🛠️ Groundcrew SEO

Groundcrew SEO is a provider-flexible pack of 20 plain-Markdown skills for any capable AI agent: site auditing, repository fixes, keyword and competitor research, AI/answer-engine visibility, authority and backlink review, content strategy, credibility review, evidence-labeled reports, standups, and a survey → fix → verify → report loop. No lock-in: every skill works with public/local evidence and validated imports; TrustGrowth adds the one thing a skill pack cannot carry — measured history and a real verify loop.

Published before v0.5.0 as groundcrew; old GitHub URLs redirect.

curl -fsSL https://raw.githubusercontent.com/techwright-lab/groundcrew-seo/v1.1.0/install.sh | bash

Then ask a coding-capable agent:

Inspect this site and repository, fix one verifiable defect, run the tests, and prepare a PR.

fix-my-site completes that without any account. The piped installer fetches the v1.1.0 payload, independently of the current working directory or default branch. Pass --ref REVISION to fetch an explicitly reviewed tag or commit. From a checked-out repository, ./install.sh --dry-run previews and ./install.sh installs; the installer refuses collisions by default (--update for a prior Groundcrew-managed install, --force after reviewing the exact paths, --skills-dir PATH to override detection).

Diagnostics, from your agent's skills directory (the doctor checks compatibility; write restrictions are skill instructions, not an interceptor of agent calls):

<skills-dir>/.groundcrew/groundcrew-doctor.py                       # install + shared-reference drift
<skills-dir>/.groundcrew/groundcrew-doctor.py --connectivity        # TrustGrowth reachability + full/read-only contract mode
<skills-dir>/.groundcrew/groundcrew-doctor.py --evidence record.json # validate an evidence record

Skills name evidence categories (~~search console, ~~page speed, ~~SEO tool, ~~link database, ~~analytics, ~~AI monitor), never vendors. shared/connectors.md maps each category once:

Groundcrew never opens with a connector menu: it detects what is present, runs, delivers, then recommends at most one missing connector with a concrete benefit (shared/provider-selection.md).

Every report follows one contract — shared/reporting.md: a fixed skeleton, a SHIP / FIX / BLOCK / UNDECIDED verdict where any veto forces BLOCK, and a Measured / User-provided / Estimated label on every figure. Its remediation rules keep observations, current policy, proposals, owner decisions, and rendered verification separate. Missing values stay missing — never zero, never interpolated.

flowchart LR
    A[survey<br/>audit · visibility · keywords] --> G1{owner gate}
    G1 --> B[fix<br/>one branch per root cause]
    B --> C[verify<br/>re-crawl & diff · or trigger audit → poll job → read fixed issues]
    C --> D[report<br/>skeleton · verdict · fact labels]

grow runs the whole loop or one --phase. It stops at a gate before anything irreversible or outward-facing, and a fix nothing re-observed stays "applied, unverified" — which caps the report's verdict below SHIP. The loop never self-restarts.

Three contracts keep the pack honest, and CI enforces all of them:

• Evidence — every normalized factual input satisfies shared/evidence.schema.yaml (source, observation time, provenance, confidence, limitations). groundcrew-doctor validates records and fails on drifted shared references.
• API surface — the TrustGrowth surface skills may call is generated, not hand-written: skills/trustgrowth/references/contract.md is rendered from the vendored capability manifest by scripts/gen-contract.py and refreshed from the live manifest before release; scripts/validate-skills.py rejects any skill referencing a path outside it. groundcrew-doctor --connectivity selects full mode for the target or a newer same-major contract, read-only feature-detected mode for an older same-major contract, and hard failure for a malformed or different major. Compatibility mode permits only operations advertised by the live manifest and blocks every write.
• Evals — every skill carries at least 5 scenario cases under tests/evals/ (open-tier behavior, connected behavior, honesty under missing data, a safety gate, routing); scripts/validate-evals.py keeps the 111-case corpus structurally sound and contract-valid. The cases are harness-agnostic and never ship in the distribution payload. Deterministic worked-example checks and the five-consumer instruction guard are separate and do not claim to execute agents.

• Own-data (Tier 2): Google Search Console and GA4 through a read-only MCP you already run; PageSpeed Insights with a free key. Recipes in shared/connectors.md.
• TrustGrowth (Tier 3): create an API key in the authenticated app, set TRUSTGROWTH_API_KEY. REST docs: https://trustgrowth.ai/developers · MCP: POST https://trustgrowth.ai/mcp. The 13+ free tools are keyless over REST.
• Paid indexes (Tier 3): Ahrefs MCP / Semrush MCP when you already pay for them; DataForSEO only behind a bounded cost preflight and explicit batch approval (shared/dataforseo.md).
• Imports: JSON/CSV-derived evidence is accepted when source, observation time, scope, provenance, confidence, and limitations are preserved.

Content inventory reads work; content generation, approval, scheduling, publication, and lifecycle writes do not — the managed Growth content engine remains waitlist-only (https://trustgrowth.ai/pricing), and Groundcrew does not expose or imply unavailable operations. No disavows, no automated outreach, no invented scores, no promised outcomes.

WHY-NOT-SLOP.md explains why evidence compounds trust while slop borrows against it. ETHICS.md enforces hard refusals, owner gates, null honesty, and claim safety.

All 20 release skills live only under skills/; marketplace and plugin metadata is generated from that canonical source:

./scripts/generate-adapters.py --check
GROUNDCREW_PRINT_PAYLOAD=1 ./scripts/validate-distribution.py
./scripts/preview-publishers.sh

CI fails when generated adapters drift. Every printed publisher command remains preview-only until a human approves publication. Skills are plain Markdown and the doctor uses only Python's standard library; document tested installation targets with exact commands as they are verified.

MIT