Prompt & behavior

Encode how your team talks about paging, and how repos and runbooks attach to each triage run. In the Exhale app this is AI → Prompt & behavior. Workspace admins can save; members and viewers can read the current values. Growth or higher; Trial and Starter see locked cards.

Related: Display & model · Quality & feedback

---

Instructions

Org-specific playbook text, severity tone, and escalation phrasing that Exhale appends to the platform triage prompt.

Use this group to encode how your team talks about paging, war rooms, and stakeholders — not to change which model or connectors run.

Custom instructions

Approved playbook text appended before the JSON schema that shapes triage output. The output schema is fixed by the platform. Please avoid pasting secrets or API keys.

Add standing guidance (naming, severity language, links to internal runbooks) that should apply to every run. Leave blank to use the platform prompt only.

Severity handling

How urgency already present in the ingest payload influences triage tone and next steps. Default follows the payload; Emphasize high leans into SEV1-style urgency; Always calm keeps a steadier voice.

Choose Always calm if the team wants a consistent tone regardless of vendor urgency. Use Emphasize high when missed high-urgency pages are a problem.

Escalation language

Short snippets for when triage recommends paging, opening a war room, or writing a stakeholder update. Blank fields keep the platform default (280 characters each).

Fill these in when your org has preferred phrases for those three moments. They steer wording; they do not page anyone by themselves.

---

Context use

How repos, runbooks, and observability context attach to each triage run: automatic runbook pull, a character budget, and per-connector opt-out.

Tighten the budget or opt out a connector if prompts are too large or a source should stay on-screen only.

Runbook attachment

Whether matching runbook documents are included in the LLM prompt (Automatic) or omitted (Off). You can also cap how many documents attach, up to the plan maximum.

Turn Off if runbooks should stay visible on Pulse detail but not go to the model. Lower the document cap when prompts are too long.

Token budget

Maximum characters of combined source-control, runbook, and observability context per Pulse before truncation. Each present section receives an equal floor of that budget; shorter sections release unused characters to the longer ones. This is a character ceiling, not a live tokenizer view.

Raise it (up to the plan maximum) when triage misses repo, runbook, or observability detail. Lower it if runs are slow or truncated awkwardly.

Context opt-out

Exclude specific context connectors from the LLM prompt while keeping them on Pulse detail for operators. Unchecked connectors still display; they are not sent to the model.

Opt out a source that is useful to read but too sensitive or noisy for inference.