SEOBuild Onpage (SEO-AGI)
One keyword command in, a complete page out — built to rank on Google and get cited by LLMs like ChatGPT and Perplexity.
Files show keys stored locally in ~/.config, no malware, covert exfiltration or destructive defaults; the changelog explicitly rejects cloaking/JS-redirect schemes and includes junk-path filtering; data flow is largely transparent. Deductions: no explicit user-confirmation step, no rollback documentation, no formal disclosure document for external services (DataForSEO, Massive, GSC); least privilege and attribution incomplete.
CHANGELOG and CLAUDE.md reference multi-version regression tests (test_dataforseo.py shows genuine parsing tests) and a --mock mode requiring no keys; key paths are self-consistent. Deductions: static review only, nothing executed; error handling and failure-feedback quality only partially evidenced; capped below 10 by static calibration.
Triggers, target audience (English SEO/GEO content producers), inputs/outputs and process boundaries are clearly stated, including non-fit ranges (CSR SPAs, multi-service local pages). Deductions: depends entirely on overseas paid APIs (DataForSEO, Massive); no Chinese-language support declared; reachability and cost create real friction for mainland-China users.
Good doc layering (SKILL.md/README/SPEC/CLAUDE.md/CHANGELOG), versioning, detailed known-limitation disclosure (e.g., Crawl Stats not in the API), visible maintenance path. Deductions: license metadata unknown; publisher unverified, so license/governance chain is incomplete.
The workflow (research→brief→page→validation) with a 66-point checklist and printed scorecard is structurally complete and the core task path is plausible. Deductions: 'every page ranking page 1' claims are author self-reports; no third-party execution evidence; direct usability of output unverifiable statically, capped at 7.
CHANGELOG documents spec-deviation corrections and regression-test motivations; fact/inference separation is fairly rigorous (not_assessed states, {{VERIFY}} tags). Deductions: key figures (e.g., the ~25% de-indexation claim) are unauditable practitioner claims with no corroborating sources, capped at 5.
- Effectiveness claims (e.g., all pages ranking page 1) are author self-reports, independently unverified; this is a static review with no code or tests executed.
- Core function depends on overseas paid APIs (DataForSEO, Massive) with BYOK; mainland-China reachability and cost are real frictions, and no Chinese-language support is declared.
- License metadata is unknown; confirm licensing terms before use. Publisher identity is unverified.
- Credentials are stored locally in ~/.config/seo-agi/.env; users must secure that file and the GSC service-account JSON themselves.
What does this skill do, and when should you use it?
SEOBuild Onpage is an agent skill (SKILL.md plus Python scripts) that automates the author's 20+ years of manual SEO workflow from the ground-transportation industry: pull live SERPs, parse competitor pages, extract People Also Ask and related keywords, then write and score a full page. Its core framework is a Two-Gate AEO model (enter the retrieval pool, then get selected for citation), a 500-token chunk architecture, and an Anti-NLP entity-stuffing protocol. Data access is BYOK: DataForSEO is the required core, with GSC and Ahrefs/SEMRush MCP optional, degrading to web search without keys. Best suited to content and marketing teams who accept its assertive SEO theories and will configure API keys.
Runs scripts/research.py to pull DataForSEO's top-10 SERP, keyword volumes, PAA questions, and competitor content structure; optionally reads GSC data for cannibalization detection and ghost-path discovery; generates Markdown/HTML pages under rules including 500-token chunks, entity-fact pairing, and strict structural entity placement, complete with JSON-LD schema, an AI Summary Nugget, an original research block, and a 'Not For You' block; for rewrites it recommends 301 or 410 for every legacy URL; finally it scores the output against a 58-point quality checklist and prints a scorecard, flagging anything below 49/58.
- A local service business generating dozens of single-service-per-city location pages, each with LocalBusiness schema and a GBP inner-page link directive.
- An affiliate site owner needing compliant monetized pages: disclosed rel="sponsored nofollow" CTAs plus an FTC disclosure, no redirect cloaking.
- A content team refreshing underperforming legacy URLs: combining their own GSC data against the current top 3, deciding item by item whether to expand, rewrite, or 410-prune.
- Competitive intelligence research: brief only, no page — word-count ranges, heading structures, topic gaps, keyword volumes.
- Optimizers targeting AI Overviews who need outbound citations, anti-paragraph-snippet rules, and DOM-flattened extraction-friendly structure.
What are this skill's strengths and limitations?
- Compresses the entire SERP-research-to-published-page pipeline into one command, with a mock mode for zero-key trial runs.
- The framework is concrete to the HTML level: 500-token chunks, entity-fact pairing, {{VERIFY}} tags — nothing left vague.
- A 58-point checklist is scored item by item and the scorecard is mandatory output, making delivery auditable.
- It draws explicit compliance lines (no cloaking, no NLP stuffing), reducing de-indexation risk.
- The author has verifiable industry background (operating ParkingAccess.com and Shuttlefare.com).
- Core DataForSEO is a paid API (~$0.002/query); without keys the skill degrades to web search with a clear precision loss.
- The repo metadata lists License as unknown while the README claims MIT — an inconsistency to verify before adopting.
- Many claims (e.g. 25% de-indexation, 'Gemini 3.5 Flash RAG shard extraction') are practitioner lore without public, independently verifiable experiments.
- The skill is clearly tuned for local-service/transport-style keywords; other verticals are not equally validated.
- Output strictly requires SSR/SSG — CSR SPA projects need prerendering work first.
How do you install this skill?
Three options: 1) In Claude Code run claude install-skill gbessoni/seobuild-onpage, or download the release zip and upload via Settings > Skills; 2) OpenClaw/Codex users: git clone https://github.com/gbessoni/seobuild-onpage.git ~/.claude/skills/seo-agi (Codex: ~/.codex/skills/seo-agi); 3) manually clone into your agent's skill directory. Then pip install requests. Optional: mkdir -p ~/.config/seo-agi, copy .env.example, and fill in DATAFORSEO_LOGIN/PASSWORD and GSC_SERVICE_ACCOUNT_PATH.
How do you use this skill?
Verify first without keys: python3 ~/.claude/skills/seo-agi/scripts/research.py "airport parking JFK" --mock --output=compact. In production, type a trigger phrase in Claude Code, OpenClaw, or Codex — e.g. "Write an SEO page for \"airport parking JFK\"", "rank for [keyword]", "GEO", or "seo-agi". The skill runs the research script, shows a brief for confirmation, writes the full page, and prints the 58-point scorecard. Use research.py "keyword" --output=brief for a brief-only handoff to a human writer.
How does this skill compare with similar options?
It takes an adversarial stance toward NLP-coverage tools like Surfer SEO and Clearscope: the skill explicitly forbids copying their entity lists into body copy, citing de-indexation-filter risk. It also positions itself beyond audit tools — the README's line is that most SEO tools tell you what's wrong with your site, while this one writes the pages.