<skill name="backlink" version="3.0" body-format="xml"> <why-xml> The frontmatter above stays YAML because the Skill loader reads it for discovery. Everything below is XML because this Skill is mostly laws and routing, and a law that is easy to skim past is a law that gets broken. Tagged blocks make "which rule did I just violate" answerable by name. </why-xml> <mission> One business Skill for the complete backlink lifecycle. Do not split it back apart, and do not create another browser-extension Skill — OpenCLI and its Chrome extension are the connector underneath this Skill, never a separate business workflow. <map> <summary> Two things live here and they answer different questions. **The data files are the asset; the references are how to use them and how not to fool yourself.** </summary> <tree><![CDATA[ backlink/ ├── SKILL.md ← you are here: laws + routing + workflow entry points ├── CONTRIBUTING.md ← how to submit a PR, the data model, the evidence rule │ ├── data/ ← THE DATABASE. Machine-readable, PR-able, CI-checked. │ ├── free-channels.json places that publish a link at no cost │ ├── submission-targets.json routes that ACCEPT a submission — first-pass library │ ├── paid-platforms.json platforms observed carrying purchased placements │ ├── index-submission.json engines that take a URL and publish NO link │ └── schema/ JSON Schema for the files above │ ├── scripts/ ← run these; do not re-derive their knowledge by hand │ ├── validate-data.mjs PR gate. CI runs exactly this. Must exit 0. │ ├── health.mjs run before ANY browser task │ ├── opencli-core.mjs ★ defaultSession(), run(), closeSession() │ ├── lib-tools-share.mjs ★ the ONE panel launcher │ ├── tools-share-open.mjs launch a tool by name; --goto for a deep link │ ├── similarweb-query.mjs performance | channels | similar-sites │ ├── similarweb-batch.mjs bulk traffic screen — one login, N domains, resumable │ ├── semrush-batch.mjs same, on the other card's quota (organic traffic) │ ├── semrush-overview.mjs AS / organic traffic / ref-domains / keywords │ ├── semrush-keyword.mjs one keyword: volume, KD, CPC, per-country split │ ├── semrush-report.mjs the OTHER four no-export reports; reuses one session │ │ table reports paginate — pass --all-pages or it warns │ ├── page-read.mjs render a public page → text, prices, paywall shape │ ├── apply-traffic-screen.mjs write verdicts back into submission-targets.json │ ├── inspect-page.mjs dump one target's form / login / CAPTCHA state │ ├── safe-fill.mjs fill a reviewed payload, never submit │ ├── release-submit-guard.mjs only after explicit per-submission approval │ ├── submit-directory.mjs the single-target driver; one session per staged site │ ├── adapter-phpld.mjs ★ reference implementation of one-session-per-site │ ├── ledger.mjs candidate → … → indexed → rel_verified │ ├── discovery-queue.mjs recursive competitor/commenter expansion │ ├── harvest-commenters.mjs pull commenter domains off an article │ ├── third-party-list-ingest.mjs someone else's list → screened leads + diff │ ├── probe-submission-targets.mjs leads → reachability, route, gate, price │ ├── merge-submission-targets.mjs fold a probe run into the two data files │ ├── targets-select.mjs pick ONE batch: --cohort open | captcha | … │ ├── paid-platform-registry.mjs merge a harvest into the paid registry │ └── harvest-*.{sh,mjs,js} bulk table extraction from logged-in dashboards │ └── references/ ← method, traps, and why the rules are the rules ├── browser-runtime.md ★★ READ FIRST for any browser work. The laws + measurements. ├── traffic-screen.md ★ the qualifying gate, and why it runs before the form ├── submission-lanes.md ★ lanes, cohorts, the three guards, staged queues ├── instant-publish.md ★ free channels: how each class behaves, what kills them ├── paid-platforms.md ★ paid: tiers, why a burst is not a purchase ├── batch-campaign.md ★ 100+ rows: queue, idempotency, resume, reporting ├── index-submission.md index-only channels; why `indexed` must name an engine ├── authorized-data-sources.md the panel, the cards, quota, expiry, the traps ├── field-notes.md what actually blocks submissions in practice ├── harvest.md scraping failures that look like success ├── safety-policy.md read before any fill / submit / logged-in action ├── acquisition-doctrine.md the standing ruling on what is worth pursuing ├── discovery-loop.md · link-quality-rubric.md · analysis-templates.md ├── outreach-templates.md · backlinkdirs.md · prompts.md · credits.md └── LICENSE-analysis-templates-Apache-2.0 ]]></tree> <path-rule>Resolve every path in this file relative to this SKILL.md.</path-rule> </map> <routing> <summary>Match the ask to a starting point. When two rows fit, take the lower one — it is more specific.