Menú
Extract structured data from an HTML table, optionally walking through paginated views. Returns a list of row dicts keyed by header cells. The demo skill (used to monitor conference paper lists).
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
| name | extract-table-pagination |
| description | Extract structured data from an HTML table, optionally walking through paginated views. Returns a list of row dicts keyed by header cells. The demo skill (used to monitor conference paper lists). |
| version | 0.2.0 |
| allowed-tools | ["wait_for_dom_ready","wait_for_selector","extract_table","click","wait","assert"] |
| metadata | {"applies_to":"any-website","url_patterns":["*"],"dom_markers":["<table","role='table'"],"flake_rate_target":0.08,"flake_rate_measured":null,"exercised_on":["wikipedia-list","arxiv-listing","hacker-news-list","openalex-search","weather-gov-forecast"],"cost_budget":{"deterministic_only":false,"max_vision_calls":1},"sensitive":false,"evaluate_success_criteria":true} |
When the user task involves reading structured data from an HTML table — the page either shows a <table> element, or a role="table" ARIA structure. If the table is paginated, the skill walks pages via a "next" link/button until either the next link is absent, the row count stops growing, or a configurable max_pages is reached.
The matcher gives this a high score when the DOM summary contains <table and the visible-text sample has table-like indicators (numeric columns, repeated short text).
For paginated tables, chain with the pagination-next-page skill in a loop. The runner's vars map preserves bindings across skill invocations so the agent can accumulate rows.
Pseudo-recipe the agent assembles:
while page <= max_pages:
invoke_skill("extract-table-pagination") → $rows_page_N
accumulated += $rows_page_N
if no "next" link visible: break
invoke_skill("pagination-next-page")
We deliberately keep extraction and pagination as separate skills so the agent can interleave with other steps (rate-limit pauses, filter applications, deduplication).
<div>-grid layout (card grid, masonry, etc.) — use the future extract-card-grid skill, or vision fallback.handle-infinite-scroll with row-extraction inside the scroll loop instead.<table> from 2003-era HTML (no <th>, no semantic structure). The extractor will produce noisy results; better to extract by text-selector and structure manually.handle-infinite-scroll or scroll-and-extract.extract_table mode="rowspan-aware".<thead>, no <th> (~5% of legacy sites): the extractor uses the first row as headers; if that row is data, the first record is lost. Workaround: pass headers=[...] arg (planned v0.2.0).pagination-next-page aborts when row hashes match.Agent task: "Get the latest cs.AI papers from arxiv."
→ invoke_skill("verify-page-loaded") → success
→ invoke_skill("extract-table-pagination") → success, $rows_page_1 = [{title, authors, abstract}, ...]
→ agent inspects $rows_page_1 and reports
verify-page-loaded — run first; tables on slow-rendering pages need readiness checkpagination-next-page — chain to walk multiple pageshandle-infinite-scroll — alternative when content lazy-loads instead of paginatingdismiss-cookie-banner — usually run before any extractionSigns into a site using credentials provided via `vars` (typically sourced from environment variables) or via a pre-saved Playwright persistent context. Never harvests or stores credentials. See docs/ethics.md.
Detects whether the current page contains a captcha challenge (reCAPTCHA, hCaptcha, Cloudflare Turnstile, GeeTest, etc.). Reports detection and stops. Never attempts to solve. See docs/ethics.md.
Confirms that the current page has finished loading and is ready for interaction. Detects loading spinners, skeleton screens, and content placeholders. Run before extraction skills.
Fills a multi-step form (stepper wizard) by routing labeled fields to values from `vars`, clicking "Next" between steps, and submitting at the end. Pairs with the `searchable-dropdown` and `date-picker-widget` skills for compound widgets.
Uploads a local file to an HTML5 file input, or triggers a download and captures the resulting path. Composes with multi-step-form for file-attaching workflows.
Picks a date in a calendar-style date-picker widget — clicks the input, navigates months if needed, clicks the target day. The famous "vision agent failure case" that motivates this whole library.