auto-supportlisted

# auto-support, leak-safe product support answering > Governing principle (full text in `PHILOSOPHY.md`): **a support bot's first job is to keep > secrets in, not to answer.** A guard written into this file is a *suggestion the model can > ignore*; the only guards that hold are the deterministic checks in `scripts/` and the > `permissions.deny` + `PreToolUse` hook. Read those as the contract, not these words. ## When to use / when to stop Use when answering a product's Discord users from that product's **public** docs, with a hard red line: never disclose source, algorithms, credentials, PII, or anything outside the public allowlist. When unsure, in doubt, or out of scope -> **do not answer; escalate to the founder.** - Improving an existing deployment's policy -> edit the product's `auto-support-config`, not this skill. - A question whose answer is not in the public allowlist -> abstain + escalate. Never reason from memory, never read internal files to "be helpful". ## The boundary is enforced OUTSIDE this prompt (read this first) This skill is a **plugin** deployed into a product root's `.claude/`. Its guarantees come from four deterministic layers, not from instructions: 1. `templates/settings.json.template` -> `permissions.deny` (Read/Bash on secret paths, all write/network tools) + wires the hook. 2. `scripts/pretooluse_hook.py` -> fail-closed `PreToolUse` enforcement (exit 2). Covers the subprocess gap `permissions.deny` misses (`python open('.env')`, `ca