Skip to content

project-context

When .claude/stack.json exists, ALL skills and agents MUST:

  1. Read it before starting work — it contains the project’s technology decisions
  2. Never ask redundant questions — if the stack is decided, use it
  3. Activate relevant reference skills — if stack says "auth": "clerk", load the clerk skill
  4. Respect the architecture — don’t suggest alternatives unless the user asks
{
"framework": "nextjs",
"styling": ["tailwind-css", "shadcn-ui"],
"auth": "clerk",
"database": "supabase",
"orm": "drizzle",
"deploy": "vercel",
"testing": ["vitest", "playwright"],
"monitoring": "sentry",
"apis": ["ga4-api"],
"packageManager": "pnpm",
"phase": "build",
"decided": "2026-02-19"
}

Each key maps to an armadillo skill name. If the skill exists, use it. If it doesn’t exist, the fresh-project orchestrator will have generated it on-demand via writing-reference-skills.

When .claude/PROJECT.md exists, it contains the project brief — who, what, why, constraints. Read it for context on ANY task. Don’t re-ask questions it already answers.

When .claude/fresh-project.json exists, the project is in an active fresh-project flow. Read it to determine which phase is current. If a phase is in_progress or pending, the fresh-project skill manages it — don’t interfere.

Normal behavior. No stack context, no project brief. Skills operate independently as usual.