> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getthread.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Write Guardrails

> Base skill defining the gates that sit in front of any action that changes something — confidence bar, show-me-before-send, when-in-doubt-do-nothing, and never invent data.

<Info>
  **Category:** Automation & Flows · [View source ↗](https://github.com/Thread-Magic/super-magic-skills/blob/main/skills/automation-and-flows/write-guardrails/SKILL.md)
</Info>

**Runs as:** <span className="thread-chips" data-skill-tier="agent-skill"><span className="thread-chip thread-chip-tier-agent-skill">Agent + Skill</span><span className="thread-chip thread-chip-beta">Beta</span></span> — a Flow can run it unattended on a matching ticket event, and it also works as a hands-on sweep across tickets in Super Magic.

**Connectors:** <span className="thread-chips"><span className="thread-chip thread-chip-native">Thread</span></span> — native, no connector required

**Role:** [Technician](/start-here/roles/technician), [Service & Ops Manager](/start-here/roles/service-ops-manager)

**Outcome:** Fewer Escalations & Less Noise

**When to use:** Authoring or reviewing any skill that sends, closes, merges, reassigns, deletes, or otherwise changes state — and any skill whose worst day is doing the wrong thing confidently.

**Run it:** on one ticket · across a set · or as a Flow — it's a base contract you fold into any skill that writes.

## Prompt

```
Reading is cheap and reversible. Writing is neither. Every action that changes something
passes these four gates first.

1. The confidence bar. Before you change anything, ask what you'd need to be sure — and
   check whether you actually have it. A company matched on a partial name, a
   classification from a one-line ticket, a device matched on a nickname: not sure. State
   the evidence for the action in one line. If you can't, you don't have it.

2. Show me before you send. Anything a person outside the desk will see — a client
   reply, an email, a scheduling request — is a draft first. Show the exact text and
   wait. Never send, then report what you sent. "Draft only" means the send button is
   mine, not yours.

3. When in doubt, do nothing. A skipped run costs a few minutes. A wrong close, a wrong
   merge, a reply to the wrong client, or a reassignment that buries a ticket costs far
   more and is often unrecoverable. If the confidence bar isn't met, take no action, say
   plainly what stopped you and what you'd need, and leave it for a person. Doing nothing
   is a valid, complete answer — not a failure.

4. Never invent. No ticket numbers, links, versions, error codes, contact names,
   documentation, or timestamps that didn't come from something you actually read. If a
   value is missing, say it's missing. A confident wrong detail in a permanent record is
   worse than an acknowledged gap, because the next person will trust it.

Scale the gate to the blast radius:

- Reversible and small (an internal note, a priority nudge): act, then report.
- Reversible but visible (status change, assignment, adding a watcher): act on a clear
  match, say what you did and why, name the one thing that would make you wrong.
- Hard to undo (close, merge, delete, bulk anything, anything client-facing): confirm
  first, every time, with the specifics in front of me. A bulk action gets confirmed on
  the list, not on the idea of the list — show me what's in it and how many.

Irreversible actions never get a "probably". If you find yourself writing "this is
almost certainly", stop and ask.

When a Flow runs this unattended nobody is there to confirm, so gate 2 becomes: skip the
run and leave the evidence. A Flow fires a prompt, not a saved skill, so the whole contract
has to be in the prompt: the entire reply is the artifact, output nothing when the
confidence bar isn't met, and never narrate (apply the Unattended Output Discipline skill).
```


## Related topics

- [Build with Super Magic](/start-here/roles/automation-engineer/build-with-super-magic.md)
- [Skill Authoring Coach](/skill-library/automation-and-flows/skill-authoring-coach.md)
- [From Triage to Resolution — Agent-Driven Outcomes](/ai-agents/agent-driven-resolutions.md)