blogger-distiller turns a creator's posts into a writing guide for your AI

The tool states its compliance boundary before its features

The README opens with a collapsed section titled a warning to read before use, a disclaimer and a note on data handling, and that ordering is deliberate.

The tool is for study and research only, and the points are specific. Data comes from public platform data through the public REST API of TikHub, for Xiaohongshu and Douyin, with no simulated login, no injected cookies and no cracking of encrypted interfaces. The stated scope is public content only, that is public accounts, public posts and public comments, and the instruction is not to use it to crawl private content, to bypass platform risk controls, or to do anything that violates the user agreements of Xiaohongshu or Douyin or relevant law.

There is a specific promise about commenter privacy. The tool automatically anonymises commenter identities as reader 1, reader 2, author, and does not save commenter nicknames, userId, avatars or IP location. The comment bodies are kept, for content research.

For commercial use, you are told to confirm yourself whether the platform policies and the TikHub service terms allow it, and the author makes no endorsement of commercial compliance. All consequences, including account risk, data accuracy and third-party API fees, are borne by the user.

The compliance boundary is then expressed as a list. The acceptable cases are personal study, needs analysis and market research, crawling only content publicly accessible without login, and using the data only for your own research without redistributing or monetising it. The risky cases are selling or distributing other people's data, crawling content that requires a login, and using large-scale crawled data as the core data source of a commercial product.

The project is MIT licensed, Python, with the last push on 2026-09-23 and no GitHub releases. The repository holds a DISCLAIMER.md, a SECURITY.md, a SKILL.md, an install.py, a run.py, scripts/, tests/, references/ and assets/.

A script does the counting and the model does the interpreting

The architecture section is the clearest statement of the design, and it is stated as a slogan. Scripts hold the floor, the model pushes the ceiling.

The split is given as a table. The script takes 30 percent and is responsible for collecting data, computing statistics, identifying title patterns, extracting the call to action and calculating the collection-to-like ratio. The AI takes 70 percent and is responsible for distilling beliefs, finding cognitive tension, writing the persona breakdown, doing causal analysis and giving personalised suggestions.

That division is the right one for this kind of tool. Counting engagement ratios over fifty posts is something a script does exactly and a model does approximately, and the statistical claims in the final report, the ones a reader might act on, come from the script. The interpretive claims, which are the point of the exercise, come from the model.

The pipeline the tool runs is a sequence of phases, and the project file names them. Environment check, then automatic endpoint detection, then collecting the posts, then statistical analysis, then cognitive layer extraction, then distillation, then producing the report.

Five core properties are claimed for that pipeline. Zero configuration, in that dependencies are downloaded automatically and the service starts out of the box. Automatic endpoint detection, where the availability of all API endpoints is checked at startup and they are dynamically ordered. Multi-endpoint fallback, with four groups of endpoint pools, web_v2, app, web_v3 and app_v2, so that if one goes down there are three more, and the pool definition files record 28 endpoints for Xiaohongshu and 19 routes across 10 function pools for Douyin. Cross-domain generality, since no content type is assumed and clustering is based on the actual content. And checkpoint recovery, with an automatic save every 10 posts so a dropped connection does not lose the data.

Three outputs, and only one of them is for a human to read

The deliverables are what make the tool more than a report generator, and there are two by default, three if transcription is switched on.

The first is a creation guide Skill folder, for your AI. It is an installable Skill, placed at a path like `.claude/skills/XX博主_创作指南.skill/` containing a SKILL.md. The claimed use is that you then say something like write a Xiaohongshu note about AI tool recommendations in this creator's style, and the AI follows the distilled formula, tone and structure rather than a vague cheerful style.

The Skill has a three-layer distillation structure. The cognitive layer answers how they think, with an example of a belief that counter-consensus drives spread, that 80 percent of their hits opened on a counter-intuitive point. The strategy layer answers how they operate, with an example of posting Wednesday at 8pm and a save-to-like ratio of 0.85 indicating the audience treats it as a tutorial to keep. The content layer answers how they write, with an example of 35 percent numeric titles and 25 percent question titles, always opening on a personal story.

The Skill document itself has eight sections: usage instructions, the cognitive layer, the strategy layer, the content layer, creation taboos, comparison examples, topic inspiration, and a self-check of limitations.

The second output is an HTML distillation report, for you. It is a single file you open in a browser, with ten modules: a one-glance card of followers, likes, hit rate and the underlying formula; a persona breakdown; the cognitive layer; the strategy layer; a breakdown of the top ten hits; a content formula with 11 title formulas, opening templates and CTA strategies; 15 topic directions sorted by difficulty and potential; a data panel with average likes, average saves, save-to-like ratio, video versus image-text, and posting frequency; a trend section comparing early and recent behaviour with a transition path and a confidence annotation; and a final three-sentence conclusion with action advice.

The third is a spoken-word transcript, on request. After distillation you say extract the spoken word and Whisper transcribes the video audio into text, saved as a markdown file you can hand to an AI. The default Base tier is about 70 percent accurate, the model can be upgraded in the Skill at the cost of doubling transcription time per tier, and only videos of ten minutes or less are supported, with longer ones skipped automatically.

