Research Social Public Data

Research Social Public Data

Build a repeatable workflow for spreadsheet-driven overseas social-platform research. Treat the job as a public-evidence pipeline, not as a writing task. This skill is topic-agnostic: the current Excel and user brief define the scope every time.

Quick Start

• If the input is .xlsx or .csv, run scripts/normalize_accounts.py first to build the normalized account list.
• Assume the user is a PM with only Codex available unless the workspace clearly shows otherwise.
• Default to a zero-install path first. Do not assume local RSSHub, Docker, browser automation, Node, or Python services are already available for the user outside Codex.
• Derive the research theme from the current spreadsheet and brief. If the user wants a generic account audit rather than a specific theme, switch to account-activity mode instead of forcing a topic filter.
• Extract and normalize the account list before writing conclusions.
• Freeze the task contract: input file, date cutoff, topic scope, platform scope, audience, deliverables.
• Build one canonical detail_rows table and treat it as the source of truth.
• Produce an internal pack first. Produce the client version only after the final row set is frozen.
• Write N/A for inaccessible public metrics. Never estimate totals that the public surface does not expose.

Workflow

1. Read references/intake-contract.md at the start of every task.
2. Read references/platform-playbook.md before collection.
3. If zero-install collection is insufficient, read references/setup-escalation.md before asking to install anything.
4. Read references/filtering-and-metrics.md after raw collection.
5. Read references/report-structure.md before building .docx or customer-facing output.
6. If the task includes .xlsx input or .docx output and the spreadsheets or documents skills are available, use them. Keep this skill as the research workflow authority.

Operating Rules

• Preserve the original sheet name, row number, region, market, platform, account name, and URL for every source record.
• Treat prior project examples in this skill as method examples only. Never inherit a default theme, keyword pack, or report angle from them.
• Use the zero-install route first: direct public pages, public RSS, public feeds, and lightweight Codex-side processing.
• Only escalate to local dependency setup when it materially improves coverage and cannot be replaced by the zero-install route.
• Ask for local setup only after a failed or incomplete zero-install pass. Do not ask preemptively.
• Before asking, summarize three things: what the zero-install pass already got, what is still missing, and what the proposed setup is likely to improve.
• Before any dependency setup, ask one explicit question in this exact pattern: 这个任务如果要提高 [平台/能力] 的覆盖率，需要在你电脑上下载并配置 [工具名]。是否现在让我下载并配置？如果不同意，我就直接跳过这一步，继续用当前可行方案。
• When describing the expected benefit of setup, use evidence from the current run or a validated prior example from this skill. Do not invent coverage percentages.
• If the user declines, skip the setup immediately and continue. Do not keep persuading or re-asking unless the user changes direction.
• Prefer stable public surfaces over ambitious scraping. If a route gives discovery but not metrics, keep the row and mark metrics N/A.
• Do not present public_video_views_min or public_interaction_min as full-platform totals. They are lower bounds from visible public data only.
• Do not put RSS failures, login walls, WAF problems, or scraping experiments into the client report.
• Do not start manual Word editing when the data model changes. Regenerate from scripts or structured tables.
• If the user asks for unsupported metrics such as total impressions, topic mention volume, negative share of voice, or sentiment split, only produce them when an actual source exists. Otherwise state that the public workflow cannot support the claim.

Deliverables

Default deliverables for this skill:

• normalized account table
• canonical detail row table
• internal report
• client report
• short summary of what was and was not publicly verifiable

References

• references/intake-contract.md: spreadsheet normalization, required task fields, canonical row schema, output contract
• references/platform-playbook.md: platform-by-platform collection order, validated RSS/public-feed fallbacks, stop rules, expected fields
• references/setup-escalation.md: when to ask for local setup, how to describe expected gains, and examples validated in prior production use
• references/filtering-and-metrics.md: topic filters, dedup content grouping, metric formulas, region participation, claim discipline
• references/report-structure.md: internal vs client report layout, column set, Word QA, phrasing rules