Skill + template
Stress-test your idea before you build it
A step-by-step method for desk research, competitive benchmarking and honest validation. It won't tell you what you want to hear. Inside is the real research I did for a Refertrack feature.
- Agent Skills format
- 7 files, all in English
- Needs web search for sources
Why I made it
I work alone. Nobody tells me "this idea is weak"
I build iOS apps on my own: no colleague to push back in a meeting, no research team. The real risk isn't ugly code. It's spending weeks on a feature the market already solved better, or that nobody feels as a problem.
When I considered adding vaccine and booster tracking to Refertrack, I did the research with this method before opening Xcode. Two things I wouldn't have found by writing code:
- Booster schedules aren't the same in every country. Both the interval and the per-vaccine exceptions change. A fixed value in the code would have been wrong by design.
- The value is knowing what's due, not archiving. The apps praised in reviews get credit for "what's overdue, what's due soon". History comes second.
I used the same method for a competitive benchmark on Quillino, my rental management app. Then I cleaned the file up, made it general, and put it here.
How it works
From download to a sourced report
Install it
Unzip it. In Claude Code, Cursor, Copilot, Codex or Gemini CLI, install it the way that tool loads skills (folder name has to stay
product-research-validation). In Claude, Codex or ChatGPT/Gemini, pasteSKILL.mdinto a custom GPT or Project.Ask it to validate your idea
Try something like: “I want to validate an app idea: tracking vaccine and booster due dates. Walk me through the research method.” It asks a few questions, then starts sourcing data.
Get a ranked, sourced answer
Every claim comes with a link and a reliability score. It ends in a scored opportunity ranking and a report you can act on, not a vague “sounds good.”
What's inside
What's in the zip
- SKILL.md
The method your AI follows. Brief, market data, competitive benchmark across three kinds of competitor, voice of the user, a scored opportunity ranking, findings, and insights checked against five criteria. Full mode or a one-page quick scan. Every claim gets a source and a reliability score, or it's labeled as an assumption.
- references/worked-example.md
The real output on the vaccine feature: sourced market data with reliability scores, competitors' strengths and weaknesses, user data across five language markets, eight findings and six insights.
- references/report-template.md
The structure of the final report, so every run ends in the same usable document: verdict, ranking, insights, findings, sources.
- references/test-with-real-data.md
Bonus. How to validate an OCR or parsing engine with real, anonymized data instead of clean made-up examples. It comes from Refertrack's lab report parser.
- references/privacy-policy-on-device.md
Bonus. A starting privacy policy for apps that process sensitive data only on device, with a checklist to run before publishing it.
- README.md + LICENSE.md
Install guide for skills-enabled tools, chat assistants and use by hand, plus the full license terms.
How to use each file
From download to results
SKILL.md
What it is: the instructions your AI follows.
With Claude, Claude Code, Codex, Cursor, Copilot or Gemini CLI: install the folder, then ask something like “I want to validate an app idea for [idea]”. If the skill doesn't start on its own, ask for it by name: “use the product-research-validation skill”.
With ChatGPT or Gemini: paste the whole file into a custom GPT, Project or Gem, then make the same request.
By hand: work through Step 0 to Step 6 in a document, using the rules at the top as your checklist.
Tips: say “quick mode” for a one-page first answer; save the “Research state” block at the end of each step to resume later; turn on web search first.
“I want to validate an app idea: tracking vaccine and booster due dates. Walk me through the research method.”
Adult ten-year booster coverage in one EU
market studied: 10.6% nationally; no region
at 95% or above.
Grounded. High: peer-reviewed study in
Vaccines (MDPI, 2022), primary data from
regional health services.references/worked-example.md
What it is: a complete, real run of the method on a feature of a published iOS app.
How to use it: read it before your first run to see how deep each step goes. Don't reuse its numbers: use it as the standard to compare your own output against. The skill also tells your AI to read it to calibrate.
Reviews, forum threads and market data collected in the earlier steps, on why adult vaccination records get lost.
F1: The gap between child and adult
coverage is a memory problem, not a
willingness problem.
As children a system keeps count (school
requirements); as adults, nobody does.
Grounded.references/report-template.md
What it is: the structure of the final report.
