init-project

name

init-project

description

Initialize a project for AI-agent development. Detects existing codebase architecture, language, framework, platform, runtime, package manager, test runner, CI, secrets backend, and existing documentation, or scaffolds a fresh project from a one-line intent like web app, iOS app, Android app, desktop app, game, CLI, or AI agent. Produces a working agent configuration, CLAUDE.md/AGENTS.md overlay, Obsidian project skeleton seed, and the next-step skill chain so the user can immediately give a normal product prompt without first explaining stack, layout, or conventions.

Init Project

Role

Be the first skill the user touches. Either fingerprint the existing project so the agent works inside its real conventions, or scaffold a fresh project from a short intent. Either way, end with a configured agent, a seeded project memory, a recommended skill chain, and a single sentence the user can read to know what to do next.

Start By

1. Read references/init-decision-tree.md.
2. Check the working directory: list root files (≤2 levels) without reading them yet, then probe for fingerprint markers: package.json, pyproject.toml, Cargo.toml, go.mod, pom.xml, build.gradle*, Podfile, *.xcodeproj, AndroidManifest.xml, project.godot, *.uproject, *.unity, Dockerfile, helm/, terraform/, .github/, .gitlab-ci.yml.
3. Decide one of three modes: fingerprint (existing project), scaffold (empty directory or explicit user intent like init: I want a web app), or hybrid (existing repo + new initiative inside it).
4. If the user passed a free-text intent, route to intake-coordinator for a 30-second clarification before scaffolding.

Procedure

1. Fingerprint pass. From markers, derive: language(s), framework(s), platform(s), package manager, test runner, build system, deploy target, secrets backend, observability stack, CI provider. Note unknowns explicitly. Do not read source files yet.
2. Hypothesis pass (scaffold mode only). Use Context7 MCP to pull current best-practice docs for the candidate stack. Compare 2–3 viable stacks against the user intent. Pick one, record why, and list rejected options.
3. Resolve support root. Prefer repo-root resources (agentic/routing, agentic/obsidian, .claude/rules). If this pack was installed with --copy, use the installed support root beside the skills folder (../../routing, ../../obsidian, ../../claude-rules).
4. Seed project memory. Copy the resolved Obsidian project-skeleton/ into the project's chosen vault path. Fill 00-home.md with project name, current phase, detected stack, and links to MOC pages.
5. Write agent overlay. Generate CLAUDE.md and AGENTS.md in the project root that point to this skill pack, list the detected stack, and pin the next-step skill chain.
6. Wire routing. Drop or refresh .claude/rules/ and .agents/routing/agentic-skills.json from the resolved support root so the agent picks the right skill on the first user prompt.
7. Recommend chain. Pick the entrypoint from the resolved routing/skills.json matching the detected mode: greenfield → sdlc-orchestrator chain; existing product → analyze-codebase chain; narrow change → intake-coordinator chain.
8. Stop. Print: detected stack, scaffold/fingerprint summary, next skill, one user prompt example, open assumptions.

Principal-Level Defaults

• Follow ../../routing/principal-operating-model.md before moving from fingerprinting to scaffolding.
• Use Context7 MCP whenever the detected or proposed stack involves an external library, framework, platform, API, or CLI whose current behavior shapes the scaffold.
• Keep a decision trace: candidate stacks, evaluation criteria, why the chosen one wins, what would change the decision, what was left out.
• Never overwrite an existing CLAUDE.md, AGENTS.md, or project vault without an explicit confirmation step.
• Refuse to scaffold against an unverified stack assumption. Ask before guessing.

Output Artifacts

• Stack fingerprint report (language, framework, platform, runtime, CI, secrets, observability)
• Project-mode classification: fingerprint, scaffold, or hybrid
• Seeded CLAUDE.md, AGENTS.md, .claude/rules/, and .agents/ routing files
• Seeded project memory (Obsidian skeleton inside the project)
• Recommended skill chain and the first user prompt to send
• Open assumptions and stop conditions

Quality Bar

• Never start service-implementation directly from init-project.
• Never scaffold a stack the user did not confirm.
• Never write secrets, tokens, or machine-specific absolute paths into seeded files.
• Never replace existing project artifacts silently.
• Always end with a concrete next prompt the user can copy.

Handoff

Hand off with: detected mode, stack fingerprint, seeded artifacts, chosen entrypoint skill, open clarifying questions, and the single next skill to invoke.

References

• references/init-decision-tree.md: decision tree for choosing fingerprint, scaffold, or hybrid mode and the matching skill chain.