</summary> <route ask="Somewhere I can post without registering"> `data/free-channels.json` filtered to `account: "none"` and `status: "live"`, then <ref file="references/instant-publish.md"/> for that class's mechanics. Directory submission does NOT satisfy this ask; burning a campaign discovering that is the common failure. </route> <route ask="What paid options exist / where did this competitor buy its links"> <ref file="references/paid-platforms.md"/>, then `data/paid-platforms.json` sorted by how many independent sites were observed using each. </route> <route ask="Find me new opportunities"> <ref file="references/discovery-loop.md"/> — merge whatever you harvest back into the registry. </route> <route ask="Where can I submit this site"> `node scripts/targets-select.mjs --stats`, then one cohort at a time per <ref file="references/submission-lanes.md"/>. </route> <route ask="Is this link profile any good"> <ref file="references/link-quality-rubric.md"/> </route> <route ask="Get these numbers out of a dashboard with no API"> <ref file="references/harvest.md"/> </route> <route ask="Here are 300 directories, submit to them / a campaign that must survive interruption"> <ref file="references/batch-campaign.md"/>. The single-target loop is correct per target and wrong per campaign. </route> <route ask="Someone published a list of backlink sites, is it useful"> `scripts/third-party-list-ingest.mjs` to normalise and diff it, then the "Reading a third-party list" section of <ref file="references/instant-publish.md"/>. </route> <route ask="Submit our pages to Brave / another engine, why is our index count low"> <ref file="references/index-submission.md"/>. It publishes no link, so it never enters the placement ledger. </route> <route ask="Should we post here at all — off-topic host, low DR, known nofollow"> <ref file="references/acquisition-doctrine.md"/> BEFORE rejecting anything. </route> <route ask="Just open this page and tell me what is on it"> <workflow-ref id="explore"/> — still OpenCLI, still a script. </route> <query-the-data> Query the data rather than reading JSON by eye. <cmd><![CDATA[ node -e 'const d=require("./data/free-channels.json");console.log(d.channels.filter(c=>c.account==="none"&&c.status==="live").map(c=>`${c.id}\t${c.kind}`).join("\n"))' node scripts/paid-platform-registry.mjs list --min-sites 2 ]]></cmd> </query-the-data> </routing> <browser-runtime> <summary> `$backlink → scripts and policy → OpenCLI → the owner's authorized Chrome → website` </summary> <default-driver> OpenCLI is the default for **everything**, including a quick ad-hoc look at one page. It reaches the owner's Chrome through an extension plus a local daemon, and because it is a CLI, any agent runtime that can run a shell command gets the identical capability — Claude Code, Codex, anything else. Work done through a runtime-specific tool cannot be replayed from a script or from another agent later, which defeats the reason this Skill has scripts. <law id="one-session-one-tab" weight="load-bearing"> <statement> `opencli browser &lt;session&gt;` is a one-page abstraction. **A session name owns exactly one tab.** Different names never steal from, switch, or pollute each other. So N pages need N session names. </statement> <why> This inverts the intuition most people arrive with, which is why it is stated first. Measured 2026-08-21 under three concurrent agents: distinct session names produced **zero** cross-agent thefts across 4 rounds × 3 pages; three agents sharing the name `work` produced 3, 12, and 2 thefts, one of them missing on every check it made. Re-confirmed the same day against this Skill as written: three agents told only to follow it scored **36/36 clean with zero leaked tabs**. </why> <correct><![CDATA[ opencli browser recon-sw-notion --window background open "https://..." opencli browser recon-sw-figma --window background open "https://..." opencli browser recon-sem-rival --window background open "https://..." ]]></correct> </law> <law id="no-multi-tab-api" weight="load-bearing"> <statement> Do not use `tab new`, `tab select`, or `open --tab` to hold several pages under one session. All three fail, and every one fails **silently** — the command reports success and the next read returns the wrong page. </statement> <why> Measured 2026-08-21 on opencli 1.8.6: a session tracks only its newest tab, so earlier ids drop out of `tab list`; `tab select` returns success with no effect on reads; `open --tab &lt;id&gt;` opens a **new** tab and leaves the named one untouched; and `get` does not accept `--tab` at all, so a run using `get url` to confirm its position cannot be right about it. One three-agent run took the owner's Chrome from 11 tabs to 30 orphans. </why> <instead> `--tab` works