Cost per creator is a few yuan, and there are two analysis modes

There are two modes and they change what the tool is for.

Mode A is analysing a comparable creator, for learning their approach, and it produces an HTML report plus a `{creator}_creation guide.skill/` folder. Mode B is distilling your own account, for building your own creation workflow, and it produces an HTML report plus a `{username}_creation DNA.skill/` folder. The number of posts collected has three settings: 30 for a quick run, 50 which is recommended, and 80 for a deep run.

Two advanced analyses become available after the first distillation, by saying a trigger phrase, with no re-collection because they run on the existing data. Analysing covers does an 8-dimension visual breakdown, cover style type, composition, title hook, text design, whether a person appears, decorative elements, information density and visual consistency, and outputs a cover formula, a content match assessment and three targeted suggestions, with zero extra API calls. Keyword trends analyses the heat trend and audience profile of the creator's core keywords to surface content direction opportunities.

The cost table is per creator and reflects pay-as-you-go billing at TikHub. 30 posts is about 1 to 3 yuan, 50 posts about 2 to 5 yuan, and 80 posts about 4 to 8 yuan. The real figure depends on comment volume, endpoint availability and retry count, and the documentation says to treat the TikHub pricing page as the source of truth. It also notes the system has several built-in saving strategies, startup detection to avoid requests going to invalid endpoints, dead-link caching within a session, and skipping search supplementation by default.

So the cost of the tool is not the software, it is the API. That is worth saying plainly, because it makes the trade-off concrete: a few yuan per creator against the value of the Skill file you end up with.

Getting set up is three commands and a TikHub token

The quick start is short:

git clone https://github.com/otter1101/blogger-distiller.git
cd blogger-distiller
python install.py

Then in your AI programming assistant you say something like analyse the Xiaohongshu or Douyin creator XX, and the system asks you for the mode, A or B, the platform, Xiaohongshu or Douyin, and the number of posts, after which it is fully automatic.

There is also a manual path for the Skills themselves. Download the skills archive, extract it into the right folder and follow the prompts.

The TikHub setup is four steps, and the third one is the one that catches people out. Register at user.tikhub.io and verify by email. Then top up in the console under the recharge or plan page, and pay-as-you-go is recommended, billed per actual call. Then in the console under API permissions, tick all the Xiaohongshu and douyin endpoints in one go, because the system probes which endpoints are available at startup and the more you enable, the stronger the automatic fault tolerance. Finally generate a token in the API Token page and give it to your agent.

The documentation even addresses the case where the agent cannot parse the API key, with three suggestions: retry and check for network problems, change the key without topping up again, or use a different agent, and it names several that work.

There is one optional tuning knob. Different TikHub plans have different requests-per-second limits, the default is RPS=10, and you can raise it in the environment:

export TIKHUB_RPS=20

The author states there is no financial interest between them and TikHub and that the recommendation is simply one workable option.

The script layout shows where the phases live

The project structure in the documentation is a phase map, and reading it tells you what the tool actually does rather than what it claims to.

At the top level there is the SKILL.md, which the AI programming assistant reads, a run.py one-click entry point, an install.py automatic install script, and a scripts/ directory. Inside scripts/, crawl_blogger.py is the Phase 1 collection entry point that routes automatically to each platform, with crawl_xhs.py for Xiaohongshu and crawl_douyin.py for Douyin. crawl_common.py holds the shared collection utilities for directories, JSON, progress and rate limiting. analyze.py is Phase 2, the data analysis with clustering, tags, the top ten and the cognitive layer. deep_analyze.py is Phase 3, the data groundwork plus AI distillation task generation. verify.py is Phase 4, the data validation for both platforms.

Under scripts/utils/ the pieces are the TikHub API client with multi-platform routing, rate limiting and degradation, an endpoint pool router with an automatic fallback engine, the two endpoint pool JSON files, response data normalisation adapters for both platforms, the cover analyser and the keyword trend client. There is also a platform registry with shared helper functions, a privacy module doing the data anonymisation for both platforms, a first-run compliance confirmation, a data quality checking tool, and the Whisper integration for spoken-word extraction.

The first-run compliance confirmation module is worth noticing. It is a script rather than a paragraph in a document, which means the compliance boundary is checked at the point of use rather than left for the reader to have noticed.

The version history table runs from v0.1 to v1.0, moving from SKILL design through the phase scripts, a generalisation refactor, automatic environment preparation and a stabilised input method that used a QR code login and has since been deprecated. v1.7 added the cognitive layer extraction of opinion sentences, thinking patterns and value words, and v1.8 changed the deliverables to the HTML report plus the Skill.

Editorial conclusion

Adopt blogger-distiller if you are studying how a specific creator's content works and want the analysis in a form your AI assistant can actually use, because the Skill file is a machine-readable set of rules rather than a summary you would have to retype. Skip it if you need a general content analysis tool, since it is built for one workflow, and think carefully about the platform terms, because the project itself sets out which uses it considers safe and which it does not. Verify first that you have a TikHub account with the endpoints enabled, that the creator's posts are publicly visible, and that you are willing to pay per call, since the documented cost is a few yuan per creator depending on how many posts you collect.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. otter1101/blogger-distiller on GitHub
4. README