JustFlows

Documentation is English-only for now. The rest of the site follows your language.

Build with AI agents

How repository-native skills help AI coding agents understand, change, and verify Justflows safely.

8 min read

Justflows is AI-ready at the repository level. The codebase includes machine-readable instructions that give compatible AI coding agents the same architectural map, public contracts, safety boundaries, and verification expectations a human contributor needs.

Skills guide the agent; they do not add AI to your published site

Agent skills are development context stored with the source. They help an AI agent work on Justflows. They do not send site content to an AI provider, run a model in production, or require an AI service for Justflows itself.

What an agent sees

The root `AGENTS.md` is the entry point. It describes the monorepo, the ownership boundaries between packages, repository-wide invariants, and the expected checks. It tells the agent to load the matching file under `.agents/skills/` before changing code.

SkillUse it for
justflows-platformCross-cutting features and architectural work spanning packages or surfaces
justflows-server-apiExpress routes, middleware, installation, public APIs, and server rendering
justflows-extensionsPlugins (create them in plugins/<name>/), the public SDK, hooks, manifests, permissions, and .jfpkg packages
justflows-dataSchemas, queries, adapters, and PostgreSQL/MySQL/MariaDB migrations
justflows-admin-uiThe React/Vite administration UI, forms, builders, API calls, and localization
justflows-themes-blocksThemes, blocks, serialization, sanitation, CSS providers, and public rendering
justflows-deploymentDocker, shared hosting, startup, packaging, updates, and production operation
justflows-qualityTesting, debugging, review, regressions, and release-readiness checks

How the skills improve agent work

Less architectural guesswork

The agent knows where framework-neutral logic, HTTP wiring, public SDK types, admin interactions, and rendering belong.

Stable extension points first

Plugins live in plugins/<name>/ and are steered toward the typed SDK, documented hooks, permission checks, and plugins/hello-world instead of core patches.

Product constraints stay visible

Browser-first installation, three database dialects, backward compatibility, localization, security, and distribution are part of the task context.

Verification is proportional

Each skill names focused typechecks, tests, builds, and runtime paths so the agent can prove the relevant behavior.

A typical agent workflow

  1. 1

    Open the repository root

    Start the agent in justflows-platform-ce so it can discover AGENTS.md, .agents/skills/, package manifests, and the working tree.

  2. 2

    Describe the outcome

    Ask for the user-visible result and name important constraints. The agent maps the request to one or more focused skills.

  3. 3

    Let it trace before editing

    A well-scoped agent reads the existing contract and its callers, then chooses the smallest coherent edit across domain, runtime, UI, docs, and examples.

  4. 4

    Review the diff and checks

    Require a concise handoff listing changed files, behavior, checks run, and anything that could not be exercised. Human review remains the final authority.

Prompts you can use

Plugin using the stable extension surface
text
Build a Justflows plugin that adds a typed action after content is published. Follow the repository agent guide, use the extension skill, declare the narrowest permissions, update the canonical example where appropriate, and run the focused checks.
Cross-cutting platform feature
text
Add scheduled publishing end to end. Start with the platform skill, preserve PostgreSQL, MySQL, and MariaDB compatibility, enforce capabilities on the server, localize the admin UI, document the public behavior, and verify each affected package.
Review and diagnosis
text
Review this change for Justflows invariants. Use the quality skill plus any domain skill that applies. Focus on security boundaries, compatibility, missing failure states, and tests. Do not modify files; report evidence with file and line references.

Use skills in your agent

Use an AI coding agent that supports repository instructions, or explicitly tell your agent to read `AGENTS.md` and the relevant `.agents/skills/<skill>/SKILL.md` at the start of the task. The skill files are plain Markdown, so their guidance remains inspectable and portable even when an agent does not auto-discover them.

Keep the skills with the code

When platform contracts or workflows change, update the matching skill in the same pull request. That keeps future humans and agents aligned with the version they are editing.

Responsible use

  • Review changes before merging, especially authentication, permissions, uploads, archives, migrations, updates, and secret handling.
  • Give an agent only the credentials and external permissions the task actually needs.
  • Keep unrelated working-tree changes in place and inspect the final diff.
  • Treat passing tests as evidence, not as a substitute for product and security review.
  • Never assume an agent capability from marketing copy; inspect the agent, model, and tool configuration you use.