What happened
The GitHub repository describes Not Obvious as an MCP server for brainstorming tasks such as naming, positioning, side projects, campaign ideas, research angles, and talk topics. It aims to show the common answers models give to a particular prompt before helping a user explore alternatives. The project is published under the MIT license and provides both MCP tools and a command-line interface. The repository page lists zero stars and zero forks.
Its baseline tool asks configured models a plain question and returns a “cliché cloud,” with answers cached per prompt. diverge generates alternatives using SCAMPER, TRIZ, a morphological matrix, far analogies, or constraint flips. The tool can run all five techniques in parallel. It retains ideas with an embedding-based Cliché Score below 50 and a usefulness rating of at least 6; if too few qualify, it regenerates ideas using rejection reasons. Survivors are searched on the web, receive a web-prevalence component, and are sorted lower when the system flags them as an existing common product or practice.
remix combines two ideas into hybrids and identifies which components—such as mechanism, audience, channel, or constraint—came from each parent. score rates a list of ideas from 0 to 100, where lower scores mean greater distance from model consensus and from what already exists. Its parts include embedding distance, web prevalence rated 0–10 with sources, a model judge’s account of the closest cliché, and a separate usefulness critic rated 1–10. The judge explains a score but is not blended into the final formula, because the project says it did not improve held-out accuracy. With web checking disabled, the score uses embedding distance alone. explain offers a reason for a score and a concrete rewrite intended to push the idea further.
The repository says the score combines embedding distance (60%) and web prevalence (40%) when a web check is available. Web checks use Claude web search when an Anthropic API key is configured, or Brave Search with a Brave API key; the README estimates Claude web search at about $0.01 per idea. Gemini embeddings are available with a Gemini key. Alternatively, an optional Transformers.js installation can run the multilingual E5 Small model locally; the README describes it as a 118 MB download, running on CPU without a key or quota. Without either embedding option, the tool falls back to a less semantic offline lexical mode.
Setup uses npx and supports Anthropic and Gemini model keys; one key is enough, while the README says two model families produce a better cloud. The project reports calibrating the score on 270 labelled ideas across 20 prompts: 70 handwritten anchors and 200 generated ideas rated blindly by three LLM judges. In leave-one-prompt-out cross-validation, it says a more clichéd idea ranked above a less clichéd one in 86% of pairs, versus 82% using embeddings alone; reported agreement with judges’ originality ratings is Spearman 0.73. Human ratings are listed on the roadmap.
What the archive adds
We have not covered this before; it opens a new thread.
Trend signal
This does not yet fit a pattern New Runtime tracks. Not Obvious applies a score and generation workflow to creative ideation, but the source does not establish a broader shift in agent architecture, interfaces, or economics.
Why it matters
The tool makes a useful distinction between producing more ideas and producing ideas that are less similar to familiar model outputs. For teams using an assistant in naming, positioning, or early-stage concept work, a baseline cloud can make the model’s default suggestions visible instead of leaving users to mistake fluency for originality. The diverge and explain steps then offer a structured way to challenge those defaults.
That workflow may also help teams make creative review more specific: rather than asking whether an answer “feels original,” they can inspect the nearest cliché, compare alternatives, and decide whether a direction deserves human research. But the score is a proxy, not a verdict. Its value depends on the prompt, the configured models, embedding behavior, and the web check. Teams should treat it as a brainstorming aid and verify promising ideas independently, especially before making claims about novelty or market whitespace.
Limits and uncertainty
The reported calibration uses LLM judges, not human ratings; the repository lists human originality ratings as future work. The figures are the project’s own results, and the README does not provide independent validation. A high distance score does not establish that an idea is useful, original to people, or commercially available nowhere; the tool’s separate usefulness critic and web-prevalence search address parts of that problem, not all of it. Search coverage and score behavior may vary with configuration. The repository’s zero stars and forks indicate little visible GitHub adoption at the time represented by the page, not evidence about the quality of results in use.
Skill to build now
Try the tool on one real ideation task this week: write a specific prompt, run baseline, and save the cliché cloud before asking diverge for alternatives. Keep the same personal context across tools, as the README recommends, so the comparison uses one cloud. Choose one alternative, run score and explain, then manually check whether its purportedly unusual mechanism or audience already appears in your field. Record which suggestions a teammate finds useful; that human check is precisely what the project’s current calibration does not yet establish.
Reader takeaway
Not Obvious turns “give me a more original answer” into a small inspectable workflow: show the common model response, generate away from it, and score the distance. The score can help expose cliché, but it cannot certify creativity. Use it to widen a team’s options, then rely on human judgment and independent research before treating an idea as genuinely new.