ScrapeCreators social-media-research-skills: agent workflows that end in an artifact, not a JSON dump

The gap between an endpoint and a research artifact

Most social scraping tools hand you a function and stop. You call it, you get a payload with view counts, captions and comment bodies, and then you spend the afternoon deciding what counts as interesting. The README for this repository names that gap directly: the project is described as "not just endpoint routing", and the stated goal is to give an agent complete research workflows. That is the whole pitch. The unit of delivery is a skill that ends in a document, not a response object.

The thirteen skills listed in the README each name an output. outlier-post-finder returns an outlier table plus repeatable patterns and hooks. comment-mining returns a VOC report with themes and quotes. ad-library-teardown returns messaging angles, hooks, CTAs and offers. The audience is whoever already runs social research by hand: content marketers, influencer marketers, competitor analysts, and the founders who do all three. It is not aimed at data engineers building a pipeline, because the output is prose and tables meant for a human to read.

Thirteen skills, one data layer, and a deliberate split

The architecture in the README is a two-tier structure. At the bottom sits scrapecreators-api, described as the skill that routes a raw scraping request to the right ScrapeCreators endpoint and carries the endpoint references and pagination guidance. Above it sit twelve workflow skills, from outlier-post-finder through content-repurposing. The README states the rule plainly: workflow skills should use scrapecreators-api as the data layer when they need endpoint details.

That split matters more than it looks. Endpoint names, parameter shapes and cursor handling change; research procedures change less often. Putting pagination in one place means a fix to how the agent walks a video list does not have to be repeated across twelve directories. The cost is an extra hop: a workflow skill that needs a specific endpoint has to consult the API skill first, and if the agent skips that step it will guess at a URL. There is no evidence in the README of a validation layer that catches a guessed endpoint before the call is made.

The names are also unusually literal. transcript-intelligence, competitor-social-research, social-listening-brief and product-demand-research describe the task rather than the data source, which is what you want when the agent is picking a skill from a request like "why is this account working".

Baseline-aware analysis is the one design choice worth copying

Of the six design principles in the README, the fourth is the one that separates this from a generic summarizer: judge performance against a creator's own normal performance, not only raw vanity metrics. A post with 40,000 views is unremarkable for an account that averages 35,000 and remarkable for one that averages 2,000. The outlier-post-finder skill is built around that idea, and its stated output includes repeatable patterns and hooks to steal, which only makes sense if the outliers have been separated from the account's ordinary output first.

The same principle shows up in audience-research, which the README says produces an audience-fit report with market and country notes and confidence labels. Confidence labels are a small thing that changes how you read the output: the agent is being asked to mark how sure it is, rather than presenting a demographic guess as a finding. The README does not explain how those labels are assigned or what the scale is, and that is a real gap. You would need to open the skill directory to see whether the labels are defined or left to the model's judgement.

Install and setup: two commands and one environment variable

Installation is a single command from the README:

npx skills add ScrapeCreators/social-media-research-skills

Setup is one environment variable:

export SCRAPECREATORS_API_KEY=sk_...

The key comes from scrapecreators.com. The README lists compatibility with Claude Code, Cursor, OpenAI Codex, GitHub Copilot, Gemini CLI, Windsurf, VS Code, and other agents that support the Agent Skills spec at agentskills.io. Because the install path goes through the skills CLI rather than a language package manager, the Python listing on the repository is about the skill implementations rather than something you pip install. There is no requirements.txt step in the README, and no version pinning guidance.

The prompts in the README show the intended level of specificity. One example asks for outlier posts for @starterstory on YouTube Shorts from the latest page of videos. Another asks to mine comments on a viral Instagram Reel for objections, questions, buying intent and exact audience language. Note that both name a platform, an account or post, and the fields wanted. Vague requests will get vague artifacts, and nothing in the material suggests the skills compensate for that.

The API key is the boundary, and it is a hard one

The second design principle is public-data only, with an explicit instruction not to promise logged-in or private data. That is a constraint you should take at face value. Anything behind a login, any private account, any ad set that is not in a public ad library, is out of scope. If your research question depends on private audience data or on a platform's internal analytics, this repository will not answer it and no amount of prompt engineering will change that.

The harder dependency is the key. Every workflow skill ultimately calls ScrapeCreators, which is a commercial service. The repository is MIT licensed, but the licence covers the skill definitions, not the data access. Running these skills without a paid or trial key produces nothing. That also means your cost scales with research volume in a way that a self-hosted scraper does not, and the README gives no rate limits, quota figures or pricing. You have to check scrapecreators.com for that before you plan around it.

A third limitation is quieter. The skills produce cited outputs with source URLs, which is good practice, but citation quality depends on the underlying payload including those URLs. Where a platform's response omits a permalink or a stable identifier, the agent has less to cite. The README does not discuss how the skills behave when a field is missing.

What this replaces, and what it does not

The obvious alternative is calling the ScrapeCreators API directly, or calling any social data API directly, and writing your own prompts around it. The difference is where the research logic lives. With a direct API integration, every analysis starts from raw fields and the analyst supplies the framing: what counts as an outlier, which comments are objections versus praise, how to compare two brands' content pillars. With these skills, that framing is written into the skill definitions once and reused. The trade-off is control. A skill that decides for you what an outlier is will be wrong for some accounts, and overriding it means editing the skill rather than adjusting a prompt.

A second alternative is a general-purpose research agent with web search and no social-platform integration. That approach can read a blog post about a trend but cannot pull the last page of a creator's Shorts and compute a baseline. The distinction is access to structured per-post metrics, which is what the ScrapeCreators layer provides and what makes baseline-aware analysis possible at all.

A third path is a full social listening platform, which typically bundles ingestion, storage, dashboards and alerting. Those are built for ongoing monitoring with a persistent dataset. This repository is built for one-off research runs that produce a brief. If your need is a dashboard that updates daily, the skills are the wrong shape.

Maintenance, licence and what to check before you commit

The repository is MIT licensed, which is permissive and places no obligation on how you use or redistribute the skill definitions. That licence does not extend to ScrapeCreators data or to the API itself; those are governed by the service's own terms, and the README does not restate them. Treat the MIT grant as covering the code in this repository only.

On maintenance, the last push recorded is 2026-08-26 and no releases were retrieved, so there is no versioned changelog to track. Updates arrive as commits to main. Because the skills depend on an external API, the failure mode you should expect is not a bug in the skill logic but a drift between the endpoint references in scrapecreators-api and the live API. The README points to docs.scrapecreators.com and to dedicated Agent Skill and MCP integration pages, which is where endpoint changes would surface first.

Before adopting, open skills/outlier-post-finder/ and read how the baseline threshold is defined, since that single choice determines whether the outlier table is useful or noise. Then open skills/scrapecreators-api/ and confirm it documents pagination for the platforms you care about, because a workflow that only reads the first page of results will miss exactly the older outliers you are looking for.

Editorial conclusion

Adopt it if you already pay for ScrapeCreators and want an agent to produce a competitor brief or a VOC report without you writing the prompt scaffolding each time. Do not adopt it if you need logged-in data, if you want a self-hosted scraper, or if you are not willing to hold a ScrapeCreators key. Before committing, read skills/outlier-post-finder/ and skills/scrapecreators-api/ to see how much of the baseline logic is written down versus left to the model, and check the Agent Skill docs for whether your client reads SKILL.md files from the installed path.

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. ScrapeCreators/social-media-research-skills on GitHub