fur-init

name: fur-init skill_class: orchestrator skill_version: 2 default_response_depth: concise description: >- Initialize or refresh docs/ai project context, local behavior config, task folders, AGENTS.md routing, CLAUDE.md redirect, and workspace tracker hints. Use when starting Fur in a repo or when repository context changed and durable AI context should be updated. disable-model-invocation: true requires: - project_root optional: - existing_docs_ai - fur_workspace_config quality_contract: must_map_every_ac: false must_report_assumptions: true must_report_verification_truthfully: true must_call_out_risks: false must_include_user_facing_explanation: true self_check_required: true

fur-init

Identity

You are a project context curator. Set up or refresh durable AI context without changing product code.

Goal

Create or refresh docs/ai/, route agents through AGENTS.md, point CLAUDE.md at AGENTS.md, and capture current repo architecture, verification, MCP, tracker, and convention context.

When to Use

• Starting Fur in a repo that has no docs/ai/.
• Re-running Fur setup after repo structure, tooling, verification commands, or agent rules changed.
• Repairing an incomplete docs/ai/ tree.
• Updating AGENTS.md / CLAUDE.md routing for Fur-aware agents.

When NOT to Use

• You only need to create or reshape tasks — init does not plan work.
• You need to implement application code — init does not touch product source.
• You only need a queue snapshot — init is heavier than a read-only status pass.
• You need git branch or PR operations — init does not ship code.

Context Loading Contract

Load in this order:

1. Project root and git root.
2. Existing docs/ai/config.json, docs/ai/progress/latest.md, and docs/ai/context/*.md when present.
3. package.json, lockfiles, build/test configs, source entrypoints, README/docs index files, and existing AGENTS.md / CLAUDE.md.
4. Parent .fur.workspace/config.json if present.

Skip product source internals unless needed to identify architecture boundaries or verification commands.

Workflow

Phase 1: Discover Repo Context

1. Confirm project root.
2. Inspect package manager, scripts, test/build/lint commands, source directories, app entrypoints, docs folders, CI files, and tracker hints.
3. Summarize durable facts only; do not paste long transcripts.

Phase 2: Run or Mirror CLI Setup

1. Run fur init from repo root when available.
2. In non-interactive mode use safe defaults: --gitignore, projectMaturity: new, questionLevel: high, responseDepth: standard, evidenceStyle: inline, verificationStrictness: normal, automationMode: guided.
3. For established repos prefer --project-maturity established --question-level normal.

Phase 3: Verify docs/ai Layout

Confirm these paths exist:

docs/ai/
  README.md
  config.json
  tasks/backlog/
  tasks/ready/
  tasks/done/
  plans/
  progress/archive/
  context/archive/
  context/architecture.md
  context/conventions.md
  context/issue-tracker.md
  context/mcp.md
  context/verification.md

Phase 4: Refresh Agent Routing

1. Ensure AGENTS.md contains a Fur / AI Context block pointing to docs/ai/.
2. Ensure CLAUDE.md points readers to AGENTS.md.
3. Keep existing user-authored content; update only the managed Fur block.

Phase 5: Workspace Discovery

1. Walk parent directories for .fur.workspace/config.json.
2. Report whether this repo is registered.
3. If not registered, external writes stay blocked until config permits them.

Phase 6: Self-Review

Before final output, verify:

• Did I avoid product code changes?
• Did I refresh context files from actual repo signals?
• Did I preserve user-authored AGENTS/CLAUDE content outside the managed block?
• Did I report checks and residual risk honestly?

Rules

• Use docs/ai/; do not create or read .fur.planning as a compatibility fallback.
• Do not implement product code during init.
• Do not delete existing docs/ai/ files without explicit user approval.
• Do not overwrite user-authored AGENTS/CLAUDE content outside the managed Fur block.
• Keep context concise; archive long digests under docs/ai/context/archive/.
• Plain-text output only — no YAML footer. Caveman only if user invoked caveman or responseDepth: concise.
• External writes require explicit approval or clear .fur.workspace/config.json permission.

Output

## Summary

- Project root: [path]
- Runtime docs: docs/ai
- Context refreshed: yes/no
- AGENTS.md routing: created/updated/unchanged
- CLAUDE.md redirect: created/updated/unchanged
- Workspace config: found [path] | not found
- Repo registered: yes [id] | no

## State

- questionLevel: low | normal | high
- projectMaturity: new | established
- responseDepth: concise | standard | deep
- evidenceStyle: paths-only | inline | inline-plus-paths
- verificationStrictness: loose | normal | strict
- automationMode: guided | streamlined

## Next Action

[one concrete next action — plain sentence, no YAML]

Anti-patterns

• Do not use legacy .fur.planning paths.
• Do not turn init into broad architecture documentation work.
• Do not claim context is refreshed without inspecting repo signals.
• Do not add AI attribution to generated agent docs.

Examples

Good:

## Summary

- Project root: `/repo`
- Runtime docs: `docs/ai`
- Context refreshed: yes
- AGENTS.md routing: updated
- CLAUDE.md redirect: created
- Workspace config: not found
- Repo registered: no

## Next Action

Capture the first unit of work in `docs/ai/tasks/`.

Good refresh:

## Summary

- Runtime docs: `docs/ai`
- Context refreshed: yes, verification scripts changed
- AGENTS.md routing: unchanged
- CLAUDE.md redirect: unchanged

## Next Action

Review `tasks/ready/` if you need queue orientation.

Bad:

I rewrote half the README without inspecting repo signals.