コンテンツにスキップ
pentaglyph

AI-target reference

このコンテンツはまだ日本語訳がありません。

What this is: the exact list of files each --ai=<target> installs, derived from the CLI source (cli/src/commands/init.ts). For how to use a target day-to-day, see the matching how-to guide (use-with-claude-code.md for the Claude flow).

Terminal window
pentaglyph init <target> --ai=<claude|cursor|copilot|generic>

If --ai is omitted, the default is generic. The other targets install editor-specific auto-load hooks that point at the same underlying docs/AI_INSTRUCTIONS.md.

What every target installs

Section titled “What every target installs”

Regardless of --ai, every project gets:

FilePurpose
docs/AI_INSTRUCTIONS.mdThe protocol the AI agent must follow when touching docs/. This is the source of truth — the editor-specific hooks just point at it.
docs/WORKFLOW.mdThe canonical placement / lifecycle rules AI_INSTRUCTIONS.md resolves to.

The editor hooks below are additive. They make the AI agent automatically read the protocol whenever its working directory matches docs/**, without the user having to remind it.

claude — most fully integrated

Section titled “claude — most fully integrated”
<target>/
├── .claude/
│   ├── rules/
│   │   ├── documentation.md      ← auto-load when editing docs/**
│   │   ├── version-control.md    ← auto-load globally (Git Flow + Conventional Commits)
│   │   └── dialogue-style.md     ← auto-load globally (consultative dialogue)
│   ├── agents/
│   │   ├── adr-writer.md
│   │   ├── architect-agent.md
│   │   ├── completeness-auditor.md
│   │   ├── discovery-agent.md
│   │   ├── doc-orchestrator.md
│   │   └── spec-writer.md
│   └── skills/
│       ├── diagram-render/
│       ├── doc-fill/
│       ├── doc-init/
│       ├── doc-status/
│       └── explain/

Use when: Your team uses Claude Code as the primary AI assistant.

What it gives you beyond the auto-load rule:

  • 6 specialist agents the orchestrator can dispatch — discovery, architecture, ADR drafting, spec writing, completeness auditing.
  • 5 skills that wrap multi-step doc work behind a single /doc-init, /doc-fill, /doc-status, /diagram-render, /explain command.
  • Three auto-load rules covering documentation, version-control conventions, and dialogue style.

This is the most powerful target by a wide margin. If you have a choice, pick claude.

cursor — auto-load only

Section titled “cursor — auto-load only”
<target>/
└── .cursor/
    └── rules/
        └── docs.md     ← copy of .claude/rules/documentation.md

Use when: Your team uses Cursor.

What it gives you:

  • The auto-load rule wakes up whenever Cursor’s working context overlaps docs/**.
  • The content of docs.md is bit-identical to Claude’s documentation.md. The protocol is the same; only the file path the editor reads from differs.

What it does NOT give you: agents and skills are Claude-specific. There is no equivalent surface in Cursor today (as of 2026-05).

copilot — auto-load only

Section titled “copilot — auto-load only”
<target>/
└── .github/
    └── copilot-instructions.md     ← copy of .claude/rules/documentation.md

Use when: Your team uses GitHub Copilot Chat as the primary AI assistant.

What it gives you:

  • .github/copilot-instructions.md is GitHub’s native auto-load surface for repository-level Copilot context. Every chat message in this repo will include the protocol.

What it does NOT give you: agents, skills, or the broader .claude/ rule set.

generic — no editor hook

Section titled “generic — no editor hook”
<target>/
└── (no editor-specific files)

Use when:

  • You use an AI agent that has no editor-integrated auto-load surface (e.g. a CLI agent, a custom MCP client, a self-hosted assistant).
  • You want to keep the project’s editor configuration uncluttered.

What it gives you: the docs/AI_INSTRUCTIONS.md file ships as usual; you must instruct the agent to read it manually at the start of each session:

“Read docs/AI_INSTRUCTIONS.md and follow pentaglyph rules whenever you touch docs/.”

Target capabilities at a glance

Section titled “Target capabilities at a glance”
Capabilityclaudecursorcopilotgeneric
docs/AI_INSTRUCTIONS.md installed✅✅✅✅
Auto-load on docs/** edits✅✅✅—
Auto-load globally (not just docs)✅—✅—
Version-control + dialogue-style rules✅———
Specialist agents (6)✅———
Skills (5)✅———

The asymmetry is unavoidable: cursor and copilot only have auto-load rule surfaces. Claude Code is the only target today that supports agents and skills.

Switching targets after init

Section titled “Switching targets after init”

You do not have to re-init to switch targets. Re-run with --force --ai=<new-target>:

Terminal window
bunx --bun @uyuutosa/pentaglyph init . --profile=standard --ai=cursor --force

Re-running with a new target installs the new target’s files but does not remove the old target’s. Delete the unused .claude/ / .cursor/ / .github/copilot-instructions.md manually if you want a clean switch.

Caveats

Section titled “Caveats”
  1. Targets are not mutually exclusive. If your team is mixed (some Claude, some Cursor), run init multiple times with different --ai flags. The targets do not conflict.
  2. cursor and copilot copy the same content as Claude’s documentation.md. This is intentional — the protocol is editor-agnostic. If a future protocol revision diverges per editor, this page should be updated to reflect that.
  3. Agents and skills are not yet portable. The 6 specialist agents under .claude/agents/ reference Claude Code’s Task tool surface. They would need substantial adaptation for non-Claude environments.
Section titled “Related”