Practical guides for AI product builders
Battle-tested methodology on multi-video research, CLAUDE.md structure, validation, and the research-to-shipped-SaaS workflow. No fluff.
Is there a tool that turns YouTube tutorials into a product roadmap?
Summarizing ten tutorials gives you ten summaries. A roadmap only appears when you read them against each other and let the disagreements decide the order.
The spreadsheet is the competitor: spotting manual workarounds in creator content
Nobody makes a tutorial about their workaround. They perform it mid-sentence, apologise for it, and cut away — which is exactly why it is the part worth watching.
Setting your product benchmarks from what practitioners call good
A dashboard full of numbers with no threshold is a dashboard that cannot tell you anything. The threshold comes from the people already doing the work.
Designing your data model from the vocabulary practitioners already use
Every schema is a bet about which things are the same thing. The domain already decided — in the words people use when nobody is designing a database.
Should this feature use AI at all? Deciding from practitioner evidence
If five competent practitioners reach the same answer from the same input, you have found a rule, not a job for a model.
Designing your free trial from how practitioners actually evaluate tools
A fourteen-day trial for a monthly workflow is a trial nobody can finish. The right length is a fact about the job, and practitioners state it on camera.
Validating a two-sided marketplace when only one side makes videos
Freelancers make videos. The people who hire them do not. Build your corpus without noticing that and you will validate the abundant side of your own marketplace.
Building honest comparison pages from video research
A comparison that finds you better on every row is read as marketing and discounted entirely. The page earns its credibility on the row where you lose.
When your sources disagree: resolving contradictions in video research
The instinct when two sources disagree is to pick a side or split the difference. Both destroy the signal. A contradiction is usually a boundary condition wearing a disguise.
How to research a market in a language you do not speak
The best-served markets are the ones every founder can read. The under-served ones are frequently just one language away — and the barrier is lower than it was two years ago.
Using video research to decide a pivot
Most pivots are researched backwards: the destination is chosen in a room, then evidence is gathered to support it. The corpus always agrees, and it never predicts anything.
Workflow mapping from day-in-the-life and over-the-shoulder videos
Interviews give you what people notice about their work. Screen recordings give you what they actually do — including the ten minutes of copy-paste nobody would ever mention.
Reading vendor content without absorbing vendor bias
Official content is the best record of what a company claims and the worst record of what its customers experience. Most corpora mix the two and call the result a market view.
Name your product from the vocabulary in your research
Founders agonise over the brand name and choose the category noun in a sentence. It is the wrong way round — customers forgive strange names and never forgive not knowing what you are.
Is this demand B2B or B2C? Reading the buyer out of creator content
Founders usually decide B2B or B2C from the subject matter, which is the one piece of information that does not distinguish them. Who feels the pain and who controls the money does.
Estimating your support burden before you have customers
A validated problem with a support cost you never modelled is how a product with good retention still loses money on every customer it keeps.
How to search YouTube like a researcher, not like a viewer
Most research corpora are weak because of the searches that built them. Six query families, run to saturation, beat fifty scrolls through one.
Thin-niche research: what to do when your market has almost no videos
Almost no videos on your niche is common and rarely fatal. The signal has moved one layer up — into the tools the niche runs on, and the trade next door.
Build, buy, or integrate: deciding from practitioner content
The cheapest roadmap decision is the one that removes work. Practitioner content tells you which components your market treats as plumbing.
Keeping CLAUDE.md current as your research changes
Writing a good CLAUDE.md is the easy half. Keeping it true while the research underneath it moves is what decides whether the agent helps or confidently misbuilds.
Writing a demo script from video research
A feature tour asks the prospect to translate every screen into their own life. A researched demo starts inside it.
Citing creators properly when your research is built on their work
Research built on other people's public work carries two obligations: make every claim checkable, and never pass off their material as yours.
The post-launch research review: did your evidence hold up?
Teams review launches and almost never review the research that shaped them. The scorecard takes an afternoon and changes every pass afterwards.
Timing signals: is now the right moment to build this?
Most idea validation asks whether the problem is real. The harder question, and the one that kills more products, is whether it is real yet.
Designing onboarding from the moments people stall on camera
Onboarding usually gets designed from the inside out. The better draft comes from watching where people already get stuck, on camera, in someone else's product.
Scoping a v1 from build-along videos instead of a feature list
Feature lists describe one team's answer to a job. Build-along videos describe the job — which is a much better place to draw a v1 boundary.
Reading switching costs: why people stay with software they complain about
A loudly-hated incumbent looks like an opportunity. Whether it is one depends on a number nobody measures: what it costs to leave.
Writing your docs from the questions people actually ask on camera
Most docs are a description of the software. The useful version is an answer to the questions people were already asking before they met it.
Feature or product? Telling the difference before you build
Most ideas that fail this test fail it loudly and early, in public content, months before anyone writes code against them.