on `open`, `state`, `extract`, `find`, and `click`. When a read must name its target, use `state --tab &lt;id&gt;`. <confirm-identity> The canonical check after every navigation, and the one Law 4 exists to make possible: <cmd><![CDATA[ opencli browser "$S" --window background get url # one page per session: safe opencli browser "$S" --window background state # same, plus title + elements ]]></cmd> </confirm-identity> </instead> </law> <law id="no-literal-session-name"> <statement> Never write a literal session name as a default. In JS use `defaultSession(base)` from `scripts/opencli-core.mjs`; in shell use `SESSION="backlink-$$"`. </statement> <why> "Another task stole my tab" is never the CLI round-robining — it is always two tasks that picked the same name. The commonest source is documentation: `opencli browser --help` opens with `opencli browser work open https://x.com`, so every agent copying the example lands on `work`. This Skill caused the same failure itself when `tools-share-open.mjs` defaulted to `backlink-panel`. </why> <code><![CDATA[ const session = flags.session ? validateSession(flags.session) : defaultSession('backlink-work'); ]]></code> <subagent-trap> Subagents inherit the parent's environment, so several agents spawned inside one conversation resolve to the same default. When fanning browser work across parallel agents, give each an explicit `--session` or a distinct `OPENCLI_SESSION_SUFFIX`. </subagent-trap> <naming> Make names **describe the work**: `backlink-probe-&lt;suffix&gt;` beats `bl-1`. The session name is the only identifier that exists — the Chrome tab group carries one extension-supplied label for every session and cannot be changed from here, so never ask a person to tell tasks apart by looking at it. <law id="claim-handles-first"> <statement> Open every session you need up front and capture every handle, then start the work loop. Do not interleave creation with use. </statement> <why> Every driver tested shares one race window: the stretch between creating a page and holding a stable handle to it. Two independent runs lost pages in exactly that gap, because a bare `open` with no established handle resolves against whatever "current" happens to mean at that instant. </why> </law> <law id="background-by-default"> <statement> Default every session to background mode. The flag sits **between** the session name and the subcommand: `opencli browser &lt;session&gt; --window background &lt;command&gt;`. </statement> <why> Background mode runs the owner's real logged-in Chrome without raising the window. It is **not** headless — `navigator.webdriver` is `false`, the UA carries no `Headless`, `plugins.length` is 5, and `visibilityState` is `visible`. So "background will trip the site's bot defences" is not a real concern, and there is never a reason to reach for foreground to look more human. <other-drivers> <driver name="agent-browser" verdict="no logged-in identity, ever"> It attaches over CDP, and CDP cannot reach the owner's Chrome: Chrome 136+ silently ignores `--remote-debugging-port` on the default user-data-dir (verified on 151 — the flag is passed, no port is opened), and macOS TCC blocks copying the profile out. Relaunching Chrome is wasted effort; do not suggest it. Its `--profile` means a separate directory you log into once, unrelated to the owner's sessions. Use it only for tasks needing **no** logged-in identity, and address tabs by `--label`, never by the `t1`/`t2` positional index, which is a shared namespace across agents. </driver> <driver name="Claude in Chrome" verdict="single agent, ad-hoc, prefer OpenCLI anyway"> It reaches the owner's Chrome but has no isolation boundary of any kind: one flat tab group shared by every concurrent agent, and omitting `tabId` resolves to "first tab in the shared group". It also has a reproducible bug where closing one of your own tabs tears down your session's tab-group tracking and orphans the rest. It is Claude-only, so anything built on it cannot be replayed from another runtime. </driver> </other-drivers> <preflight> <cmd>node scripts/health.mjs</cmd> Run before browser work. Use `--check-update` only when the user asks about versions; an available update is informational, and upgrading OpenCLI needs a separate request. <data-sources> <terminology lang="zh"> **当用户说「数据面板」「数据勘测」「查一下数据」「用 Similarweb 看看」「Semrush 拉一下」， 指的都是同一件事：走那个共享账号的代理面板，用 Similarweb 或 Semrush 查。** 这两个产品是这里唯一的第三方数据源，没有别的候选，不需要反问用户指的是哪个平台。 </terminology> <division lang="zh"> 分工固定，按问题类型选，一次只开一个： <workflows> <workflow id="explore" when="the user just wants to see what is on a page"> <statement> Ad-hoc looking is still scripted work. Use OpenCLI so the look is replayable. </statement> <cmd><![CDATA[

