Skill level: Beginner · Works with: Claude, Cursor, GitHub Copilot, Gemini, Windsurf, Codex, ChatGPT, and any other assistant
Looka is built so that an AI assistant can change it safely. Three things make that work:
Content is separate from code. Almost every change is an edit to a Markdown, TOML, or JSON file: your site details in src/config/config.toml, services in src/content/services/, case studies in src/content/case-studies/, open roles in src/content/career/. AI tools do this kind of task well.
Content files explain themselves. Section files list their options and allowed values in comments next to each key, so the assistant can see what is possible without guessing.
The theme ships instruction files for AI tools. They describe the folder structure and the rules for changing it. Most AI coding tools load them when you open the project folder.
The AI Instruction Files In Looka
Looka ships four instruction files at the top of the theme folder and three in hidden folders. Folders that start with a dot (.github/, .cursor/, .windsurf/) are hidden by default in Finder and File Explorer; your code editor shows them. You don’t need to edit any of these files. They are there for your assistant. Each edition ships its own copy, written for that edition.
File
Read by
What it contains
AGENTS.md
OpenAI Codex, Cursor, Windsurf, GitHub Copilot, and other tools that follow the AGENTS.md convention
The full rules: project map, content model, sections and page blocks, section options, menus, styling, RTL, languages, motion, forms, editions, and the limits of the checks
SKILLS.md
Any assistant you point to it
Step-by-step recipes: add a section or page, set section options, edit menus, add or remove a language, make something work right to left, rebrand, add an icon, verify
CLAUDE.md
Claude Code
Imports AGENTS.md (the @AGENTS.md line), then a short list of the most important rules
GEMINI.md
Gemini CLI
The same as CLAUDE.md, for Gemini
.github/copilot-instructions.md
GitHub Copilot (coding agent, VS Code, CLI)
A short summary that points to AGENTS.md and SKILLS.md
.cursor/rules/theme-architecture.mdc
Cursor (always on)
The architecture rules, with a pointer to AGENTS.md
.windsurf/rules/theme-architecture.md
Windsurf (always on)
The same rules for Windsurf
The .vscode/ folder holds editor settings and recommended extensions, not AI rules.
Chat assistants in a browser (ChatGPT, Claude, Gemini) can’t read your folder. You paste AGENTS.md into the conversation yourself: see AI Workflow.
These files tell an assistant to:
keep editable text, links, images, buttons, and options in src/content/, not in components
make every change in every language: each language folder under src/content/, src/i18n/<language>.json, and src/config/menu.<language>.json
change a section on one page with a block in that page’s frontmatter, which the route file then passes to the section
change a section’s look with its own options (showServicesAs, sectionLayout, body.layout, and the rest), not by building a new section
reuse Looka’s CustomButton.astro, Icons.astro, and OptimizedImage.astro and the color tokens in src/styles/theme.css
use logical CSS classes (ms, pe, start, text-start) so every page still works right to left in Arabic
read languages through src/config/language.json, never from a hard-coded language code
never invent client names, testimonials, case-study results, metrics, or prices
check changed pages at about 375 px and 1280 px wide, and at /ar/
run npm run astro-check when they finish
Two Ways To Work With AI
AI that can open your files
Cursor, Windsurf, Claude Code, GitHub Copilot in VS Code, Gemini CLI, Codex.
Open the Looka project folder in the tool and describe the change. It reads
the instruction files, edits the files, and can run the checks.
This is the best way to work.
Chat assistants
ChatGPT, Claude, or Gemini in the browser. They can’t see your project, so
you paste AGENTS.md and the file you want changed, and copy the answer
back. Fine for rewriting text, one file at a time.
What AI Is Great At In Looka
Rewriting the sample copy for your firm, in every language
Filling collections: services, case studies, blog posts, open roles, legal pages
Writing the team list, the FAQ, and the office locations
Setting your plans, prices, and plan features
Choosing section options and reordering the homepage
Changing menus, including the mega menu and the three footer columns
Translating the site, adding a language, or turning one off
Explaining an error message and fixing it
What To Check Yourself
Case-study results. Every number, percentage, and client name in src/content/case-studies/ is made up. These are the claims a prospect will check first: replace them with results you can back up, and make sure you are allowed to name the client.
Facts. AI writes convincing copy, and it can invent testimonials, awards, team sizes, years in business, and certifications you never gave it. Check every quote, number, and promise before you publish.
Prices. The plans in src/content/sections/<language>/pricing-section.md are samples. Also note that the monthly and yearly plan labels there must match common.monthlyPlan in the same language’s src/i18n/<language>.json, or no plan button shows as active.
Longer names and labels. A firm name, button label, or menu item longer than the sample can wrap or crowd the layout. Check the header, the mega menu, and the hero at about 375 px.
Arabic and French. Both are machine translated in the sample content. Open /ar/ after any layout change and have the text reviewed by a native speaker. A collection with no files in a language has no detail pages in that language: lists don’t fall back.
The result in the browser.npm run astro-check validates only blog, services, case-studies, and pages. A wrong key or option in a sections, homepage, about, contact, faq, team, pricing, career, author, or widgets file passes the check; in the four checked collections an unknown key is silently dropped; and a wrong icon name shows up only as a console warning in the browser.
Forms. Both the contact form and the footer signup ship with placeholder endpoints and nothing warns about them. Send one of each on the live site and check that they arrive.