backlink_skills: 743 candidate sites, two submission postures, and a status called outcome unknown

The candidate list is 743 entries, and the project says many of them are not usable

The headline asset is a single markdown file holding 743 websites or submission entry points, each with its own Chinese description classifying it as a product directory, an AI tool listing, a founder community, a software review site, a content platform, a company directory or a plain submission form. The accompanying text is careful about what that number means. The list has had historical submission status stripped out, so results from one product cannot be carried over into another task. Old traffic figures with no source and no month were removed. And the file name is inherited from the team's internal naming, which the project warns should not be read as a promise that every entry can currently be submitted to for free: the list explicitly contains sites that have shut down, entry points that have closed, directories that require payment, sites that demand reciprocal links, channels suitable only for articles, duplicates, and entries nobody has verified. The recommended use is to filter, then check a handful properly, then decide.

Two submission skills, and batch is defined as queue management rather than concurrency

There are two submission skills with genuinely different postures, and the document tells you which to use. The batch skill is for a set of URLs you have already vetted, and the prompt it expects looks like this:

使用 $submit-product-directories-v1-batch 处理我提供的目录 URL。
先规范化、去重并建立执行分片；先检查登录和验证码，把人工步骤集中成队列。
只使用已确认的产品资料，不付费、不添加互链、不绕过验证。
每个站点完成后立即记录证据，结果不明确时不要重试。

In outline it optimises for coverage, throughput and resumability: it normalises URLs, strips tracking parameters, deduplicates ahead of time, assigns stable queue identifiers, idempotency keys and execution shards, does read-only checks first to surface login and verification steps before touching a form, batches those manual steps into one queue so the run does not stall on the first challenge, submits site by site in the same browser profile, writes a result after each site before advancing the cursor, and records seven distinct outcomes so an interrupted run can continue. The quality skill is the opposite: it researches a site's real audience and editorial governance first, filters out low-quality networks and paid-link sellers, runs at most ten sites per batch, and checks authorisation separately for each action.

An unknown outcome is a state you may not retry, which is the mechanism that prevents duplicates

The status vocabulary is the most reusable thing in the repository. A submission counts as submitted only when there is a reliable receipt or sent-mail evidence. A separate state means waiting on email verification, another means the site has said it is under review, and published means a public, non-preview page shows the correct product identity. Then come the three states that matter for restraint: outcome unknown, meaning the final action was performed but the receipt is uncertain and must not be retried until it is investigated; submission failed, requiring an explicit rejection, a bounced mail or reliable failure evidence; and ineligible or unavailable, for a site that does not suit the product or cannot be used at all. The document's framing is that submitting is not the same as succeeding, and the batch skill's cursor behaviour, writing a result before moving on, exists so an ambiguous run does not replay work.

The banned techniques are named, and they are exactly the ones an anti-detection browser would sell

One paragraph does the reputational work for the project. Captchas, challenge-response tokens, two-factor codes, passkeys and email or phone verification must be completed by the site's own flow or by the user. The skills state that they do not use captcha answering, stealth browsing, proxy rotation or fingerprint evasion. Read that list as a boundary marker, because each item is a purchasable capability that other tools in this category provide. The batch example prompt repeats the same constraints in the instructions it hands the agent: use only confirmed product material, do not pay, do not add reciprocal links, do not bypass verification, and when a result is unclear do not retry. The quality skill adds its own exclusions, filtering sites that only sell ranking links, that force reciprocity, that are unrelated to the product, or that offer nothing a reader could genuinely find useful.

Credential hygiene is an instruction to the user, not a design in the tool

The preparation list asks for a specific set of inputs and the omissions are as informative as the inclusions. You supply a product name, a canonical site URL, positioning and target users; short, medium and long real product descriptions; public and verified facts such as category, price, launch date, company and founder; assets cleared for upload; and then, notably, aliases for accounts and contact information. The instruction attached to that last item is that passwords, one-time codes, cookies and recovery codes must not be written into the task record. So the model is that the human holds the credentials and completes the steps that need them, while the skill holds an alias it can type. It is a sensible division and it is enforced by convention, since nothing in the visible text prevents a user from pasting a real password into a chat with the agent. There is no secret store, no credential vault and no redaction pass mentioned anywhere.

Four writing skills, and none of them will log in and publish on your behalf

The writing half is larger than the submission half: a general search-oriented writer plus three platform-specific entries for LinkedIn, Medium and WeChat. The document is emphatic that the platform variants are not the same skill with a different output folder, since each is redesigned around a different audience and platform ecosystem, adding topic research, narrative structure, publication matching, image captioning and, for two of them, disclosure of AI assistance. What they produce is specified tightly: a local draft or a publishing package. The statement that matters is that writing an article does not cause the skill to log in and publish, and that external publishing, account actions and image uploads remain independently authorised steps. The general writer adds a fact-checking pass over capabilities, prices, dates, statistics, policy and comparative claims, an SEO audit, a humanisation pass applied after the factual and structural fixes, and 16:9 image planning with local relative paths by default, with an optional upload to object storage only when configured.

The documented tree lists two submission skills, and the repository contains a third with a misspelled name

The repository structure block is worth comparing against the actual listing. The tree describes the Chinese main document, a language-prefixed set, the candidate list, an assets directory, a prospecting matrix folder with a YAML source table and its own skills, and the two submission skills with their agent configuration, assets, references, scripts and tests. The real root has all of that, plus a third submission-related directory whose name carries a typo, reading the product word with an extra letter, and plus a separate skill matrix directory that the tree never mentions. There is also a redundancy in the readme files: the language list presents the unsuffixed file as the main Simplified Chinese document, and a Chinese-suffixed file also exists at the root. Each skill does have its own SKILL.md, and the documentation warns against treating the whole repository as one skill, which is right, since two submission modes and four writing entries behave very differently.

Editorial conclusion

This is worth reading if you do product marketing and want an auditable process rather than a spamming script, because the parts that make it defensible are the parts it is explicit about: verification steps are handed to a human, paid links and forced reciprocity are filtered out rather than tolerated, an ambiguous outcome is never retried automatically, and every submission action asks for authorisation separately. The documentation also disclaims indexing, follow links, traffic and rankings outright, which tells you what the authors think the tool is worth. Two caveats before you use the list. The 743 entries are candidates, and the project says so plainly, listing shutdown sites, closed entry points, paid-only directories, article-only channels and duplicates among them, so the list needs re-checking against live pages rather than being worked through in order. And the repository is maintained by an agency that sells this kind of promotion, so the workflow is both their method and their marketing; there are no releases, so what you get is whatever the branch holds, and the writing skills produce local drafts rather than publishing anything themselves.

Frequently asked questions

Official sources

1. flaqai/backlink_skills on GitHub
2. Issues
3. License: MIT
4. Project website
5. README