Agent config — the consistency layer

AI agents build alongside us on Cynario. For that to stay coherent, they all read the same brief before they act. Three things make up that brief — a project context file, a portable map, and a set of task-specific skills — and together they’re the project’s consistency layer: the reason two different agents (or an agent and a person) arrive at the same conventions instead of each inventing their own.

This page summarizes what each one is for. The files themselves are the source of truth — follow the links to read them.

---

CLAUDE.md — the project brief

The entry point an agent reads first. It’s the standing context for working in this repo: what Cynario is, the tech stack, the strict rules (no AI attribution anywhere; encryption is mandatory for sharing), the data model and node types, file-naming and styling conventions, the knowledge-base gradient, and the index of which skills and agents exist.

Think of it as the onboarding doc an agent re-reads every session. When a convention changes, it changes here, and every agent picks it up on the next task.

→ Read it: CLAUDE.md

---

AGENTS.md — the portable map

A condensed, tool-agnostic version of the same knowledge — structured as a “where to look” map: which file to open for a given task, the conventions that deviate from framework defaults (Svelte 5 runes only, path aliases, 100-char line width), the project’s anti-patterns and gotchas, and the security non-negotiables.

It exists so the guidance isn’t locked to one assistant. CLAUDE.md is the detailed brief; AGENTS.md is the portable quick-reference any agent runtime can load.

→ Read it: AGENTS.md

---

.claude/skills/* — auto-loading playbooks

Skills are task-scoped playbooks that load automatically when the work matches them, so an agent gets the right conventions exactly when it needs them — without front-loading everything into one giant prompt. Each is a folder under .claude/skills/ with a SKILL.md.

Skill	Loads when you’re working on…
scenario-conventions	Scenario JSON, the schema, validation, nodes/edges
svelte-flow-patterns	The graph editor — custom nodes, edges, handles
store-patterns	State management, stores, undo/redo, the operation log
p2p-sharing	Trystero, WebRTC, encrypted share links
astro-svelte-patterns	Astro pages, client directives, routing
testing-workflow	Testing, debugging, logs, guiding manual tests
dev-workflow	Build, lint/format, Docker, the dev server
knowledge-curator	Creating or moving knowledge-base/ content

→ Browse them: .claude/skills/

---

.claude/agents/* — explicit specialists

Where skills auto-load, agents are invoked on purpose for a focused job. The project ships one:

• scenario-architect — designs complex interactive training scenarios (phishing simulations, tabletop exercises, branching decision trees), structuring the narrative, decision points, and outcome paths.

→ Read it: .claude/agents/scenario-architect.md

---

How they fit together

CLAUDE.md      ─ the standing brief (read every session)
AGENTS.md      ─ the portable "where to look" map
skills/*       ─ auto-loaded when the task matches
agents/*       ─ invoked on purpose for a focused job

All four are authored config: public-safe once scrubbed, and kept clean afterward by the same guard described in How we work. That’s why we can summarize and link them in the open — the rules an agent follows are written down and reviewable, not improvised.