content-forecast: An AI Content Performance Forecasting Tool Built on Your Own Baseline

Three historical posts are enough for content-forecast to set a temporary baseline. That is the entire premise of the tool: instead of scoring a script against a generic virality model trained on other people's channels, it scores against the only dataset that actually predicts what will work for you, which is your own back catalogue. The repository, built by a developer going by Colin, has grown from 83 GitHub stars to 196 in the space of a week, the fastest percentage riser in this week's open-source creator-tool radar and its first appearance on AIOS Guide.

content-forecast is an agent skill, meaning a saved playbook that Claude Code or Codex follows rather than a standalone app. It installs with a single script (bash install.sh claude or codex, with a PowerShell equivalent for Windows) and needs nothing beyond a host agent with file access, Python 3 for its calculations, and optional web access for research steps. It is MIT licensed, free, and currently at version 0.3.0, with support for English and Simplified Chinese.

How content-forecast actually works

The tool runs in three stages, and the first one is the part most forecasting tools skip entirely: building a profile of what you specifically know that your audience doesn't. A creator introduces themselves, and content-forecast maps their knowledge and their audience's gaps onto a four-quadrant grid. Common Ground is what both the creator and the audience already know. Gold Mine is the creator's unique, defensible knowledge. Blind Spot is where the creator has a real gap. Frontier is emerging territory neither side has settled yet.

Topic ideas come from intersecting quadrants rather than from trend data. Cross Common Ground with Gold Mine and you get the question a creator is uniquely positioned to answer that their audience is already primed to ask. That is a deliberately different generation method than scraping what is currently trending and reskinning it, and it is the reason the tool's own description frames its philosophy as answering the questions only you can answer, not chasing what is already working for someone else.

The second stage is script diagnosis. Feed content-forecast a draft and it flags the strongest elements, the likely drop-off points, and what revisions would matter most before filming. The third stage is the forecast itself: a prediction of how the finished piece will perform relative to the creator's own historical baseline, expressed as a direction and a range rather than a fake-precise number. The tool is explicit that three historical posts are the minimum to get a working baseline, and that forecasts widen or narrow their range as more real data accumulates. Once a piece is actually posted and the real numbers come in, the forecast for that piece locks and cannot be revised after the fact, which is the detail that keeps the tool honest rather than quietly grading itself on a curve.

Who this is actually for

content-forecast is built for creators and content-led marketers who already have a catalogue worth learning from. A brand-new account with two posted videos does not have the three-post minimum the baseline needs, and the tool will not have much to say to someone who hasn't shipped anything yet. Where it fits is the operator who has been posting consistently, has a rough sense of what has worked before, and wants that intuition formalized into something a script can be checked against before the camera turns on, rather than after the view count comes in.

It is not a replacement for actually watching your own analytics dashboard, and it will not tell you why a specific post underperformed in the way a human reviewing comments and retention curves might. What it adds is a repeatable, structured pass that happens before publishing rather than after, at the moment a script or an idea can still be changed. That is the gap between content-forecast and most of the AI writing and repurposing tools in this category, which generate output but stop short of judging whether that output is likely to land.

Setting it up and what using it looks like

Getting started is a short conversation, not a form. After installing the skill, a creator tells Claude "Initialize Content Forecast" and then introduces themselves in plain language: what they make content about, who watches it, and what they know that their audience generally doesn't. content-forecast turns that into the four-quadrant map and stores it in a separate content-forecast-data/ directory that stays out of any public repository, which matters for anyone using this on a machine where the rest of their work is version-controlled and shared.

From there, the working loop is: bring a script, get a diagnosis, revise if needed, get a forecast, film, post, and let the real numbers feed back into the baseline for the next round. Because the calculations run on Python's standard library with no external dependencies, there is nothing to configure beyond having Python 3 available, and the optional web access is only used for research steps a creator can skip entirely if they'd rather work from what they already know.

The honest limitation

A tool that forecasts against your own baseline is only as useful as that baseline is representative. A creator who just changed format, platform, or niche is effectively starting over, and content-forecast's early forecasts for that pivot will carry wide, low-confidence ranges by design rather than false precision. That is the correct behavior for a tool built this way, but it means the payoff is back-loaded: the more consistent history you feed it, the more the forecasting stage actually earns its keep. For a creator still finding their format, a generic content-strategy checklist will do about as much good. For someone with a real back catalogue who wants their gut feeling about "this one will work" checked before they spend a day filming, content-forecast is doing something none of the more generic AI content-repurposing tools in this space attempt.

Source: content-forecast on GitHub.