Finding your churn reasons before you have a single customer
Churn research normally starts after the cancellations do. Most of the reasons were legible in public content long before that.
Spotting the constraints that quietly kill a v1
The requirement that sinks a first release is rarely a missing feature. It is a rule the team never knew the market had to follow.
Where to launch: reading distribution signals out of creator content
Most launch plans are a list of places founders go. The useful version is a list of places your customers already go when they have the problem.
Turning video research into a customer interview script
The point of a research pass is not to skip interviews. It is to stop wasting the first five of them on questions the public record already answered.
Finding the objections your buyers already voice
You do not need customers to learn what will stop customers buying. People say it out loud, unprompted, every time a tool in your category gets recommended.
Defining your ICP from what a segment already says out loud
Job titles are a weak way to cut a market. How a group talks about the problem is a much stronger one, and it is sitting in public.
Choosing a stack from what practitioners say after year one
Every technology looks good in a twenty-minute demo on an empty project. The information you need is in the videos recorded eighteen months later.
Writing a market memo you can actually defend
Anyone can assert that a market is large and underserved. A memo where each claim can be opened at its source is a different document entirely.
Making research survive contact with the rest of your team
Research nobody else can verify becomes one person's opinion within a fortnight. The fix is structural, and it is mostly about what you hand over.
You cannot size a market from video. You can sanity-check one.
The number has to come from somewhere with a methodology. What public research can tell you is whether the world behaves as though the number were true.
Picking your first integrations by watching where the copying happens
Integration roadmaps usually get written from a list of popular products. The useful version gets written from watching what people copy and paste.
How to kill a SaaS idea early: the signals worth looking for
Most idea research is built to succeed, which is why so much of it does. Running the pass designed to kill the idea gets you to a real decision weeks earlier, and the ideas that survive it are worth considerably more.
Choosing between two SaaS ideas without guessing
The problem with researching one idea is that there is nothing to compare it to, so your existing preference quietly supplies the verdict. Running two in parallel to the same depth removes that comfort.
YouTube vs Reddit for product research: what each one actually shows you
Reddit is faster at telling you what annoys people. Video is far better at showing you why, because demonstrating a workflow takes an effort nobody spends on a problem they do not really have.
Turning research into landing page copy that sounds like the customer
Most landing page copy is written from the inside out — feature, benefit, adjective. Copy built from research runs the other way, starting with the sentence the buyer already uses, which is why it stops feeling like marketing.
A one-week research plan for a vertical you know nothing about
Outsiders usually enter a vertical at the opportunity question, which is the last one they are equipped to answer. Doing vocabulary and workflow first takes a week and changes every conversation that follows.
Prioritizing a backlog with evidence instead of opinion
Every prioritization framework has two inputs filled in from memory, and those two inputs decide the ranking. Counting independent sources instead is unglamorous and changes the order of the list more than any scoring formula does.
Conference talks and podcasts as research sources: what each one is good for
Most research corpora are built from tutorials, which means they describe problems that have already been solved well enough to teach. Talks and podcasts sit earlier in that cycle, and the difference is roughly a year.
What AI video research gets wrong, and how to catch it
Automated research fails in specific, predictable ways, and none of them look like errors. They look like fluent, confident summaries — which is exactly what makes them worth learning to spot.
A SaaS idea validation checklist that fails ideas early
Most validation is confirmation-shaped: look for evidence the idea works, find some, start building. This checklist inverts it — nine checks ordered so the cheapest ones have the best chance of killing the idea first.
General AI summarizers vs a research system: where the gap actually is
Summarizing and synthesizing are different jobs. General assistants do the first one well and have no mechanism for the second, which is why a folder of twenty excellent summaries still leaves the week of work undone.
Tutorial videos are a workaround archive — mine them for micro-SaaS ideas
A tutorial only exists because a product does not do the thing. That makes how-to content the densest available record of manual workarounds — and the friction people apologise for on camera is the shortlist.
Hype vs signal: reading creator content without getting fooled
A corpus where every source agrees enthusiastically is usually a narrative propagating, not a problem recurring. Here are the checks that tell the two apart before you build against the wrong one.
How many videos does a research corpus actually need?
Picking a round number before you start is the wrong move in both directions. Corpus size should be decided by when new sources stop producing new claims — a threshold you can actually watch happen.
Pricing research from public content: what people already spend
You cannot read a price point off public discussion, but you can read the two things that determine it — what buyers already spend, and what they mentally compare you to. Both are stated freely and neither shows up in a survey.
Client market research without the two-week desk-research bill
Discovery phases get compressed because they are hard to bill. Working from what a client's customers already say in public produces a sourced findings document in days, which makes the phase worth charging for again.
