rendering should be saved/exported as Markdown or HTML first.
• Private account pages are supported through a local Markdown/HTML snapshot; browser cookies and platform credentials do not enter the plugin.
• Local Markdown/HTML inputs do not send extracted terms to public search; their keyword plan stays seed-only unless you explicitly provide external keyword data yourself.
• Search results provide qualitative topic signals only. Search volume, ranking difficulty and traffic require a separate data source and are not fabricated by this plugin.

Start with a readiness check when installing into a new environment:

Check whether the configured knowledge-base root is ready for a local scan.

Audit this Markdown note for SEO, GEO and AEO. Show scores, evidence and the top five actions.

Scan my knowledge base and list notes with missing sources, orphan links and the lowest overall scores.

Create a project report with average SEO, GEO and AEO scores, governance gaps and priority files.

Create a content brief from this note, including audience, intent, outline, questions and source gaps.

Run the complete workflow for https://example.com, use "product education" as the seed keyword, and generate a draft plan without writing anything.

Use focus=seo, focus=geo or focus=aeo when you only want one assessment pillar.

3. Preview before writing

Content changes are preview-only by default. A reliable first-use loop is:

1. Ask for an audit of one note and select one or two high-impact findings.
2. Ask the agent to rewrite only that note while preserving factual claims, sources and useful local links.
3. Ask for a preview and review the diff, not just the score.
4. Explicitly confirm the preview. Harness asks for approval again before the write.
5. Apply the preview. The agent can use path + previewToken; repeating the full Markdown content is optional because the token already binds the exact preview.

The write operation uses a version guard and refuses to overwrite a file changed after the preview. If that happens, audit the current file and create a new preview instead of forcing the old change through. To save a draft as a new note, use createIfMissing=true on geo_preview_content; the destination must still be inside defaultRoot.

For a practical first request, use:

Check the configured root, audit notes/launch.md, fix only the three highest-impact issues, show a diff, and wait for my approval before writing anything.

Configure

The bundle ships with these defaults:

defaultRoot: "<your-knowledge-base-root>"
maxFiles: 500
maxFileBytes: 1048576
maxTextChars: 180000
maxResultChars: 50000

Tool paths may be absolute or workspace-relative. Relative paths resolve under defaultRoot, and every read or write is checked against that boundary.

Example prompts

请审计一篇 GEO.md，分别给出 SEO、GEO、AEO 的问题和优先级。

扫描我的知识库，找出来源缺失、孤立笔记和综合评分最低的 10 个文件。

先预览对 GEO.md 的优化结果；只有我明确确认后，才应用修改。

Development

pnpm install
pnpm run typecheck
pnpm test
pnpm run build

The plugin follows the DeepSeek Harness Cordis model: it exports apply(ctx), injects tools, fs and web, and registers model-facing tools through the normal tool pipeline. Core scoring and file analysis are self-contained; public URL retrieval uses the official Harness web capability and does not require GEO-PRO.

DeepSeek Harness is currently a developer preview, so the plugin pins the current release-candidate peer range and should be tested against the Harness version used for deployment.

Images and contact

The following images are included from the project-level img folder and are intentionally displayed in this README:

微信联系

微信支付

Contact: 2712192471@qq.com

License

MIT. See LICENSE.