What happened
The README for the public GitHub repository tipsai/product-seo describes an MIT-licensed skill for running SEO on a product website. It works from the site's own Search Console data, starting with a CSV export of performance for the previous three months. The skill sorts queries into two groups: topics where the product is a plausible answer and topics likely to attract readers who will not sign up. It then proposes the first three actions, with pages ranking around positions 8–20 among the initial candidates. The README says paid SEO tools are optional; a new site without Search Console history can instead connect Search Console and Bing, check indexing and signup measurement, and plan an initial three to five pages.
The workflow covers query triage, content ideas, audits of existing pages, rewrites, titles and descriptions, task pages and free tools, comparisons, listings, and AI-search visibility. It can recommend consolidating duplicate or thin pages, but it does not publish, redirect or delete without the user's approval. It also does not submit forms or send emails: it drafts those materials for a person to send. For rewrites, the README describes a separate review by a second model and a reader check on screenshots. Product facts are kept in a dedicated file, and competitor claims are collected from official pages with a quote, link and date; the README says those facts are checked again after two weeks.
The skill records a change, its backup and its planned review in a project seo/ folder. It calls for checks at one and three weeks. Its stated rollback rule is to restore the prior version if clicks are down by 30% or more and the position has not recovered. The README also describes why it waits: one batch of rewrites initially fell for weeks before finishing above its starting point. The skill writes operating notes into the project so later sessions can read site decisions, change history, lessons, product facts and competitor sources. It requires Node.js 18 or newer for its helper scripts, which use no packages; without Node, the agent can work from exports.
The repository says the skill was developed while working on HypeScribe. From August to early November 2026, that site recorded about 1 million Google impressions and 9,000 clicks; its click-through rate rose from 0.4% in spring to 0.9%, and clicks were almost twice those of the preceding three months. The README explicitly frames those figures as the outcome on one site, not a promise for other users.
What the archive adds
New Runtime's earlier coverage of Geo-sleuth and print3d examined skills that package specialist work as reusable agent capabilities. This is a new event on that broader thread: product-seo applies the same packaging model to SEO operations and makes its review and rollback process explicit.
Trend signal
This supports skills as a portable capability layer. The repository offers installation paths for Claude Code, Codex, Cursor and agents that read a shared skills folder, while keeping the operating procedure in project files that later sessions can reuse. The portability claim is practical rather than universal: the README provides different installation routes and does not demonstrate identical behavior across every supported agent.
Why now
The repository says the skill grew out of three months of SEO work on a live product site. Its process addresses a timing problem from that work: a rewrite can dip before it improves, so an early signal alone can trigger a premature reversal. The one- and three-week checks and preset rollback rule turn that experience into an operating procedure.
Why it matters
SEO changes are easy to make quickly with an agent and difficult to evaluate if teams have no baseline, ownership or reversal rule. This skill makes those controls part of the workflow rather than leaving them to memory: it records a backup and check date, asks for approval before publication, and sets a specific condition for restoring the old version. Its query split also directs effort toward search demand that may plausibly produce signups, not just impressions or clicks.
The approach is useful beyond SEO when an agent edits live material with delayed effects. Teams can borrow the structure without assuming the reported growth will transfer: define the source data, separate valuable work from noise, get a human decision at the release boundary, and agree beforehand on when to keep or undo a change. The repository's own evidence is one site's experience, so teams still need to validate the rules against their business and traffic patterns.
Limits and uncertainty
The repository reports results for one site and does not establish that the same workflow or thresholds will produce similar gains elsewhere. It does not provide an independent evaluation of the skill's recommendations, the accuracy of its query prioritization, or the quality of generated edits. Search performance can vary with seasonality, competition, site changes and other factors, so the stated 30% click threshold is a rule in this tool, not a generally validated SEO standard. The repository is public and MIT-licensed, but its installation guidance does not establish that every listed agent has been tested to the same degree.
Skill to build now
For any agent-assisted change to a live page, keep a dated copy of the prior version, write down the intended outcome and set a review date before publishing. Require a human approval step, then compare the agreed metric at a defined checkpoint; only roll back when a pre-agreed threshold is crossed and the recovery condition is met. The skill's one- and three-week checks and 30% click-drop rule are concrete starting points to adapt, not universal defaults.
Reader takeaway
The useful release here is not an SEO checklist but a controlled loop for agent-made changes: work from first-party data, keep humans at the publish boundary, record what changed, and decide in advance how to judge or reverse it. Treat the HypeScribe figures as a single-site case, then test whether the method's prioritization and thresholds fit your own product.