The Signal-to-Artifact pattern

Before running a full idea evaluation, I use a five-question pre-filter for a specific category of product ideas. Not all ideas fit this pattern - but the ones that do tend to share a particular structure that’s worth checking for before investing hours in deeper analysis.

The pattern: raw public signal that nobody has synthesized → LLM processing → shareable artifact → replaces a specific manual workflow.

All four elements have to be present. Remove any one and the idea degrades significantly.

The five questions

Run these in order. Fail any one, kill the idea immediately. Don’t continue to the next question.

Q1 - Where is the ignored signal?

Name a specific corpus of public data that people know exists but no product makes actionable. G2 reviews, GitHub commit history, app store reviews, patent filings, job postings - name it specifically. Generic answers (“there’s a lot of data online”) fail. The more specific and obscure the corpus, the better.

Q2 - What manual workflow does this replace?

Name the workflow, name who does it, name how long it takes. “SaaS PMs spend 2-4 hours reading competitor reviews manually before quarterly planning” passes. “Companies want competitive insights” fails. If you can’t name a specific person doing a specific thing on a specific cadence - kill it.

Q3 - Is the output a shareable artifact?

Can the output be a URL, PDF, or document that someone sends to another person without requiring them to log in? Binary. If the value only exists inside a dashboard, fail. No exceptions. This kills roughly 70% of SaaS ideas immediately and for good reason: value locked behind a login can’t be shared, can’t be the demo, and requires a separate distribution strategy you haven’t built.

Q4 - Does the artifact explain itself?

Show the output to someone unfamiliar with the product. Do they immediately understand what it does AND want it? If you have to explain the product before showing the output, fail. The output should sell the product better than you can.

Q5 - Who has the trigger and where do they spend money today?

Name the person, name the moment of acute need, name what they currently pay for something adjacent. Must be specific. “Enterprise teams wanting insights” fails. “SaaS product marketers who run competitive analysis sprints before quarterly planning and currently pay $299/mo for Klue” passes.

---

Only ideas that clear all five proceed to the full 17-point evaluation.

The critical refinement

The original version of this pattern had four elements. The fifth - and most important - was added after testing real ideas against it:

The raw signal must come from a source that general-purpose AI research tools cannot access or index at scale.

If your signal source is publicly indexable web content - G2 reviews, Reddit threads, Trustpilot, blog posts, news articles - Claude Deep Research or Perplexity can approximate your synthesis for free. Your product becomes a formatting layer, not an intelligence layer.

Sources that pass: Google Maps reviews (behind scraping infrastructure), app store reviews (requires dedicated scraping at scale), GitHub PR and commit activity (structured data not indexed by search), Glassdoor and Indeed reviews (behind auth walls), job posting patterns (requires systematic collection), proprietary operational data.

Sources that fail: G2, Capterra, Trustpilot, Reddit, HackerNews, ProductHunt, blog posts, news articles. All publicly indexable, all accessible to Deep Research. Building on these is building on sand.

The source access is the moat. Not the synthesis quality.

What this pattern rules out

• Products requiring users to input their own data first (creates cold start and behavior change problems)
• Dashboard-only products where value is locked behind login
• Products competing on data volume or breadth (you can’t out-crawl a well-funded competitor)
• Products with unclear buyers or diffuse trigger events
• Products that require network effects before delivering any value

Why this works as a pre-filter

Running a full 17-point evaluation takes two to three hours done honestly. This takes ten minutes. The five questions above kill the ideas that would have wasted those hours - ideas with no specific signal source, no named manual workflow, no shareable output, no obvious buyer.

The ideas that clear all five tend to have a specific structural advantage: they don’t need to invent a distribution strategy because the output IS the distribution. Every artifact is a content piece that reaches the exact buyer. The product markets itself by existing.

That’s a rare property and worth filtering for specifically.