SEO Link Strategy Agent Skill

SEO Link Strategy

Build an evidence-backed backlink opportunity list and prepare local outreach drafts. Discovery, evidence capture, draft generation, and external execution are separate states.

This Skill does not ship a current directory of platforms or verified contacts. Candidates come from the user or from research performed at runtime. Never treat a guessed address, old list, search snippet, or role-based email pattern as an observed contact.

Operating boundary

Default work is research and local draft preparation. Do not send Gmail or other messages, submit forms, purchase placements, call paid/mutating services, or change remote systems by default.

Before any external side effect:

• show the exact target, payload, account, and expected effect;
• obtain explicit user authorization for that specific action;
• execute only the authorized action; and
• record the actual result and UTC time, including failures or uncertainty.

A draft is not-sent. An opportunity is not a backlink. An attempted submission is not completed unless a real result supports that state.

Evidence contract

Each opportunity or contact record must support these fields:

| Field | Requirement |

|---|---|

| name | User-provided or runtime-discovered platform/contact label. |

| source_url | Exact page where the opportunity or contact evidence was observed. |

| observed_at | Past UTC timestamp for the actual observation, or null. Future times are rejected. |

| status | observed only with captured evidence and time; otherwise unknown. |

| email | One value present in captured evidence, or null; never constructed from hints. |

| context | User-supplied limits, conflicts, access failures, or relevance notes. |

Keep sources and observation times when data is handed from research to drafting. If a page cannot be checked, preserve the candidate with status: unknown rather than filling likely values.

1. Normalize captured contact evidence

Use scripts/contact_discoverer.py after the user or an available research tool supplies candidate URLs and captured page text. The script is offline: it extracts addresses only from the supplied capture and makes no network request.

Input JSON:

{
  "opportunities": [
    {
      "name": "Example Directory",
      "source_url": "https://directory.example/contact",
      "observed_at": "2020-01-15T09:45:00Z",
      "captured_text": "Contact: [email protected]",
      "notes": "Captured from the public contact page."
    },
    {
      "name": "Unreviewed Candidate",
      "source_url": "https://candidate.example/contact"
    }
  ]
}

Run:

python3 scripts/contact_discoverer.py --input candidates.json
python3 scripts/contact_discoverer.py --input candidates.json --output contacts.json

Without --output, JSON goes to stdout. An existing output path is never overwritten. The first record can become observed; the second remains unknown with no invented email. If supplied evidence contains multiple distinct addresses, the normalizer emits one contact record per address so every record keeps a single unambiguous recipient.

2. Generate local drafts

Use scripts/email_generator.py with user-provided product, sender, and contact data. It performs no browsing and contains no sending implementation.

The normalizer's contacts array is already the exact contact schema accepted here. Copy that array unchanged into a new object with product and sender; omit only the normalizer's top-level generated_at. No field renaming or implicit conversion is required.

{
  "product": {
    "name": "Example Product",
    "url": "https://product.example",
    "tagline": "User-provided product description",
    "selling_points": ["User-provided differentiator"]
  },
  "sender": {
    "name": "Your Name",
    "email": "[email protected]"
  },
  "contacts": [
    {
      "name": "Example Directory",
      "email": "[email protected]",
      "source_url": "https://directory.example/contact",
      "observed_at": "2020-01-15T09:45:00Z",
      "status": "observed",
      "context": "User-provided reason this audience is relevant."
    }
  ]
}

Run:

python3 scripts/email_generator.py --input outreach-input.json
python3 scripts/email_generator.py --input outreach-input.json --output drafts.json

The output marks every draft delivery_status: not-sent. Invalid inputs fail before output creation, and existing output files are not replaced.

3. Review and optional execution

Review source freshness, recipient relevance, claims, product links, sender identity, and applicable platform rules. Remove unsupported personalization instead of presenting inference as observation.

If the user later authorizes sending or submitting, use an appropriate external tool only for the approved target and payload. Store a separate result record with fields such as attempted_at, status, result_id, and error. Do not rewrite the draft-generation output to imply delivery.