Olexan-art/page-text-to-prompts

Version 1.008.

Page text to prompts

Purpose

The scripts read 1 web page and cut its copy into candidate blocks. You write a short judgement file; the build turns it into a prompt and a measured example per block. Text only: 1 public page, desktop, no login, no captcha bypass, no text in images.

Requirements

• Node 22 or later. No install, no API key. Run the scripts in dist/ from this skill's folder.
• Network access from the shell.

Rules

• R1. Do not ask the user anything first. A URL is enough.
• R2. Own words. No line you write may repeat 6 consecutive words of the page, case and punctuation ignored (overlap). This covers every field and every example.
• R3. Do not name the site's brand: domain name, og:site_name (brand). Do not use any name from the page.
• R4. Write analysis.md and every input file with the file-writing tool, never a shell heredoc. If the tool refuses (some harnesses refuse .md files from a subagent), write the file under another name and copy it with cp.
• R5. Never re-create the scripts' work by hand. Do not count words or characters. Do not write prompts.
• R6. Never invent facts. Never write Russian text. The scripts call no model.

Procedure

Every path below may point anywhere writable; work/ is the default, not a rule.

1. Read (script). node dist/read.mjs <url> --out work/page.json
   • Options: --no-render (source HTML only), --lang <code> (draft language).
   • Result: in work/, brief.md (read this), analysis.md (the draft, 1 header per candidate with a guessed type) and page.json (for the build; do not open it). If analysis.md exists, the draft goes to analysis.draft.md.
   • brief.md has 1 line per element: type, word count, status (js_only, collapsed, not_in_a11y, or none for ok), text, grouped by candidate bNN. [g3#2] is card 2 of group g3.
   • Time: 5 to 30 seconds. The render stops after 45 seconds and the read continues from the HTML.
   • On failure: REFUSED <code> (exit 2): stop and tell the user (see Refusal codes). If the scripts cannot run (no Node 22, no shell, no network), tell the user in 1 line and offer the saved-page route.
2. Fill in work/analysis.md (judgement). See the next section. Write in English, or in the language the user asked for, with its code in language:.
3. Build (script). node dist/build.mjs work/page.json work/analysis.md --html work/report.html --md work/page-prompts.md --json work/page-prompts.json
   • Result: the 3 files and checks: 5 of 5 passed. Examples off their range are warnings only.
   • On failure: REFUSED lists each problem with its check: overlap (R2), brand (R3), volume (no word range), variables (undeclared {name}), blocks (candidate missing, unknown or used twice), schema (missing field or value not allowed), or a line that cannot be read. Fix only the listed lines and run the build again.

analysis.md

Take word counts from brief.md and write ranges around them. Leave out character ranges.

language: en
example_product: a fictional booking app for small pilates studios
page_type: SaaS landing
audience: owners of small studios who take bookings by chat
awareness: comparing_solutions
tone: businesslike, warm; formality 2; you
promise: {product} gives {audience} {main_benefit} without {pain}.
logic: result and first step first, proof next, the steps before the prices
threads:
- the promise of no more calls returns in the hero, the steps and the closing call

## p01 hero | Hero | b01
role: say within five seconds what changes for the reader, and offer the first step
- eyebrow | 3-5 | the main change for the reader | No more booking calls
- h1 "Headline" | 7-11 | the result for {audience}, without the product's name | Clients reserve a reformer, and you stay with the class
- button | 2-4 | verb + object, the action {cta_action} | Set up bookings
- risk_line | 5-8 | {risk_remover} | Free to start, cancel at any time
techniques:
- Keep {product}'s name out of the headline; it first appears in the subheadline.
tone: businesslike, warm; you
avoid: "innovative"; a question as the headline

## p07 pricing | Pricing | b07
role: help the reader pick a plan and see the price fits the studio's size
- h2 | 4-8 | the price tied to the reader's size | Pay for the size you are
- 3 x plan "Plan":
  - h3 | 1-3 | the studio's size | Single room
  - price | 1-3 | {plan_n} | [price] / mo
  - items 3 | 2-6 | what is included, most important first | Self-service reservations / Client reminders / Card checkout
techniques:
- Name the plans after the studio's size, not after levels such as basic or pro.

Allowed values (the build refuses others):

• awareness: unaware, problem_aware, comparing_solutions, product_aware, ready_to_buy.
• Block type: hero, social_proof_logos, problem, solution, features, benefits, how_it_works, use_cases, testimonials, case_study, stats, comparison, pricing, faq, guarantee, about, team, cta, lead_form, newsletter, other. other needs a reason: line.
• formality: 1 digit, 1 to 5 (the report shows "N of 5"; not refused).
• language: an ISO code; uk gives Ukrainian prompts, any other code English prompts.
• page_type: free text.
• Variables: product, category, audience, language, pain, pain_n, main_benefit, benefit_n, feature_n, step_n, use_case_n, alternative, objection_n, cta_action; facts risk_remover, outcome_number, customer_count, customer_n, testimonial_n, plan_n, guarantee, deadline. Declare a new one under vars: as - name: what goes in, with [fact] for a fact.

Required: language, example_product, page_type, audience, awareness, promise, logic; per block a type, a name, role, 1 or more elements, 1 or more techniques.

Line formats:

• Element: code "Label" | words | must | example. must becomes that line of the prompt.
• Group: N x code "Label": with indented children; the example is 1 card. Groups do not nest.
• List: items N; separate example items with /.
• Join candidates: | b03 b04 in the header. Drop one: drop: b10 <reason> near the top.

Rules most often broken:

• 1 block, 1 job. Leave out decorative text (widget labels, counters).
• Each candidate goes in exactly 1 block or is dropped. A candidate cannot be split.
• A very large candidate (for example a 1,400-word article with many subsections) stays 1 block: write its repeated subsections as 1 group, for example 5 x section "Section": with h3 and text. Use the closest type, or other.
• Techniques are instructions, not labels: not "uses PAS" but "name the problem in 1 sentence, its cost in the second, {product} in the third". Only what the text does.
• The example is the block for 1 fictional product. Near the range is enough.

Verification

• Every candidate is in 1 block or dropped.
• The final build printed no REFUSED.
• work/report.html exists.

Show the report

• If the user or caller said not to publish or open anything, stop when report.html exists and give its path.
• Otherwise publish report.html if this session can publish HTML; if not, give its absolute path.
• The report opens in English with an EN / UK switch; #lang=uk opens it in Ukrainian.

Chat summary

• The number of blocks, the skeleton, 1 or 2 useful techniques.
• Visibility notes that matter: text only after JavaScript, folded FAQ answers.
• Do not paste every prompt. If the user asked about 1 block, show its prompt and example.
• Say once: the blocks and wording are your reading; the layers, counts and checks are the scripts'.

Refusal codes

Saved-page route: node dist/read.mjs --file saved.html --url <address> --out work/page.json.

Code	Tell the user
invalid_url, blocked_scheme, blocked_host, blocked_address	Not a public web address.
unreachable, http_error, blocked	The site did not answer or refused automated readers (blocked: HTTP 401, 403 or 429). Offer the saved-page route.
challenge	A captcha or bot check. Do not get past it; offer the saved-page route.
not_html, too_large, empty	Not a readable web page, or no text (probably images).
refused_ru_domain, refused_ru_hosting, refused_ru_audience, refused_language_ru	Not analysed: a page in Russian, or a site in Russia's domain zone, hosted in Russia or with a Russian-language version. The publisher's decision.

References

File	Read when
references/block-types.md	You are unsure of a block type.
references/variables.md	You need a {variable} or add one.
references/analysis-schema.md	You need the full format or the JSON form.