Content operations is three gates in sequence: what to make, whether it is good, and whether it ships. Each gate is a decision with a different cost of error, and each benefits from being written down as criteria rather than left to taste.
The tools below do not write the content. They rate it, route it, and check whether the claims hold. That is the right division of labor: a model that generates prose can also generate the justification for the prose, which makes review harder rather than easier.
Briefs are decision states
A brief is not a topic. It is a state: audience, search intent, the evidence available, the claims that are off-limits, the format, and the action the reader should take. Once those are fields, the interesting questions become typed: does this draft address the stated intent, does it contain an unsupported claim, is it differentiated from the existing page, and is it ready to publish.
The use case map lists ten decision shapes, including classification, ranking, verification, and structured extraction, across nineteen industry sections. It is a map of examples, not products or benchmarks, but it is useful for naming the gate you are building. A content brief usually needs a mix: a Choice for intent, Scores for differentiation and evidence, and Nouls for specific policy checks.
Editorial QA is a rubric, not a review
jev-writer rates each published post with thirty Jev judgments against a pre-registered rubric, then correlates those ratings against impressions, reach rate, and conversion rate. It enforces an observational power gate, date-confound controls, and Benjamini–Hochberg corrections, and it refuses to state findings the sample cannot support.
That statistical layer is the difference between a rubric and a vibe. The README stresses that the tool analyses posts and does not write them, that ratings are Jev judgments correlated with observed engagement rather than a causal model, and that LinkedIn is the only platform integrated today. It is MIT-licensed with 83 passing tests, but the README's known-limits and costs sections were truncated in the captured copy, so budget and setup need checking against the repository.
The composite scoring pattern is the right way to build the rubric underneath. Ask four Score questions in one parallel request, each with named levels, normalize each by dividing by the number of levels minus one, and apply weights in code. The pattern's resume example uses 40/10/40/10 for a senior IC and 15/40/20/25 for an engineering manager. The weights are policy, not model output, and should be revisited when the role or the cost of a bad hire changes. The same is true of a content rubric: clarity, originality, evidence, and usefulness are separate scores combined by an editorial policy you can inspect.
Publishing gates and the launch post
The Jev Typewriter launch is a useful caution about what a launch post proves. Steve Krouse posted a single line linking to a live demo and invited readers to try it. The captured thread contains no description of the questions asked, the state sent, or the results shown. The only cost reference is a reply mentioning a $2 budget, and the only substantive criticism (that it could be faster) goes unanswered.
That is not a reason to avoid demos. It is a reason to separate the demo from the documentation. A publishable launch needs the question set, the state, the failure cases, and a reproducible way to run it. Otherwise the post is a link, and the reader cannot tell what was decided or why. The same standard applies internally: a publishing gate that says "looks good" is not a gate.
SEO: code finds, the model judges
jev-seo audits a business site with a deterministic crawler and narrow typed Jev questions. The project's rule is explicit: code finds, Jev judges, the dashboard shows the probabilities. Findings are ranked only when the answer is decisive.
The boundary is the important part. Jev has no index, no search-volume data, and no backlink graph, so nothing in the tool invents a monthly search count. That is a deliberate constraint, not a missing feature. The package is jevseo for Node 20.11 and later, the repository badge marks it UNLICENSED, and no benchmarks are published. Use it as a pattern for a crawl-plus-judgment audit, not as a source of market data.
The internal linking tool applies the same split to link placement. It reads robots.txt and the sitemap, fetches up to 2,000 pages, maps existing internal links with anchor text and position, and uses Jev to classify each page and its search intent. Candidates come from TF-IDF similarity: for each source page, the eight most similar pages are considered, and pages with fewer than three links get a boost so orphan pages gain links. Runs up to 100 pages are free; larger runs cost one credit per 500 pages. The page publishes no accuracy figures for Jev's classifications, so treat the suggestions as a ranked reading list.
Distribution: find demand before writing
The Reddit Content Opportunity Finder is a two-stage pipeline. Stage one cleans an Apify export: drops bots and deleted comments, strips Markdown, URLs, and HTML entities, and converts timestamps to UTC. Stage two embeds comments with all-MiniLM-L6-v2, reduces with UMAP, clusters with HDBSCAN, and names clusters with Gemini or a TF-IDF fallback. Jev then scores each comment from 0 to 1 for whether the brand could publish new, useful content about it.
A comment is flagged when the score clears 0.8 and no competitor is mentioned. The README warns that Jev answers a yes/no question, so scores tend to bunch high, and recommends labeling about 100 comments and choosing the threshold from precision and recall instead of trusting 0.8. The repository has no license and depends on an Apify export plus Gemini, Google Natural Language, and TypeSafe keys. The pipeline is a useful template; the default threshold is not a policy.
The GTM cookbook takes a similar approach across fifteen outbound recipes, from ICP scoring and job-change interpretation to cold-email reply classification and lead routing. Its headline figure (16,711 connections scored for $0.73) is the author's own run, not a bill or an independent benchmark. The framing is the durable part: code does retrieval, counting, and enrichment; Jev makes the judgment calls.
Measure the outcome, not the score
A rubric score is an input, not a result. jev-writer's power gate and false-discovery corrections exist because it is easy to find a correlation between a writing dimension and engagement in a small sample and call it a finding. The same discipline applies to SEO and distribution tools: decide in advance what outcome you are measuring, how many observations you need, and what result would change your mind.
The confidence page adds the operational version. Confidence is a statistic derived from the probability distribution, not a guarantee, and thresholds should be set per action. A publishing gate can auto-approve a low-risk page at one threshold and require a person for a claim-heavy page at another. The scoring without measuring article covers how to keep those bands honest.
Keep the weights and thresholds visible
A content system accumulates policy: rubric weights, opportunity thresholds, link-suggestion cutoffs, and the point where a draft goes to a human. Put them in one place, version them, and review them when the strategy changes. The search visibility article covers the SEO side.
The test of a content operation is not whether the model agrees with the editor. It is whether the system can explain why a piece was approved, why a suggestion was ranked first, and why a claim was flagged. If those answers live in code and labeled examples rather than in a prompt, the operation can improve without rewriting its own justification. That is what makes the gates reviewable.
