Insta parcel skill

Insta Parcel Skill

Build an audit-ready creator candidate dataset. Treat every discovered profile as a candidate until evidence, metrics, validation, and human review pass.

Set The Research Contract

Before collecting anything, define:

• campaign purpose and audience segment;
• allowed public, official, exported, or permissioned sources;
• minimum follower threshold;
• accepted account types;
• required metrics and ER sample size;
• output CSV path outside version control;
• human reviewer and publication destination.

If these are missing, create a conservative draft contract and mark assumptions.

Enforce Boundaries

• Use public pages, approved APIs, official exports, or a browser session where the user logged in manually.
• Stop at login prompts, CAPTCHA, checkpoints, or rate limits.
• Do not bypass controls, rotate proxies, fingerprint browsers, call private endpoints, or extract cookies and tokens.
• Do not collect private profiles, minors, or sensitive personal attributes.
• Do not automate likes, follows, comments, messages, or account changes.
• Use neutral campaign segments. Treat location and audience fit as hypotheses that require direct public evidence.
• Keep raw exports, browser state, and production data outside public Git.

Read references/evidence-policy.md before designing a source strategy or approving profiles.

Discover Candidates

Prefer narrow, auditable sources:

1. official venue, event, association, or campaign pages;
2. explicit public creator bios or posts;
3. permissioned platform exports;
4. approved analytics providers or official APIs;
5. search results used only as candidate discovery, never as proof.

For each candidate, preserve the originating URL, discovery context, and parse date. Normalize handles to lowercase and exclude previously reviewed handles.

Search queries, hashtags, recommendations, visual similarity, and a source account's general theme do not prove that an individual belongs to a segment.

Review Evidence

Approve segment evidence only when a public source directly supports it, such as:

• explicit text in the creator's public bio;
• explicit text in a specific public post;
• a visible tag or collaboration with an independently verified source;
• an official or permissioned export that names the relationship.

Record a compact evidence excerpt or paraphrase and its URL. Mark ambiguous, stale, unavailable, or indirect evidence as needs_manual_review.

Calculate Metrics

Use visible or authorized metrics only. Do not invent unavailable values.

Calculate post-based ER as:

ER % = (average likes + average comments) / followers * 100

Prefer at least three recent, representative posts. Flag one-post samples, hidden metrics, viral outliers, ER above the configured ceiling, and mismatched snapshot dates.

Build The CSV

Use the schema in references/data-contract.md. Keep candidate, reviewed, approved, rejected, and published states distinct.

Run the bundled deterministic validator from the installed skill directory:

python3 "<skill-dir>/scripts/validate_batch.py" \
  "/path/to/batch.csv" \
  --min-followers 3000 \
  --allowed-account-type personal \
  --require-er \
  --require-evidence

Use --json for a machine-readable report. Do not publish when validation returns a non-zero exit code.

Apply Human Review

For every accepted row, confirm:

• the profile is public and currently available;
• the handle and profile URL agree;
• the account type and follower threshold pass;
• source URL, evidence, and parse date are present;
• metrics are visible or authorized and ER is reproducible;
• the segment claim is direct and auditable;
• the profile is relevant to the stated campaign purpose;
• the row does not describe a private person, minor, business-only page, aggregator, or unrelated source unless the contract explicitly allows it.

Never infer sensitive traits from appearance, neighborhood, associations, or other proxies.

Preview And Publish

Before merging into an existing database:

1. back up the current dataset;
2. compare by normalized handle;
3. require no unexpected removals or changes;
4. preserve existing workflow statuses and metric history;
5. quarantine rejected rows with reasons;
6. publish only the explicit reviewer allowlist;
7. verify totals, segment counts, evidence coverage, and access controls.

Read references/publish-checklist.md before any production update.

Handle Failures

• Login, CAPTCHA, checkpoint, or rate limit: stop and request manual action.
• Missing metrics: keep ER empty and withhold approval when ER is required.
• Weak evidence: retain as a candidate or reject; never promote automatically.
• Duplicate handle: merge only through an explicit preview.
• Validation error: fix or quarantine the row.
• Interrupted publication: inspect the live dataset and run identifier before retrying.

Report

Report:

• candidates discovered and reviewed;
• accepted, rejected, and manual-review counts;
• rejection reasons;
• follower and ER coverage;
• duplicates and validation errors;
• publication diff and backup location when applicable;
• unresolved evidence gaps and the next approved source set.

Do not claim completion until the target dataset or production destination proves it.