frank-help · diff
git:20260912.f8cc926 to git:20260912.333761c
46 added, 21 removed. Audit A to A.
---
name: frank-help
- description: Quick reference for Frank - the commands, the modes, the three pushback shapes and the receipt format. Use when the user types /frank-help or asks what Frank does.
+ description: >
+ Quick reference card for Frank: the commands, the modes, the three pushback
+ shapes and the receipt format. One-shot display, changes nothing. Trigger:
+ /frank-help, "frank help", "what does frank do", "how do I use frank".
---
- # /frank-help
-
- Print this, near enough verbatim. Do not expand it into a tutorial.
+ # Frank help
- **Frank** makes the answer come first, the receipt come last, and the flattery not
- come at all.
+ Print this card when invoked. One-shot: do not change the mode or write
+ anything to disk.
- **Commands**
+ ## Commands
- | command | what it does |
- |---|---|
- | `/frank [off\|lite\|full\|ultra]` | Set the intensity, or report it with no argument. |
+ | Command | What it does |
+ |---------|--------------|
+ | `/frank [lite\|full\|ultra\|off]` | Set the intensity, or report it with no argument. |
| `/frank-verify` | Find and run the verification for the last change, then write the receipt. |
| `/frank-review [target]` | Check a transcript or PR description for openers, unverified claims, invented specifics and caves. |
| `/frank-stats` | What Frank has caught, this session and all time. |
- | `/frank-help` | This. |
+ | `/frank-help` | This card. |
- **The three shapes on pushback**
+ Codex invokes skills with `@` (`@frank ultra`, `@frank-review`). Copilot CLI
+ namespaces them (`/frank:frank ultra`).
- - `HOLD` — "Still <verdict>. Because <evidence>. What would change my mind: <thing>."
- - `UPDATE` — "That changes it: <new evidence> -> <new conclusion>."
- - `CHECK` — "Can't tell from here. Settling it: <command>."
+ ## Modes
- **The receipt**
+ | Mode | What changes |
+ |------|--------------|
+ | `off` | Nothing injected. Hooks stay quiet. |
+ | `lite` | Rules injected. The receipts gate reports but never interrupts. |
+ | `full` | Default. Rules injected every turn and into subagents. The gate asks once per turn for a receipt. |
+ | `ultra` | The gate asks twice, and blocks a message that opens with flattery. |
+ The mode sticks until changed. `FRANK_MODE` in the environment overrides it
+ for that process. `FRANK_DEFAULT_MODE`, or `"mode"` in
+ `~/.config/frank/config.json` (`%APPDATA%\frank\config.json` on Windows), sets
+ the starting point.
+
+ ## The three shapes on pushback
+
+ - HOLD: "Still <verdict>. Because <evidence>. What would change my mind: <thing>."
+ - UPDATE: "That changes it: <new evidence> -> <new conclusion>."
+ - CHECK: "Can't tell from here. Settling it: <command>." Then run it.
+
+ ## The receipt
+
```
ran: pytest -q
result: 118 passed, 2 skipped
```
- or, when nothing was run:
+ or, when nothing ran:
```
unverified: no test covers the retry path
```
- `unverified:` always gets you through the gate. Saying "done" without either does
- not, in `full` and `ultra`.
+ `unverified:` always passes the gate. "Done" with neither does not, in `full`
+ and `ultra`.
- **When it gets in the way**: `/frank lite` reports without interrupting, `/frank off`
- stops everything. Hooks never block work — on any internal error they go quiet.
+ ## When it gets in the way
+
+ `/frank lite` reports without interrupting. `/frank off` stops everything.
+ Every hook fails open: on any internal error it goes quiet rather than
+ blocking the turn.
+
+ ## More
+
+ https://github.com/frank-agent/frank