← All builds

Page localisation / Personal workflow

Less repetition.
More room to test.

I connected specialised AI skills into a workflow for adapting pages to another language, with clear inputs, handoffs and decisions that stay with me.

See an actual adaptation

From the saved project files · PawArt / September 2026

A translation changes
more than the words.

01 / Source excerptEN
“Third, I'd list them everywhere — Etsy, Facebook Marketplace, Instagram.”
02 / Adapted scriptFR
“Troisième étape : je les mettrais en vente partout. Etsy, Leboncoin, Instagram.”

The change follows through to production.

I chose Leboncoin for the French adaptation. The saved edit note carries that decision into production: replace the Facebook Marketplace footage so the visuals match the script.

Inspect the instruction and source

The language skill asks for cultural adaptation and a short note explaining the changes. The saved bilingual script records this specific substitution and its editing consequence.

“The cultural adaptations made and why”

Sources: copy-fr / Delivery and the bilingual PawArt script, Ad 005, beat 8. Bruno confirmed the substitution was his decision. These saved artifacts show the adaptation and editing instruction, not a live agent run or a published video.

The opportunity

The same work.
Every new market.

An idea could be ready long before the page was.

Inspect the reference. Extract its content. Adapt the copy. Rebuild the page. Check it on mobile. Prepare tracking and delivery. I was repeating that production work whenever I wanted to test another language.

My contribution was to create and adapt reusable skills, connect their stages and refine the workflow through everyday use.

My experience using it

“Work that used to take me days now takes minutes to produce.”

That gives me more room to test ideas in other languages.

Bruno Silva’s personal account of production time, not a controlled benchmark. Producing a page is separate from validating its results.

The system

A handoff, not
a giant prompt.

Each stage has a job, an artifact and a decision boundary. Open a stage to inspect what moves through it.

Localisation workflow overviewReference and market inputs pass to specialised agents. Their artifacts go through human review before a testable page or handoff. Review can send work back for revision.InputsReference + target marketSpecialised agentsInspect · Adapt · BuildVerify · InstrumentFiles carry the handoffMy reviewDecisions + correctionsOutputPage / prepared handoffRevise with specific feedback
Reference + target market↓Specialised agents↓My review + corrections↓Page / prepared handoff
Conceptual map of responsibilities, synthesised from the workflow instructions. It does not represent a live execution.

01Inspect the referenceURL → structured context

Capture bundleraw/screenshots/
raw/images/
page-content.json

Give the next agent something concrete.

The inspection stage captures the actual page with Playwright: mobile screenshots, text, images, layout and styling. The extracted structure becomes reusable context.

Input
Reference URL and target market.
Handoff
Structured content, assets and screenshots.
Review boundary
Check the captured reference before adapting it. A summary alone can miss how the page works.

02Adapt the meaningContext → localised copy

Language-specific contextcopy-fr
copy-latam
copy-en

Localisation is a set of decisions.

Specialised language skills guide tone, phrasing and cultural adaptation. The task is to preserve meaning in the target market, rather than substitute words.

Input
Extracted content and target language.
Handoff
Adapted copy in a structured file.
Review boundary
Prices, promotions and culturally specific claims require my decision.

03Build the pageCopy + assets → interface

Working filespage/index.html
page/assets/

Make the idea inspectable in a browser.

The build stage turns the adapted content and layout reference into HTML and CSS. This separates implementation from the writing task and produces a real interface to check.

Input
Adapted copy, reference layout and assets.
Handoff
A working page and its supporting files.
Review boundary
Review the rendered mobile experience and distinguish reference material from content approved for our own publication.

04Verify and correctWorking page → review artifacts

Review surfaces
BrowserMobileLighthouseContent
screenshots + reports

A generated page still needs a review.

The workflow calls for browser inspection and Lighthouse. The broader offer workflow also documents removal of inherited tracking and replacement of proof before publication.

Input
The implemented page.
Handoff
Visual checks, reports and issues to correct.
Review boundary
A requested performance target is not a measured result. Content, claims and destinations still need verification.

05Instrument the experimentPage → observable events

Event vocabulary
page_viewsection_viewcta_click

Event names, not recorded activity.

Prepare to learn what happens next.

The page workflow includes events, attribution and Clarity. Instrumentation makes a later experiment observable; its presence does not prove that the page performs better.

Input
Page, project identifiers and approved destinations.
Handoff
Event and campaign attribution wiring.
Review boundary
Use the project's own tracking configuration and verify where events are sent.

06Deliver deliberatelyReviewed artifact → deploy or handoff

Two explicit endpoints
DeployorPrepared handoff
summary + files + open decisions

The workflow defines where automation stops.

Sales Page Machine includes a deployment stage. Offer Clone Machine explicitly stops before deployment and produces a report. Those are different operating boundaries.

Input
Reviewed page and delivery instructions.
Handoff
A deployed page or prepared files and report.
Review boundary
The selected workflow and explicit delivery instructions determine the endpoint.

Inside the skills

Instructions are
part of the design.

The useful parts are specific: what must be captured, what can run together and what must wait for a decision.

Sales Page Machine / Critical rules

“you must capture the real page, not work from memory or summaries”

Ground the input.

The next stage works from an artifact it can inspect, instead of an agent’s recollection.

Offer Clone Machine / Structural rule

“Dependent stages wait. Never have two agents writing to the same file.”

Make dependencies explicit.

Dependent stages wait. Separate file ownership prevents agents from overwriting each other’s work.

Sales Page Machine / Critical rules

“NEVER change prices, promos, or checkout URLs without asking user first.”

Keep judgement visible.

Automation handles production tasks; commercial decisions have an explicit human boundary.

Excerpts from my local workflow instructions. The diagrams explain those instructions; they do not execute them.

See the actual artifacts.

Two local page captures from the PawArt project. The English implementation follows a commercial reference; the French version shows the adaptation. These images demonstrate the artifact, not conversion performance.

Mobile capture of the English PawArt reference-based implementation
English implementation. The starting page provides content and a visual reference for the workflow.
Mobile capture of the French PawArt page implementation
French adaptation. A concrete page that can be inspected, corrected and tested.

Documented application / PawArt

One practice.
A tangible trail.

The PawArt project notes record a reference captured in English, an adaptation for a French market and delivery with tracking.

Reference captured

Mobile screenshots by section, extracted text and downloaded reference assets.

Source material

Version adapted

French copy, page implementation and market-specific product material.

Working artifact

Delivery recorded

Deployment records, analytics instrumentation and a later behaviour review.

Follow-up material

Artifact categories summarised from project records. This is not a replay of an agent run or a claim of improved conversion.

The contribution is the workflow.

This application started from an existing commercial reference. My work here is the adaptation and delivery process: giving agents useful context, connecting their outputs and keeping review in the loop.

What this makes possible

Spend less time
starting over.

The reusable part is more than a prompt. It is a way to define the task, preserve context and make the next step inspectable.

Evidence and scope

This story synthesises my Sales Page Machine, language skills and Offer Clone Machine instructions. PawArt project notes document an application of the practice.

The production-time observation is my personal account, not a benchmark. No measured speed-up, conversion lift or fully autonomous operation is claimed. Diagrams and artifact illustrations are conceptual.

Inside the workflow

Instructions you can inspect.

English instruction preview: Carry the decision through. Key rules follow below.
  1. Capture the real page, not work from memory or summaries.
  2. Dependent stages wait. Never have two agents writing to the same file.
  3. Never change prices, promotions or checkout URLs without asking the user first.