Communitygithub.com

prfaq-smoke-test

Shows an agent how to art-direct and build "prfaq smoke test" so the concrete output is a landing page, homepage, or portfolio. The procedure moves through PR/FAQ Smoke Test, Steps, 1. Pin the bet. Use it when a new assignment calls for this craft.

Qu'est-ce que prfaq-smoke-test ?

prfaq-smoke-test is a Claude Code agent skill that shows an agent how to art-direct and build "prfaq smoke test" so the concrete output is a landing page, homepage, or portfolio. The procedure moves through PR/FAQ Smoke Test, Steps, 1. Pin the bet. Use it when a new assignment calls for this craft.

Compatible avec~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/robfold/prfaq-smoke-test/tree/HEAD/skills/prfaq-smoke-test

Demander à votre IA préférée

Ouvre une nouvelle conversation avec cette compétence d'agent déjà préchargée.

Documentation

PR/FAQ Smoke Test

Find out whether a product deserves to be built, and which reason to buy it wins, before writing product code. One chain:

PR/FAQ → three angles → three pages → small equal ads → compare → build the winner, or nothing.

The PR/FAQ is the cheap filter: if the press release cannot make the target customer care, stop there. The smoke test is the behavioral filter: three pages that differ only in the message, and real visitors deciding which message they act on.

Work done on the user's machine is free to do. Three steps reach the outside world: publishing pages, collecting emails, and spending ad money. Each one waits for the user's explicit go-ahead in this task.

Steps

1. Pin the bet

Collect the customer, the problem, the product idea, the price (or [TBD]), what the test will decide (build, change, or drop), and the money and days the user will spend. Read the conversation, files, and repo first. Ask the user only what they alone know, one question at a time.

Done when: one line reads "We are deciding whether to build [product] for [customer]. The test spends at most [budget] over [days]." Unknown values stay [TBD].

2. Write the PR/FAQ

Read references/prfaq.md and fill its template. The press release is dated on the imagined launch day and fits on less than one page. The FAQ has an external part (what customers ask) and an internal part (cost, risk, economics, the riskiest assumption).

Then judge the release as the target customer would. Name its weakest line. If the headline and first paragraph would not make that customer want the product, say so plainly, and stop or rewrite before going further.

Done when: the PR/FAQ is saved as prfaq.md, your verdict is written under it, and either the verdict is "compelling" or the user chose to continue anyway.

3. Pull three angles

An angle is one reason a customer would pick this product. Draw each from a different line of the PR/FAQ. Common sources: the outcome they get, the pain that goes away, the mechanism that makes it believable, speed or price, who it is for.

Three angles are three different reasons to buy. The test: a real customer could prefer A and shrug at B. Three phrasings of one reason fail that test, so replace one.

For each angle write: headline (10 words or fewer), subheadline, three supporting points, CTA label, the PR/FAQ line it came from, and the assumption it tests.

Done when: angles.md holds a three-row table, each row traced to a PR/FAQ line, and every pair passes the prefer-A-shrug-at-B test.

4. Build three pages

Read references/landing-pages.md. Fill the product JSON block of assets/landing.html once, then copy the file as a.html, b.html, c.html and edit only the angle JSON block in each. The product block, layout, price, form, and footer stay byte-identical, so the message is the only variable.

The CTA opens a fake door: an honest panel saying the product is not available yet, with an optional waitlist email. It never takes payment.

Then run python3 scripts/check_pages.py a.html b.html c.html. It confirms the files differ only inside the angle block, then opens each page headless, clicks the CTA, submits a test email, and reads the events. It intercepts every POST, so no test event reaches the user's storage.

Done when: the script prints structure: PASS and events: PASS. If it prints events: UNVERIFIED because Playwright or its browser cannot be installed, click through each page by hand as the reference describes and write in test-plan.md that the event check was manual. Step 5 starts after two PASS lines or a recorded manual check.

5. Fix the rules before launch

Write test-plan.md before any traffic: primary metric, traffic source, budget per variant, run length, minimum visitors per variant, the bar a winner must clear, and stop conditions. Size the test with python3 scripts/compare.py plan (see references/reading-results.md). If the budget cannot buy enough visitors to detect the difference the user cares about, say so now, while the money is unspent.

Done when: every field in test-plan.md is filled or marked [TBD], and the user has approved the launch.

6. Launch

Deploy the three pages on the user's static host, one URL per variant. Give each variant its own ad set with the same audience, budget, schedule, and format, and let each ad repeat its page's headline (message match). Run all three at the same time. Before spending, confirm the privacy notice, the ad platform's policy on not-yet-available offers, and any rules for regulated categories such as health or finance.

Done when: three live URLs load, events from each arrive where the user stores them, and the ads are approved.

7. Read the result and decide

Export per-variant counts to a CSV and run python3 scripts/compare.py results.csv with the minimums and bar from test-plan.md. Report each variant's rate with its interval, cost per signup, the verdict line, and any note: line under it. A note means the winner's interval reaches below the bar, so the build decision is provisional. A variant under its minimum visitors makes the whole test UNDECIDED, and the test plan's rules decide whether to extend it.

Write the decision into prfaq.md: build with angle X, change the offer, drop it, or undecided. If building, the winning angle becomes the PR headline. Keep the losing angles and their numbers in the record.

Done when: prfaq.md has a dated decision with the counts it rests on, and the verdict matches what compare.py printed.

Honesty

  • A fake door measures interest. To learn whether people will pay, show the price before the waitlist form, as Buffer did with a pricing page. Taking real money is a pre-sale, a different test with its own delivery and refund promises.
  • The customer quote in the press release is imagined and marked so. Pages and ads carry no testimonials or user counts that have not happened.
  • Emails go only to the user's own store, with a one-line promise of how they will be used. Analytics events carry the variant and event name, never the email.
  • Small samples mislead. Report the intervals, name the winner only when the pre-set rule names it, and keep the losers' numbers.

Credits

  • PR/FAQ: Colin Bryar and Bill Carr, Working Backwards (2021), on how Amazon writes a press release and FAQ before building. Amazon's excerpt.
  • Landing page before product: Joel Gascoigne, Idea to Paying Customers in 7 Weeks (Buffer).
  • Further reading: David J. Bland and Alexander Osterwalder, Testing Business Ideas (2019); Ash Maurya, Running Lean, 3rd ed. (2022); Alexander Osterwalder et al., Value Proposition Design (2014); Tim Ash et al., Landing Page Optimization, 2nd ed. (2012).

The text is new. The three-angle chain, the byte-identical page rule, and the comparison script are this skill's own combination, not an official method from these authors.

Skills associés