Standing research: monitor a niche instead of researching it once
Research is usually treated as a project with an end date, which is why findings quietly expire without anyone noticing. Standing research swaps the output from what is true to what changed — a far more useful thing to receive.
How indie hackers read YouTube demand signals
Validation advice usually optimizes for whether a problem is real. For an indie hacker the harder question is whether it is small enough to ship alone and whose audience you can actually reach — and YouTube answers both.
Finding trending SaaS opportunities in creator content
A topic trending is not an opportunity. The gap between them is whether the audience is still asking questions the content did not answer — and that is a different measurement, needing different tooling.
Researching a niche by analyzing YouTube channels at scale
Analyzing channels at scale fails in a predictable way: the loudest channel becomes the consensus. Sampling across sizes, capping per-channel weight, and mapping contradictions is what makes the output a market read rather than an echo.
Tracking which SaaS tools YouTubers actually recommend
The tools mentioned most are usually the tools paying the most. Filtering for unpaid, mid-workflow mentions — and recording why the creator uses them — turns a vanity count into competitive intelligence.
Tearing down a competitor using their own YouTube footprint
A competitor's own channel shows the happy path. Third-party tutorials show the friction, and the comments under them name the objections and the products people switched to — unsolicited and specific.
What YouTube research actually costs
Tooling costs are trivially small next to the time cost of manual research. A twenty-video corpus is roughly a working week by hand — worth pricing honestly before deciding which side of the build-versus-buy line you are on.
Turning research notes into a spec a coding agent can build
The gap between good research and working software is a compilation step. Findings become constraints, constraints become a context file, and the non-goals section does more work than any feature list.
Build in public with a research trail, not just a changelog
Most build-in-public content reports what shipped. A research trail reports why — sourced, specific, and impossible to copy — which is both better content and a better filter for the audience you want.
How to extract SaaS product ideas from YouTube content automatically
Feed 15-25 long-form videos on one topic into a structured extraction pass, then synthesize across all of them. The buildable idea is the manual workflow step that repeats across creators — here is how to surface it automatically instead of by hand.
How to validate a startup idea using YouTube search and view data
Views per subscriber, view concentration, and comment-to-view rate turn YouTube's public numbers into a demand signal you can rank topics with — and reveal exactly where quantitative validation runs out and behavioural evidence has to take over.
How to find underserved niches by analyzing YouTube content gaps
Content gaps show up in three shapes: unanswered recurring questions, stale videos ranking on live queries, and workflows that exist only as fragments. Each maps onto a different kind of product opportunity — here is how to find and score them.
How to mine YouTube comments for product pain points and feature ideas
Most YouTube comments are reactions to the video and worth nothing. A small fraction describe what the commenter actually tried and where it failed — those are unsolicited bug reports on a workflow. Here is how to isolate, cluster, and rank them.
The AI video summarizer test: does it keep timestamps and citations?
Most AI summarizers flatten the transcript before summarizing, which destroys time codes in step one — nothing downstream can restore them. Here is the test for tools that keep timestamps and citations, and why it matters more than summary quality.
YouTube as a feature-discovery channel for product managers
Creator videos are unmoderated usability sessions with no observer effect. Extract the workflow each one performs, count the deviations from the intended path, and you have a feature backlog ranked by how often real users go off-script.
YouTube research tools vs manual note-taking: what actually replaces the notebook
Manual note-taking is not slow because typing is slow. It is slow because human working memory cannot hold twenty videos at once. Here is what each alternative actually replaces, and where the notebook still beats the software.
Turn your YouTube watch history into a searchable knowledge base
Watch history is a record of research you already did and cannot retrieve. Export it, filter aggressively, group into topic corpora, and extract structured notes — and hundreds of hours of watching become something you can question.
Turn YouTube videos into a SaaS in 7 days (a builder's playbook)
Seven days isn't a marketing number — it's the actual cadence we've watched indie hackers run when the research-to-build handoff is clean. Here's the day-by-day breakdown.
How to write a CLAUDE.md file (with examples Claude Code actually respects)
Claude Code reads CLAUDE.md every conversation. Most files developers write are shopping lists. Here's the structure that turns it into a binding contract — what to build, what to refuse, and what specific failures to plan for.
How to use Claude AI for product research (the methodology that actually works)
Generic AI summarizers lose at product research because they collapse contradictions instead of surfacing them. Here's how to use Claude differently — to find the wedge, the failure modes, and the exact MVP scope.
How to validate a SaaS idea without surveys
Surveys ask people what they would do; SaaS validation needs to know what they actually do. The watch-what-they-do framework looks at five behavioral signal sources and tells you whether a problem is real, painful, and pay-worthy.
Best YouTube research tools for builders in 2026
There are roughly three jobs people lump together when they say 'YouTube research tool': transcript extraction, single-video summary, and multi-video synthesis. Different tools win at different jobs. Here's the honest split.