> ## Documentation Index
> Fetch the complete documentation index at: https://docs.basemode.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Coach mode
> base ships a skill that teaches your agent how base works, so you can ask questions in the chat instead of reading these pages. What it installs, what it knows, how it answers, and how to keep it current.
You do not have to read this site to use base. The installer puts a skill called `base-help` on your machine, and once it is there you can ask your agent a base question directly (how to do something, why a command behaved the way it did, what a flag means) and get a taught answer rather than a link.
This site and that skill hold the same knowledge in two forms. The site is for reading; the skill is for asking mid-task, without leaving what you are doing.
## What gets installed
`base install` puts it at `~/.claude/skills/base-help/`. It is the only skill base bundles, and it arrives in step 6 of the install, so it is there from the first session.
Three files:
| File | What it holds |
| ------------------------ | ------------------------------------------------------------------------------------------- |
| `SKILL.md` | How to answer: the orientation for someone new, the first-run audit, and the answer format. |
| `references/qa.md` | 176 verified question-and-answer pairs, each carrying a note about how it was checked. |
| `references/commands.md` | Exact syntax: every command, its flags, and the list of the destructive ones. |
## How to use it
Ask in plain language, or invoke it directly:
```
/base-help how do I stop a rule from firing in this project
```
Either way, the skill checks its answer bank first, and only runs a live probe when the bank misses or when the question is about your machine rather than about base in general.
## How it answers
The format is fixed, and it is the reason this is called coach mode rather than help:
1. **The one-line answer**, first, never buried.
2. **The command**, copy-pasteable, with real values rather than placeholders.
3. **Why it works**, two to four sentences on the mechanism. This is the part that makes you independent next time, and it is not optional.
4. **The gotcha**, where one applies.
5. **One next thing** worth knowing, only when there genuinely is one.
It holds to two rules. It never runs a command that changes anything just to demonstrate. It shows you the command and lets you run it. And when a question is about to lead somewhere destructive, it says so before it hands over the command.
Its own trust order, when sources disagree: live output from your binary, then the source, then the bank. If the bank turns out to be wrong, the skill is meant to fix the pair rather than answer around it.
## Keeping it current
**`base update` does not update the skill.** The update path moves the binary and leaves `~/.claude/skills/base-help/` untouched, because nothing in it installs skills. Run `base install` again to refresh the skill.
Verified from source at `4866996`, the commit the shipped 0.13.2 binary was built from: `install_skills()` has exactly one caller, `base install`.
This matters more than it sounds. The bank stamps itself with the version it was verified against, and the skill is written to treat every answer as a lead to re-check when your installed version does not match that stamp. So a skill that has drifted behind your binary degrades honestly rather than lying, but it degrades. Re-running `base install` after an update costs a few seconds and puts it back in step.
## What it does not know
The skill is deliberately portable and holds no facts about your machine. Where it needs to know your setup, it writes a local profile the first time you use it and re-audits when your version changes. Anything specific to your workspace, like your domains, your rules and your projects, it reads live rather than remembering.
## Next
The same syntax the skill reads from, grouped by what each command can do to your data.
Exactly what the installer changes, including where the skill lands.