Lesson 3 · ClaudeTag

Ambient Mode: the Guardian that Speaks First

Letting prism-ops act unprompted — without turning it into noise.

~7 min · builds on Lesson 2 · Mission: better Prism/agent prompts

So far prism-ops only moves when tagged. Ambient mode flips that. With it enabled, ClaudeTag "will proactively keep you updated about whatever it thinks you might need to know … flag relevant information from across the channels it's in … and follow up on threads or tasks that have gone quiet without being resolved."[ref §4]

You already run this pattern — prism-guardian polls live state, compares against guardian-baseline.yaml, and raises the anomaly. Ambient ClaudeTag is that guardian, except it lives in the channel and you can talk back to it.

1. The two ambient behaviours

Behaviour	What it is	Your existing analogue
Flag	Notices something off and posts about it unprompted	guardian anomaly → Telegram
Chase	Follows up on a thread/task that went quiet unresolved	the dead-watchdog problem: a known issue nobody closed

2. The whole game: define "normal"

An agent that can speak unprompted is only as good as its definition of normal. Get it wrong in either direction and ambient mode fails:

No baseline → spam. Every wiggle in spread or P&L becomes a post. The channel learns to ignore it — and then misses the real one.
Baseline too loose → silence. It sits quiet through the thing that actually mattered.

Your scar tissue, generalised
Your monitoring_e2e_alert_blindspot note: alerts that fire only on a change of failure-signature go silent on a persistent unresolved problem — it pages once, then nothing. Ambient mode's "chase threads gone quiet" is the built-in fix if you tell it that an open incident stays open until a human closes it. Bake that into the instructions or you'll re-live the dead-watchdog.

3. Design the escalation ladder

The craft is matching loudness to severity. Three rungs — decide where each signal lands before Claude does:

Rung	When	Example
STAY SILENT	Within baseline. Log to memory, say nothing.	Spread 110% of normal; one losing trade inside variance.
POST A NOTE	Outside baseline, not yet dangerous. Channel-visible, no ping.	Bar age creeping to 60s; DD at 60% of the soft limit; a backtest finished.
@MENTION A HUMAN	Risk-relevant or unresolved. Ping a person, keep chasing until closed.	DD nearing FTMO floor; EA silent > threshold; data stale > 90s; an open incident no one acked.

This is just severity routing — the same logic as your medic fleet's severity-ranked findings, expressed as "how loud should Claude be."

A flagged incident sits in the channel for an hour with no human reply. Ambient mode should:

"Follow up on threads gone quiet without being resolved" is the whole point — and the antidote to your page-once-then-silent blind spot. It must NOT self-resolve a risk item (that's the destructive/confirm rule), and it must NOT go quiet.

4. The ambient policy block

Add this to the prism-ops instructions from Lesson 2 — it's the system-prompt half of ambient behaviour:

ambient:
  baseline: reference guardian-baseline.yaml; treat anything within
            it as normal and STAY SILENT (log to memory only).
  post_note_when:   metric drifts outside baseline but not risk-critical
                    (bar age > 45s, DD > 60% of soft limit, BT done).
  mention_human_when: risk-relevant OR unresolved
                    (DD near FTMO floor, EA silent, data stale > 90s).
  chase: an open risk item stays OPEN until a human acks it in-channel;
         re-ping on a backoff. NEVER self-resolve a risk item.
  dedup: don't repeat an identical flag; escalate loudness instead of
         reposting (avoid the page-once-then-silent failure).
  every alert: prefix with a UTC timestamp.

What single input makes ambient mode trustworthy rather than noisy?

Loudness without a baseline is noise. The baseline is what lets Claude decide silent / note / mention. Spend cap and tool breadth don't make it judge severity — they just let it run.

5. Your win

You can now turn a scoped channel into a guardian: a baseline, a three-rung escalation ladder, a no-self-resolve rule, and dedup. That's a complete ambient policy — and it's exactly the spec you'd hand any always-on agent, ClaudeTag or your own launchd guardian.

Read this next ~5 min, primary source on ambient framing: Introducing Claude Tag (the "ambient" and "proactive" passages).

Ask your teacher We've built the contract (L1), scoped it (L2), and made it proactive (L3) — that's a full prism-ops ClaudeTag on paper. Next options: Lesson 4 — multiplayer & separate identities (a prism-ops Claude vs a finance Claude, and the memory walls between them), or bring me a real guardian recipe and we'll port it to this ambient-policy form.