The useful product is the process, not the backlink list
Backlink Skills packages several Codex instruction sets rather than a conventional application. Two workflows handle product-directory submissions. A general writer produces website articles, while separate writers target LinkedIn, Medium, and WeChat. The submission skills define states, authorization, evidence, recovery, and handoff rules so an agent can resume work without treating a filled form or redirect as proof that a listing exists.
The headline list contains 743 sites or submission routes. The maintainers warn that it includes closed services, paid entries, reciprocal-link demands, duplicates, article routes, and items that have not been verified. Its status notes are historical and the individual site summaries are Chinese. That makes the file a candidate pool, not a ready queue. Any campaign that imports all 743 entries without fresh research would ignore the repository's own strongest advice.
V2 limits a campaign to qualified destinations
SPD V2 Quality is the better starting point for most brands. It caps a batch at 10 sites, requires an approval reference for each one, and separates research from form execution. The first pass checks audience fit, governance, link-selling claims, reciprocal requirements, and terms. Only candidates that pass every mandatory quality dimension proceed to the second pass, where fields and final actions need matching authorization.
The distinction is practical. A directory can technically accept a URL while offering no discovery value, or it can exist mainly to sell ranking claims. V2 marks those routes ineligible instead of rewarding the agent for completion. It also limits anchors to brand, product, or plain URLs and requires appropriate link attributes for paid or incentivized placements. These rules steer the work toward accurate listings and referral value rather than a count of links.
V1 is for a list that has already been checked
SPD V1 Batch handles a larger approved source list through normalization, tracking-parameter removal, duplicate detection, stable queue identifiers, resumable shards, and immediate result recording. The README is careful with the word batch: it refers to planning and queue progress, not uncontrolled parallel form submission. Native CAPTCHA, Turnstile, email verification, 2FA, and passkeys stay with the user or the destination's own flow.
V1 and V2 solve different problems, and the repository keeps them in separate SKILL.md folders. Starting with 5 to 10 V2 candidates is sensible when channel quality is unknown. Moving those findings into a larger V1 queue is reasonable only after product facts, authorization, and destination fit are settled. Otherwise, faster queue mechanics merely process weak or ineligible sites more efficiently.
What happened when we ran it
We did not run backlink_skills at commit 9290e55. The lab classified the repository's language as Python but found no supported executable ecosystem and no Dockerfile, so it produced no install, build, test, dependency, disk, timing, or vulnerability measurement. This is an unmeasured review, and no paragraph on this page should be read as evidence that the included scripts pass.
The README does document 2 Python record-audit commands and says both submission skill folders include tests. It also describes Markdown checks for the LinkedIn, Medium, and WeChat writers. Those are useful signs because instruction-heavy workflows need deterministic checks around states and deliverables. They are still repository claims until executed. A future harness should run the two unittest discovery paths and the auditors against valid and deliberately invalid records.
The authorization model is stricter than typical SEO automation
V2 treats submission as a state machine. Read-only inspection comes first, registration needs its own permission, field entry uses approved facts, and the final submit action requires valid scoped authorization at that moment. The workflow records events and evidence references while excluding passwords, cookies, OTPs, recovery codes, raw email addresses, private paths, and live session links from the public campaign record.
An ambiguous click is not counted as success. The instructions require checks against the backend, mailbox, and public page before any retry. The status model distinguishes submitted, waiting for verification, waiting for approval, published, unknown, failed, ineligible, and unavailable outcomes. That detail is the repository's best work. It prevents duplicate submissions and gives a reviewer a chance to understand what happened without inheriting sensitive browser state.
The writing bundle is broad enough to need editorial judgment
The general writer covers briefs, outlines, fact checking, SEO audits, revisions, images, metadata, and file packaging. The three platform writers alter the process for their destination rather than changing only an output directory. LinkedIn receives business-oriented research and publishing settings, Medium gets publication fit and disclosure rules, and WeChat gets Chinese mobile-reading and evidence requirements. Local output is separate from external publishing.
That breadth can also make the instructions heavy. The general writer's task card alone specifies 10 lines or fewer, articles usually between 800 and 1,500 words, at least 3 subheadings, and 3 FAQ items. Fixed shapes help an agent finish a package, but they can flatten subjects that do not need an FAQ or standard search article. Editors should keep the factual safeguards while overriding formulaic structure when the destination or story calls for something else.
Active files are not the same as current destinations
The repository was pushed on 2026-08-24, and GitHub showed 0 open issues or pull requests and no published release. The MIT license is present. Current source activity is encouraging, but the empty issue queue provides little evidence about how quickly broken destinations, platform changes, or skill failures are handled. Directory availability can change without a code defect, so repository activity and route freshness are separate questions.
Backlink Skills is most useful as a policy kit for Codex. Its 10-site V2 cap, action-level approval, evidence model, and refusal to bypass verification make it safer than generic submission automation. The 743-entry list is the least trustworthy part if taken literally because the maintainers say it contains stale and unsuitable candidates. Start small, recheck every destination, record public outcomes, and judge the campaign by relevant referrals and durable accurate profiles.