Use programmatic-seo for related pages such as integration guides, template directories, or local service information. The central question is whether each address answers a distinct question with maintainable evidence, not how many URLs can be generated at once.
Install and invoke
Run this in your project terminal to install only this skill for Claude Code:
npx skills add coreyhaines31/marketingskills --skill programmatic-seo -a claude-code
Node.js and npx must be available. For Codex, replace -a claude-code with -a codex. In the agent conversation, explicitly ask it to use programmatic-seo and supply your material. The installation command belongs in the terminal, not in your content brief.
Start with records, not a huge keyword list
Give the agent ten to twenty real records. Explain their origin, permitted use, and update frequency, then describe why a reader would search by this dimension. An integration page needs distinct connection requirements and instructions, rather than the same product description under another name.
Example brief:
Use programmatic-seo to evaluate the integration directory in integrations.csv.
Choose three contrasting records and design sample pages before generating the full set.
For each, show the unique task, required fields, missing evidence, title, sections, and related links.
Flag records that cannot yet answer the reader’s question as unsuitable for publication.
Check whether the search pattern corresponds to a real task and whether differences between pages are meaningful. Directories, comparisons, templates, locations, and glossaries require different inputs. Start with the pattern your evidence can support. Leave unresearched demand estimates unknown.
Turn the sample into a quality gate
Define the target question, unique introduction, main data section, prerequisites, and relevant links. Request three populated examples and identify which passages change with the record. If removing the name makes the pages nearly identical, improve the underlying data or consolidate the content.
Specify which missing fields prevent publication, who maintains stale records, and what happens to URLs when records disappear. Important pages should be reachable through a directory or contextual links, not solely through a sitemap. Unfinished combinations should not enter the public sitemap automatically.
Google emphasizes accuracy and added value when using generative content. Producing many unhelpful pages can fall under scaled content abuse. AI assistance with organization does not remove the need to verify facts. After a small release, check whether readers can complete the intended task, then use indexing, queries, and page behavior to decide what deserves expansion. A functioning template is not proof of a validated search opportunity.
Test a search pattern with a few sample pages
Choose a repeated question with real demand, such as differences between two tools or a template for a recurring task. Then verify that each variable has distinct, accurate data. Keyword count is not user value. If pages would only swap a city or product name around identical prose, do not generate them yet. Ask the skill for a sample record listing page-specific facts, the decision the reader must make, data source, update process, and internal links.
An example brief: “We have official capability and limitation data for thirty integrations. Select three that differ meaningfully and draft sample pages. Each must explain who it fits, how to connect it, and important limits. Separate documented facts from open questions; do not scale to thirty before human review.” Three pages are a sensible test size, not a mandatory step built into the skill. If one integration lacks substance, keep it in a directory or research queue instead of publishing a repetitive profile.
Before launch, inspect URL rules, directory-to-detail links, breadcrumbs, and the sitemap. A page should be reachable through real navigation. After publication, examine specific queries, indexation, and whether readers continue their task before expanding. Without unique data and a maintained source, more pages only multiply the original gap.
Example: decide which integration pages should not publish yet
Suppose a spreadsheet lists one hundred integrations, but only twelve have setup steps, required permissions, and limits. Give the skill the source table and three existing pages. Ask: “Use programmatic-seo to define a minimum evidence bar for an integration guide. Classify six sample records as publishable, research needed, or directory-only, and identify the user question each lacks. Improve one sample page and its internal links before generating anything in bulk.”
For each page, ask whether a reader could connect the integration and understand when it does not fit. “Supports X integration; contact us” is not a useful standalone page simply because X is a distinct keyword. If the agent proposes filling eighty-eight thin entries with the same template, stop scaling and return to data collection. Reconsider the page family when more integrations have distinct instructions and a maintenance owner.
What the public audit actually shows
In a public walkthrough, Eric Siu ran programmatic-seo against 440 Single Grain location pages and Search Console data, found weak clicks for those pages, and received a new persona-by-industry direction. It shows a diagnosis of an existing page family using search data, not later rankings for new pages. Before scaling a proposed pattern, sample rendered pages for empty fields, repeated prose, internal links, and whether each page answers its own user question.