With AI: used automatically in the last step. Ask “write the final report as a Markdown file following the template” to get it as a file.
By hand: copy it into your doc and fill each section as you finish the matching step.
Every finding and insight from Steps 1 to 6, still scattered across your notes.
## Verdict
[3 to 5 sentences. Say plainly whether the
original idea is the best option, and why.]
## Opportunity ranking
| Opportunity | Market | ... | Total |
| Original idea | 1-5 | ... | /25 |references/test-with-real-data.md
What it is: bonus. A pattern for testing OCR, receipt, PDF or form parsers with real, anonymized data.
How to use it: when your product extracts data from real documents. With a coding agent: “set up parameterized tests for my parser following references/test-with-real-data.md”. By hand: follow the four steps and the checklist.
A real photographed document, anonymized, transcribed exactly as OCR would output it, noise and all: headers, disclaimers, lines to ignore.
Field not recognized → alias missing
from the dictionary. Add it, plus an
explicit test case. Never fix it silently.
Value extracted wrong → parser has an
unhandled case. Extend the sub-parser,
keep old fixtures green.references/privacy-policy-on-device.md
What it is: bonus. A plain-language privacy policy template for apps that keep sensitive data on the device.
How to use it: replace every [square bracket], remove what doesn't apply, run the checklist, and have a lawyer review it. With AI: “adapt this template to my app, which [what data, where it goes]”, then check every claim against your real implementation.
“Adapt this template: my app photographs blood test reports and stores them only on device, synced through the user's own iCloud.”
[What the user photographs] stays on the
user's device. The only copy outside the
device is the backup in the user's
personal cloud. The app has no server of
its own and never receives this data.README.md and LICENSE.md
README.md: installation for every kind of tool, example requests and this per-file guide. Start here.
LICENSE.md: use and adapt the files for your own projects, commercial ones included; don't resell or redistribute them.
A taste
The rule everything starts from
The job is not to validate the user's idea, it is to stress-test it.
SKILL.md
Adult coverage of the ten-year booster in one EU market studied: 10.6%, against 92.8% average pediatric coverage in the EU/EEA. Nobody keeps count.
references/worked-example.md, section 1.2
Is it for you
Honest fit
Yes, if
- you build alone or in a small team
- you want to decide what's next with data, not gut feeling
- you'd rather hear "no" now than after launch
Not if
- you want quick confirmation of the idea you already have
- you need research with real users (this complements it, it doesn't replace it)
- you need legal, medical or tax advice
The worked example is real
What I removed, what I kept
None of the competitors I found sell a method backed by a real case you can check. Usually the example is invented to make the method look brilliant.
Here it's the research I used for a decision on Refertrack, which is on the App Store. I removed competitor names and private references. Sources, numbers and links are the originals, dated September 2026: check they're still current before reusing them.
License
- You can use and adapt the files for yourself and your own projects, commercial ones included.
- You can't redistribute, resell or republish them, in full or in substantial parts, or present them as your own product.
A Markdown file can't be protected technically, so I'm trusting you. Full license
Questions
Can I get a refund?
No. The packs are digital products you download instantly, so there are no refunds. If a file is missing, broken or doesn't match the description, write to quillino.support@gmail.com and it will be fixed.
What if my AI can't browse the web?
The skill says so at the start instead of inventing sources. It works with the sources you paste in, labels everything else as an assumption, and picks up the sourced steps when browsing is available. There's also a quick mode for a first one-page answer.
Can I stop and continue later?
Yes. Every step ends with a short research state summary: paste it into a new conversation and the wizard resumes from there.
Does the skill find sources on its own?
If your AI has web search turned on, yes. Without it, it can't find sources. The method requires direct links, no invented sources, a reliability score for each, and an explicit label when a claim is only an assumption. You still check the sources: findings only start after your review.
Is it only for iOS apps?
No. It works for an app, a feature or a small business. The example is an iOS app because that's what I build.
Which AI do I need?
One with web search, for this pack. It loads as a real skill in Claude, Codex, Cursor, Copilot, Gemini CLI and other tools that support SKILL.md. In ChatGPT or Gemini you paste it as instructions. Without AI, it's a framework you follow by hand.