backlink-analyzer

browser-harvest

opencli

get url

--tab

backlink-probe-$$

backlink-probe-p1-$$

-p2-$$

-p3-$$

opencli browser --help

opencli browser work open https://x.com

--help

--window background

opencli browser <session> close

close

tab list

unknown command: <yoursession>

opencli browser s open URL --window background

--help

about:blank

lib-tools-share.mjs

DEFAULT_DASHBOARD

.env

lib-tools-share.mjs

similarweb-query.mjs

<workflow id="discover" when="the user wants new opportunities"> <read><ref file="references/discovery-loop.md"/></read> <method> Recursive discovery: seed competitors → get their backlink rows from an authorized Semrush/Ahrefs export or logged-in browser → classify source URLs (editorial, resource, directory, profile, comment, login wall, paid, CAPTCHA, rejected) → harvest commenter domains on real article pages → feed those back into the queue → repeat to a bounded depth. Rank by topical fit, page quality, moderation, public visibility, and referral potential. Low-quality comment volume is auxiliary, never the goal. </method> <cmd><![CDATA[ node scripts/discovery-queue.mjs seed --file .backlink/discovery.json --domain competitor.com

page-read.mjs

curl | grep

scripts/

<check level="1" compares="rawText vs parsed.rows.length"> Count the record-shaped lines in `rawText`, compare with `parsed.rows.length`. A gap here means **your regex has a blind spot** — the rows arrived and you dropped them. This is the silent, dangerous one. Fix the parser. </check> <check level="2" compares="the page's own headline count vs rawText"> Semrush prints its own total (`自然搜索排名: N`). If that exceeds what `rawText` even contains, the rows **never reached you**: these tables are virtual-scroll and only mount a fraction at a time, so a full pull needs the export, which costs quota. This is a known ceiling, not a bug — say so rather than "fixed the parser". </check> <workflow id="screen" when="before filling anything, always"> <statement> The qualifying test is real traffic (`&gt;= 100` monthly visits), never DR, and it runs BEFORE the form does. </statement> <read><ref file="references/traffic-screen.md"/></read> <cmd><![CDATA[ node scripts/similarweb-batch.mjs --domains-file domains.txt --out sw.jsonl node scripts/apply-traffic-screen.mjs --in sw.jsonl --source similarweb node scripts/targets-select.mjs --cohort open --min-traffic 100 ]]></cmd> <headline> Measuring a domain costs one query; filling its form costs two orders of magnitude more. One run filled every form across a 73-domain family and only then sampled five for traffic — every filled form was discarded. </headline> </workflow> <workflow id="submit" when="a route exists and the target passed the screen"> <read><ref file="references/submission-lanes.md"/></read> <inspect> Inspect every target independently. Never infer a form from a sibling site. A page is fillable only when there is one unambiguous qualifying form and no detected CAPTCHA/login wall. <cmd><![CDATA[ node scripts/inspect-page.mjs --session "inspect-$$" --mode comment \ --url https://example.com/article --out .backlink/scan.json ]]></cmd> Modes are `comment`, `directory`, or `auto`. </inspect> <payload> Create a reviewed JSON payload with truthful values. For comment mode, `description` is the comment body. <cmd><![CDATA[ { "url": "https://owned.example/relevant-page", "name": "Real owner or product name", "email": "owner@example.com", "description": "A page-specific, useful comment or truthful listing description" } ]]></cmd> </payload> <fill> <cmd><![CDATA[ node scripts/safe-fill.mjs --session "fill-$$" \ --scan .backlink/scan.json --payload .backlink/payload.json ]]></cmd> It revalidates the URL, form identity, field semantics, login state, and CAPTCHA state, installs a submit guard, and never submits. The human reviews the rendered page and performs final submission. Only after the user explicitly authorizes one exact reviewed submission may the agent run `release-submit-guard.mjs` — and releasing the guard still does not click Submit. </fill> <staged-queue> Lane B leaves forms on screen for the owner to finish. **One session name per staged site** — a session owns one tab, so reusing one session overwrites the previous staged form while the report still says N staged. `adapter-phpld.mjs` carries the reference implementation. </staged-queue> </workflow> <workflow id="analyze" when="the user has exported backlink data already"> <statement> Analyze referring-domain quality and topical relevance; suspicious networks, sitewide links, and toxic patterns; anchor and target-page diversity; follow/nofollow/UGC/sponsored distribution **when observed**; competitor gaps and prioritized next opportunities. </statement> <read> <ref file="references/link-quality-rubric.md"/> — scoring, toxicity, disavow. <ref file="references/analysis-templates.md"/> — report shapes. <ref file="references/outreach-templates.md"/> — frameworks; sending needs the user's explicit approval per message. </read> <hard-limit> These templates assume you already have the data. They do not fetch it. **A report built from templates alone, with no observed rows behind it, is fabrication.