idea-validation-agents: A Skill Collection That Makes Claude Code, Codex or Cursor Interview You Before You Build

The problem is not idea shortage, it is unstructured optimism

Most developers do not lack ideas. They lack a forcing function that makes them write down why an idea will fail before they spend three months on it. The README frames the target user narrowly: indie app developers, solo founders exploring B2C concepts, side-project builders, and people with no idea at all who want one worth building. The stated promise is validation in about ten minutes rather than regret in six months. That is a workflow claim, not a data claim, and it matters which one you are buying. What the repository actually provides is a set of prompts that make an AI agent ask you questions in a fixed order, score the answers, and write the result to disk. The value is in the sequence and the persistence. If you have ever asked a chatbot to critique your idea and received a cheerful list of strengths, the difference here is that the process is written down in files you can read, edit and re-run.

Four workflows, one routing file per tool

The repository exposes four workflows. Idea Generation interviews you about background, skills and interests, then produces seven to ten scored ideas. Idea Validation runs a nine-step pass ending in a decision memo. Market Deep Dive researches a category before you commit to an idea, and the README is truncated mid-sentence at the fourth workflow, so anything beyond the market saturation rubric cannot be confirmed from the supplied material. Routing is handled by tool-specific instruction files: CLAUDE.md for Claude Code, AGENTS.md for Codex, and rules under .cursor/rules/ for Cursor. The README says the agent detects intent and routes you automatically, and that all analysis results are saved to a memory/ folder so nothing is lost between sessions. That is the whole architecture. There is no server, no index, no database. The skills are markdown, the state is files, and the intelligence is whatever model your editor is running. Treat the memory folder as the real product surface: it is where you will go back and read what the agent concluded three weeks ago, and it is the only part of the system you can version-control alongside your code.

The scoring stack, and where it gets its inputs

Idea Validation is the most specified workflow. The README lists nine steps: trend analysis, competitor mapping, desire scoring, pricing model, distribution analysis, retention prediction, CAC modeling, a final score from zero to one hundred, and a decision memo. Two mechanics deserve attention. First, the final score uses what the README calls a multiplicative-floor algorithm, described as one catastrophic weakness killing the overall score, which is a deliberate departure from additive rubrics where strong dimensions paper over a fatal one. Second, pricing is estimated through Van Westendorp price sensitivity analysis combined with desire-premium multipliers, with the README giving survival and status desires as examples that command roughly 1.3 to 2 times price premium. Distribution modelling covers viral coefficient across six loop types, App Store optimization scored on a five-factor rubric, and creator economy fit. Competitor work is described as systematic App Store search plus mining of one-star and three-star reviews to find gaps incumbents leave open. Idea Generation pulls trend signals from TikTok Creative Center hashtag velocity, Reddit community pain language, X/Twitter conversations, App Store new entrants and Google Trends. All of that assumes your agent has working web access. The README does not discuss what happens when it does not, and that is the largest unaddressed failure mode in the documentation.

Getting it running is genuinely three commands

For Claude Code the README gives one command, git clone https://github.com/MaxKmet/idea-validation-agents.git, then opening the project folder and starting a chat. For Codex CLI it adds npm install -g @openai/codex, then cd idea-validation-agents followed by codex. For Cursor it is the clone, opening the folder, and the AI chat via Cmd+L or Ctrl+L. The README states there are no API keys, no dependencies and no terminal commands beyond the clone and the optional Codex install. The example prompts are concrete: "I want to build a habit tracker for climbers. Is it worth it?" for validation, and "I don't have an app idea yet. Help me find one." for generation. Two escape hatches exist in the generation workflow: saying "browse topics" presents twenty product domains to pick from, and saying "skip" jumps past the interview straight to ideas. That skip option is worth understanding before you use it, because the interview is what filters ideas against your domain expertise, skills and distribution advantages. Skipping the interview removes the personalization and leaves you with generic trend-chasing output.

The verdict is only as good as the model behind it

Nothing in the repository validates the underlying data. The trend signals, competitor counts and pricing estimates are produced by the model calling web tools during the session. If the model cannot browse, or browses badly, the report still arrives with the same confident structure and the same zero-to-one-hundred score. There is no confidence field, no citation requirement and no flag for thin evidence described in the supplied material. The README also presents several named frameworks, Van Westendorp, Klein's 2007 pre-mortem, the Riskiest Assumption Test, without explaining how faithfully the agent implements them. Van Westendorp in particular is a survey technique with specific question wording and a defined respondent count; what a chat agent produces is an approximation of the price curve, not a survey result. That gap does not make the output useless, but it does mean you should read the decision memo as a structured opinion rather than a measurement. The ten-minute claim is plausible for the conversation. Whether the research behind it is ten minutes of real retrieval or ten minutes of plausible-sounding recall is something the repository does not let you verify from the outside.

When a spreadsheet beats an agent

If your question is how many people search for a keyword and what it costs to reach them, a keyword tool and an ad platform's own estimator answer it with numbers you can defend to a co-founder. Ahrefs, Semrush or the Google Ads keyword planner return volumes and cost-per-click ranges drawn from their own indexes; the agent returns a CAC model built from whatever it retrieved. The two are not substitutes. Use the agent to decide which questions to ask and in what order, then answer the quantitative ones with tools that hold real data. The same logic applies to competitor review mining: the README describes mining one-star and three-star reviews, and you can do that directly in the App Store or with a scraping script, getting the actual review text rather than a summary of it. Where this repository has no obvious substitute is the interview structure. No spreadsheet asks you about your distribution advantages before it scores an idea. That is the part worth keeping.

Maintenance cost and the MIT licence

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That is a permissive licence and it does not impose copyleft obligations on your own code. It also means the authors offer no warranty, and the supplied material shows no releases, so there is no versioned artifact to pin. Your upgrade path is git pull against main. That has a direct consequence: if you edit the skill files to match your own scoring weights, a pull can conflict with your changes, and there is no changelog described in the material to tell you what moved. Forking is the safer pattern if you plan to customize, since it makes your divergence explicit. The ongoing cost is not compute, it is the tokens your agent burns on each run and the attention required to keep your memory/ folder from filling with stale verdicts you no longer believe. Nothing in the repository suggests an expiry or cleanup mechanism for those files.

Editorial conclusion

Adopt it if you already pay for Claude Code, Codex or Cursor and want a structured interview before you commit weekends to a build. Do not adopt it if you need auditable market data, because every figure in the decision memo is model-generated from web tool calls rather than pulled from a paid dataset. Before trusting a verdict, open the skill file that produced it, run the same idea twice and compare the scores, and check whether your agent actually has live web access, since without it the trend analysis degrades into recall.

Official sources

1. Issues
2. License: MIT
3. MaxKmet/idea-validation-agents on GitHub
4. README