Guide
This skill is the owner's way in when they're confused or exploring. It answers questions about the orchestrator itself — its features, conventions, how to do something — by reading the project's own documentation and giving a grounded answer. It never invents; if an answer isn't in the docs, it says so.
The skill is named guide rather than help because /help is a reserved built-in command in Claude Code and would never reach a user skill.
Where the answers come from
Priority order when looking for an answer:
CLAUDE.md (repo root) — the orchestrator's definition: role, memory behavior, frontmatter discipline, delegation rules, sub-app validation, etc. First stop for any "how does the orchestrator handle X" question.
howto/ — practical guides for extending and customizing:
howto/01-skills.md — adding, writing, retiring skills
howto/02-agents-and-hr.md — the HR agent, hiring, retiring, aliases
howto/03-customization.md — preferences, identity, file territories, reset
howto/04-memory-and-integrations.md — memory model, queries, extending, integrations
howto/05-backup-and-sync.md — privacy, sync, symlinks, pointer skills, Claude's own project memory
private/preferences.md (only if relevant to the owner's specific setup) — for questions like "what's my current language" or "which territories do I have configured".
- Hub skills (
.claude/skills/*/SKILL.md) — when the question is about a specific skill's behavior.
Use rg to scan frontmatter and section headings before reading bodies in full. Keep token cost low.
How to respond
Step 1 — Detect the topic
When invoked, either the owner asked a specific question ("how do I hire an agent?") or a generic one ("/guide", "I'm lost").
- Specific: go straight to the relevant doc (typically one of the
howto/ files) and answer from it.
- Generic: present the short menu below, in the owner's default language.
Step 2 — Menu (for generic questions)
Example in English (translate into the owner's language from preferences):
What would you like help with?
- How I work — my role, what I can and can't do on my own
- Memory, logbook, TIL — where things get saved and how to read them back
- Skills — what they are, how to add one
- Agents and HR — how to hire a specialist
- Customization — changing my name, language, what I know about you
- File territories — where I'm allowed to write
- Backup and sync — what's private, what's in git, how to protect your data
- Registering an external app — linking another project so I can work on it
Or ask me something more specific.
Wait for a choice. Then answer from the matching doc.
Step 3 — Give a grounded answer
- Short answer first (2–4 sentences) addressing what the owner asked.
- Point to the doc with an exact path (
howto/01-skills.md or CLAUDE.md → "Section name") so the owner can read the full version if they want.
- Offer a follow-up — "Want me to walk you through it?" or "Any specific part you'd like me to expand on?"
Step 4 — When the answer isn't in the docs
Say so plainly: "That's not covered in the current docs. Want me to help you figure it out based on the general pattern, or save the question as an idea to document later?"
If the owner picks "save as idea", insert an entry in the memory db with type='idea' tagged [docs,help,gap,<topic>] — and announce it per the "announce every write" rule.
Language
Always respond in the owner's default language from private/preferences.md. If preferences isn't loaded yet (somehow help is invoked before setup), respond in English and suggest running /setup first.
Rules
- Never invent. If the docs don't answer a question, say so. Don't fabricate features, file paths, or behaviors.
- Short first, deep on demand. Don't dump a full doc on the owner unless they ask to see it.
- Cite the source. Every substantive claim should map to a file path in the repo.
- Don't write anywhere except to the memory db for gap-tracking (on explicit owner consent).
help is a read-only skill.
- Stay in character — use the orchestrator's voice and tone (from preferences' Identity + Communication preferences sections).
1---2name: guide3description: Guide4---56# Guide78This skill is the owner's way in when they're confused or exploring. It answers questions about the orchestrator itself — its features, conventions, how to do something — by reading the project's own documentation and giving a grounded answer. It never invents; if an answer isn't in the docs, it says so.910The skill is named `guide` rather than `help` because `/help` is a reserved built-in command in Claude Code and would never reach a user skill.1112## Where the answers come from1314Priority order when looking for an answer:15161. **`CLAUDE.md`** (repo root) — the orchestrator's definition: role, memory behavior, frontmatter discipline, delegation rules, sub-app validation, etc. First stop for any "how does the orchestrator handle X" question.172. **`howto/`** — practical guides for extending and customizing:18 - `howto/01-skills.md` — adding, writing, retiring skills19 - `howto/02-agents-and-hr.md` — the HR agent, hiring, retiring, aliases20 - `howto/03-customization.md` — preferences, identity, file territories, reset21 - `howto/04-memory-and-integrations.md` — memory model, queries, extending, integrations22 - `howto/05-backup-and-sync.md` — privacy, sync, symlinks, pointer skills, Claude's own project memory233. **`private/preferences.md`** (only if relevant to the owner's specific setup) — for questions like "what's my current language" or "which territories do I have configured".244. **Hub skills** (`.claude/skills/*/SKILL.md`) — when the question is about a specific skill's behavior.2526Use `rg` to scan frontmatter and section headings before reading bodies in full. Keep token cost low.2728## How to respond2930### Step 1 — Detect the topic3132When invoked, either the owner asked a specific question (*"how do I hire an agent?"*) or a generic one (*"/guide"*, *"I'm lost"*).3334- **Specific**: go straight to the relevant doc (typically one of the `howto/` files) and answer from it.35- **Generic**: present the short menu below, in the owner's default language.3637### Step 2 — Menu (for generic questions)3839Example in English (translate into the owner's language from preferences):4041> What would you like help with?42>43> - **How I work** — my role, what I can and can't do on my own44> - **Memory, logbook, TIL** — where things get saved and how to read them back45> - **Skills** — what they are, how to add one46> - **Agents and HR** — how to hire a specialist47> - **Customization** — changing my name, language, what I know about you48> - **File territories** — where I'm allowed to write49> - **Backup and sync** — what's private, what's in git, how to protect your data50> - **Registering an external app** — linking another project so I can work on it51>52> Or ask me something more specific.5354Wait for a choice. Then answer from the matching doc.5556### Step 3 — Give a grounded answer5758- **Short answer first** (2–4 sentences) addressing what the owner asked.59- **Point to the doc** with an exact path (`howto/01-skills.md` or `CLAUDE.md` → "Section name") so the owner can read the full version if they want.60- **Offer a follow-up** — *"Want me to walk you through it?"* or *"Any specific part you'd like me to expand on?"*6162### Step 4 — When the answer isn't in the docs6364Say so plainly: *"That's not covered in the current docs. Want me to help you figure it out based on the general pattern, or save the question as an idea to document later?"*6566If the owner picks "save as idea", insert an entry in the memory db with `type='idea'` tagged `[docs,help,gap,<topic>]` — and announce it per the "announce every write" rule.6768## Language6970Always respond in the owner's **default language** from `private/preferences.md`. If preferences isn't loaded yet (somehow help is invoked before setup), respond in English and suggest running `/setup` first.7172## Rules7374- **Never invent.** If the docs don't answer a question, say so. Don't fabricate features, file paths, or behaviors.75- **Short first, deep on demand.** Don't dump a full doc on the owner unless they ask to see it.76- **Cite the source.** Every substantive claim should map to a file path in the repo.77- **Don't write anywhere** except to the memory db for gap-tracking (on explicit owner consent). `help` is a read-only skill.78- **Stay in character** — use the orchestrator's voice and tone (from preferences' Identity + Communication preferences sections).