** Do not disavow links, contact site owners, or change production sites unless the user separately asks. Treat third-party authority and traffic estimates as directional and time-sensitive. </hard-limit> </workflow> <workflow id="harvest" when="the numbers are visible in a logged-in dashboard with no API"> <read> <ref file="references/harvest.md"/> before writing any scraping loop. It documents failures that produce **plausible, silently wrong output**: virtual scroll tables that are not `&lt;table&gt;` and drop rows without erroring, long URLs that make whole rows vanish, execution-channel timeouts that look like failure while the page loop is still running, and Chrome's intensive throttling stretching a four-second loop into twenty-five minutes. </read> <cmd><![CDATA[ sh scripts/harvest-collect.sh # wait for downloads to settle, then collect node scripts/harvest-merge.mjs # merge by field shape, refuse duplicate files ]]></cmd> <note> `scripts/harvest.browser.js` is the in-page collector. Its output arrives via a Blob download rather than a return value, because the execution channel truncates at roughly 1 KB. </note> </workflow> <workflow id="verify" when="closing the loop on any placement"> <states>candidate → qualified → drafted → filled → submitted → public → indexed → rel_verified</states> <cmd><![CDATA[ node scripts/ledger.mjs upsert --file .backlink/ledger.json --url https://target.example/page node scripts/ledger.mjs transition --file .backlink/ledger.json \ --url https://target.example/page --state public \ --evidence "Observed the exact public anchor on 2026-07-30" ]]></cmd> <evidence-bar> `submitted`, `public`, `indexed`, and `rel_verified` each require an evidence note. Never promote a record from a filled form, a pending notice, or a historical assumption. **`indexed` must name the engine** — `indexed@google`, `indexed@brave`. An unqualified "indexed" is a claim about the whole web built from one crawler's opinion. </evidence-bar> </workflow> </workflows> <rules type="non-negotiable"> <rule id="no-coordinate-clicking">No coordinate-based "human-like" clicking.</rule> <rule id="no-bypass">No CAPTCHA, Turnstile, login, paywall, quota, or account-scope bypass.</rule> <rule id="unmeasured-is-not-qualified"> Never treat "not yet measured" as "qualified". The traffic gate only works if unmeasured rows are excluded from a batch rather than waved through. `--min-traffic` drops them by design; `--unmeasured` lists them as the next screening queue, never as a batch. </rule> <rule id="validate-gates-against-known-bad"> A gate metric is validated against known-bad domains, never against famous ones. Any signal a link network can manufacture for itself — DR, popularity rank, index size — will pass a farm. Tranco's top-1M failed exactly this way: 48 of 73 confirmed farm domains sat inside it, from rank 134k to 998k. </rule> <rule id="no-fabrication"> No generic praise, fake identity, invented metrics, or a comment body that ignores the article it sits under. Never invent a product fact to fill a field — founder, pricing, address, launch date, user count, ownership, legal, contact. Leave optional unknowns blank and stop a row whose required field is unknown. </rule> <rule id="relevance-ranks-never-gates"> A host site on a different topic is fine. Relevance and DR **rank** candidates, they never **gate** them, and `nofollow` is an observation to record rather than a reason to skip. Read <ref file="references/acquisition-doctrine.md"/> before rejecting any target on quality grounds. </rule> <rule id="no-link-farms"> No link farms, spam generators, adult/malware surfaces, hidden reciprocal links, temporary eligibility pages, or cloaking. Two identical give-aways in one place — one site script across dozens of domains, and a promotional sentence repeated word for word — mean one operator. Submitting to N of its domains buys one link's value while accruing N times the footprint. </rule> <rule id="submission-is-not-a-backlink"> Do not record a submission as a backlink. This includes handing a URL to a search engine: that is an index-submission channel, it publishes no link, and it belongs in `data/index-submission.json` rather than the placement ledger. </rule> <rule id="observe-before-recording"> Do not record `follow`, `nofollow`, `ugc`, `sponsored`, or `indexed` without observing it for the exact URL. A click, a completed registration, a saved draft, a form that cleared itself, or a generic thank-you