Assistant Setup (Claude, ChatGPT/Codex, Gemini)

Help the user build or improve a persistent assistant setup by applying the bundled guides to their actual files. Load only the guides the request needs.

What this folder contains. references/ holds the complete Cluide guide set (00–35) and references/00_INDEX.md, which describes every guide and the reading tracks. Nothing else is needed to run the skill. The Cluide repository also has runnable tasks/ and copy-paste templates/; they are not bundled. When they are reachable (a Cluide checkout in the working folder), prefer them. When they are not, follow the guide directly and build the files from what it describes; the result is the same in substance but you write the structure yourself, so say so. Do not report the missing folders as a fault. A guide link to tasks/ or templates/ is a pointer to that optional checkout.

Which guide to load

When the request does not clearly match a row, read references/00_INDEX.md first.

How to handle a request

Clarifying questions: for any step with a fixed set of options, use the current surface's question tool when available, otherwise ask a concise question in chat.

1. Identify the surface before choosing a setup: Claude Code, Claude conversational or Cowork, ChatGPT project, Codex workspace, a Google surface (Antigravity, Jules, Gemini app), or more than one. Record file access, native skill support, available tools and scheduler access. Guide 35 owns coexistence, Guide 01 shared policy, Guide 25 the app bootstrap. A file upload is not an installed skill, and a missing Claude file is not a defect in an OpenAI or Google project. Google routes are documented in the guides but untested: say so when you rely on one, and verify any product fact against current official documentation.
2. Classify the request: start fresh, create something new, improve something existing, fix something broken, or learn. For improve and fix, ask for the relevant files and read them before recommending anything.
3. Load the guide(s). One guide for a focused request. For "set up everything", start from 00_QUICKSTART.md, then build the layers 01 → 03 → 04, helping the user finish each before the next.
4. Act. If a Cluide tasks/ folder is reachable, run the matching task (tasks/README.md routes them); it carries its own approval gates and restore points. Otherwise follow the guide. Create real files adapted to the user's domain and names; never leave placeholders such as {project_name} in a file you hand over.

Starting structures

Guide 24 gives the standard project layout, Guide 06 and Guide 07 Part 9 what a scheduled task's files hold, and Guide 14 the data-project stubs. Where the Cluide templates/ folder is reachable, copy from it instead: PROJECT_TEMPLATE/ with BLOCKS.md (any project), TASK_TEMPLATE/ (scheduled task), AI-ASSISTANT_TEMPLATE/ (Microsoft 365 business assistant), PMO_TEMPLATE/ (programme registers).

A new personal data project

Follow Guide 14: interview (purpose, where data lives, what the assistant should do); recommend which of its five patterns apply, one line of reason each, and confirm; create the structure, a stub for each selected pattern, shared policy, native adapters and .gitignore; set up git, and a GitHub repo if gh is available and the user wants one; end with numbered next steps. Keep its design rules: store facts, not computed values; always include last_updated; the assistant reads data through feeder scripts, never raw files.

Improving an existing setup

1. Ask for the relevant files (AGENTS.md, CLAUDE.md, app bootstrap, task, skill and data files), what the setup is for and where it hurts, not what the user thinks is wrong; their diagnosis anchors yours. For a review or second opinion, load Guide 27.
2. Read the files and assess both ways: what is sound and must survive editing, and what is missing or working against them.
3. Write findings down before discussing them: file and line, what is wrong, severity (blocker / material / minor), confidence, and the evidence that would make you drop it. A finding with no answer to that last one is an opinion; leave it out.
4. Say plainly when nothing needs changing. Implement only after confirmation.

Common fixes: instruction file too long or vague (01); skill not triggering (03); task loading too much context (06); data read raw instead of through feeders (14); no git for assistant state (11); no self-improvement on a running task (07, Part 9).

Style

Create files rather than describing them, use the user's real names and structures, and summarise what you did. With a less technical user, say in one sentence what each file is for before creating it.