URL is **not** evidence of a submission — those record what you did, and the ledger records what the site did. </rule> <rule id="never-retry-ambiguous"> Do not automatically resubmit an unconfirmed target. Never retry an ambiguous final action — one where the submit happened and the result was not observed. Check the account backend, then the mailbox, then the public page. That state is `outcome-unknown`, and it is not a failure. </rule> <rule id="anchor-policy"> Anchor text is the brand, the product name, or the naked canonical URL. Never request dofollow treatment, never repeat a commercial exact-match anchor across a campaign, and treat a paid or incentivised placement that publishes as a plain follow link as **noncompliant** rather than as a win. </rule> <rule id="secrets"> Records carry aliases and evidence IDs. Passwords, OTPs, recovery codes, cookies, OAuth parameters, magic links, raw session IDs, raw email addresses, and phone numbers belong in none of them. Keep raw cookies, tokens, authorization headers, and credentials out of logs. </rule> <rule id="traffic-figures-need-six-fields"> A third-party traffic figure without `source · metric · month · geography · device · date verified` is not a number. Store all six or store none. </rule> <rule id="http-over-mcp"> Prefer a documented HTTP endpoint over an MCP server when both serve the same data from the same quota — the MCP adds a connection and a process without adding capability, and a failure there is harder to tell apart from the service being down. Keep MCP where it is the only authorized channel; never retire a working path before the replacement has run successfully once. </rule> <rule id="verify-before-trusting-a-row"> Records carry `lastVerifiedAt` because this genre dies faster than it changes. A channel that worked three months ago may be gone, gated, or `noindex` today. Re-verify before a campaign; the validator warns on anything `live` older than 180 days. **Fixing a wrong row is worth more than adding a new channel.** </rule> <rule id="two-tables-two-claims"> `free-channels.json` records **a published link on a live page** and requires `relObserved`/`anchorRendered`. `submission-targets.json` records **a submission route that exists** and the validator rejects those fields there. A target graduates from the second into the first the moment an actual anchor is observed; until then it makes no promise about `rel`, anchor text, or indexability, and the report must not imply one. </rule> </rules> <escalation> <summary>Read the reference **before** acting, not after the run goes wrong.</summary> <when trigger="any browser work at all">references/browser-runtime.md</when> <when trigger="any fill, submit, account, or logged-in action">references/safety-policy.md</when> <when trigger="a supplied list of 100+ rows, or anything that must survive interruption">references/batch-campaign.md — the single-target loop deduplicates too late, stalls behind the first CAPTCHA, cannot tell an interrupted row from an unstarted one, and produces a number that counts forms instead of links</when> <when trigger="a first submission campaign">references/field-notes.md — personal-contact requirements outrank CAPTCHAs, and landing-page CAPTCHA scans give false negatives</when> <when trigger="someone hands you a 'places to get backlinks' list">the "Reading a third-party list" section of references/instant-publish.md — a Dofollow column is an assertion about a platform, never an observation of a link</when> <when trigger="the ask is about paid placement">references/paid-platforms.md</when> <when trigger="the ask is about getting pages into an index rather than getting a link">references/index-submission.md</when> <when trigger="about to reject a target on quality grounds">references/acquisition-doctrine.md</when> <when trigger="BacklinkDirs eligibility">references/backlinkdirs.md</when> <when trigger="the user wants a ready-to-copy prompt">references/prompts.md</when> <when trigger="whose work is this built on">references/credits.md</when> </escalation> <output-contract> <item n="1">data sources and authorization boundary</item> <item n="2">candidates by type, and the reason for qualification or rejection</item> <item n="3">current ledger state, never an inferred later state</item> <item n="4">evidence links or local evidence files</item> <item n="5">the next safe action, and whether human review or submission is required</item> </output-contract> <install> Source: [Skills.sh](https://skills.sh/yan-labs/yan-skills) <cmd><![CDATA[ npx skills add yan-labs/yan-skills --skill backlink -g -y # first install npx skills update backlink -g -y # update ]]></cmd> For a project-level install omit `-g`; update with `npx skills update backlink -p -y`. </install> </skill>

--session

semrush-report.mjs

sessionReused: true