Zyphr Muse — AI Authoring

Zyphr Muse is the AI assistant baked into the dashboard. It writes subject lines, drafts templates, rewrites blocks, generates whole workflows from a sentence, and analyzes broadcasts before you send — all grounded in your project's real data, not generic templates.

Muse is a paid-plan entitlement. Free accounts see the entry points with an upgrade prompt; only Starter+ plans can invoke.

What Muse can do

Surface	What it does	Model tier	Where it lives
Subject suggestions	Ranks 5 candidate subject lines for an email or push template by predicted open rate	Haiku 4.5	Template editor → "Suggest with Muse"
Draft template	Generates a full template (subject + HTML + text, or push/SMS bodies) from a one-paragraph brief	Sonnet 4.6	Templates → "Generate with AI"
Rewrite block	Rewrites a selected block per a natural-language instruction ("make it shorter", "more formal")	Sonnet 4.6	Template editor → selection toolbar
Generate workflow	Turns a sentence into a fully-structured workflow with triggers, delays, branches, and channel steps	Opus 4.7	Workflows → "Generate with AI"
Analyze before sending	Pre-send advisory citing deliverability risks, frequency conflicts, audience warnings, content notes, and a recommended action	Opus 4.7	Template editor → "Analyze before sending"

Plan tiers and monthly quotas

Each paid plan grants per-action monthly quotas. Quotas reset on the first of each calendar month UTC.

Action	Starter	Pro	Scale	Enterprise
Subject suggestions	200	1,000	5,000	Unlimited
Template drafts	25	150	750	Unlimited
Block rewrites	50	300	1,500	Unlimited
Workflow generations	5	25	100	Unlimited
Broadcast analyses	4	30	150	Unlimited

Free plan is not eligible — Muse entry points show an upgrade prompt.

Once any per-action quota is exhausted, calls to that action return a muse_quota_exceeded error until the next month. Other actions continue to work — running out of workflow generations does NOT stop you from drafting templates. The dashboard's Usage tab shows real-time progress against every quota.

Enterprise quotas are unlimited; overage above a customer-specific monthly volume is billed at a negotiated per-call rate.

Per-call cost is shown in the dashboard for transparency, but customers pay for calls, not Bedrock dollars — we absorb the underlying model cost.

What Muse can see — grounding

Muse never sees other customers' data. Every prompt is scoped to a single project_id, and the grounding context is pulled from your project's own database rows:

Subject suggestions — your project's send history (open rates by pattern, spammy-word frequency, brand-voice examples from recent subjects)
Draft template / Rewrite — your template library (available Handlebars variables, tone examples from your existing templates, optional reference template)
Generate workflow — your configured channels, known event types from existing triggers, names of existing workflows (to mirror naming style)
Analyze broadcast — your project's reputation metrics (bounce rate, complaint rate, rolling send counters), suppression list breakdown by reason, 24h and 7d send activity, topic health, scheduled-overlap window, and sender-domain verification status

Project isolation is enforced at the query layer: every grounding fetch requires project_id and every query filters on it. There is no cross-project leakage.

Privacy and PII

Muse uses AWS Bedrock as the model gateway, running Claude models in your AWS region. Anthropic does not retain prompts or completions submitted via Bedrock.

The audit layer logs only scrubbed request data — counts and shapes, not raw content. For example, the audit row for a "draft template" call includes the brief length and the number of available variables, but never the brief text or the recipient list. The full response text is persisted on the audit row because it's the generated artifact itself — review it from the dashboard audit log.

If a brief or block you paste into Muse contains real PII (a customer's name, email, support ticket text), that content is sent to Bedrock as part of the prompt. We scrub on the way in and warn you when we detect PII-shaped input in a brief, but the safest practice is to avoid pasting customer data into AI fields when you can paraphrase instead.

Invoking from the dashboard

Every Muse entry point is an "AI" sparkle button next to the existing manual control:

Template subject field → "Suggest with Muse" popover (5 ranked candidates with rationale)
Templates page → "Generate with AI" → brief-input modal with streaming preview
Template editor → "Analyze before sending" → severity-coded advisory card stack
Workflows page → "Generate with AI" → describes-in-English → step-tree preview with per-step accept/reject

Streaming endpoints emit tokens as they arrive (SSE), then a final complete event with the validated payload and cost metadata. The dashboard renders the streaming text so you see progress.

Invoking from the API

Muse endpoints live under /v1/muse/ and use the same auth as the rest of the API (JWT or API key).

Entitlement check

curl https://api.zyphr.dev/v1/muse/entitlement \
  -H "X-API-Key: zy_live_your_key"

{
  "enabled": true,
  "plan": "pro",
  "required_plan": "starter",
  "quotas": {
    "suggest_subjects": 1000,
    "draft_template": 150,
    "rewrite_template": 300,
    "generate_workflow": 25,
    "analyze_broadcast": 30
  }
}

Suggest subject lines

curl -X POST https://api.zyphr.dev/v1/muse/subjects/suggest \
  -H "X-API-Key: zy_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "channel": "email",
    "project_id": "your-project-uuid",
    "template_body": "Welcome to Acme! Click below to verify your email."
  }'

Stream a generated template

curl -N -X POST https://api.zyphr.dev/v1/muse/templates/draft \
  -H "X-API-Key: zy_live_your_key" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "channel": "email",
    "project_id": "your-project-uuid",
    "brief": "Order confirmation email mentioning the order total and estimated delivery date."
  }'

The response is a Server-Sent Events stream. Events:

token — chunked model output
variables_validated — post-processor list of auto-corrected variables and flagged unknowns
complete — final { template, metadata } payload
parse_failed — model emitted unparseable JSON; the dashboard retries once
error — plan gate, budget, throttling, or internal failure
done — stream terminator

Stream a generated workflow

curl -N -X POST https://api.zyphr.dev/v1/muse/workflows/generate \
  -H "X-API-Key: zy_live_your_key" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "project_id": "your-project-uuid",
    "description": "When a user signs up but doesn'\''t verify their email within 48 hours, send a reminder email. If they still haven'\''t verified after another 48 hours, escalate to SMS."
  }'

Returns a strict-schema workflow JSON object after Opus emits tokens. The schema validator rejects malformed output before it reaches you — the platform never auto-creates workflows from unvalidated AI output.

Analyze a broadcast before sending

curl -N -X POST https://api.zyphr.dev/v1/muse/broadcasts/analyze \
  -H "X-API-Key: zy_live_your_key" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "project_id": "your-project-uuid",
    "draft": {
      "channel": "email",
      "subject": "Welcome to Acme",
      "body": "<h1>Welcome!</h1><p>Get started by verifying your email.</p>",
      "topic_key": "welcome",
      "audience_size_estimate": 250
    }
  }'

Final advisory shape:

{
  "summary": "...",
  "recommended_action": "send" | "send_with_caution" | "delay" | "revise_before_send",
  "deliverability_risks": [ { "severity": "high|medium|low", "title": "...", "detail": "...", "fix": "..." } ],
  "frequency_conflicts": [ ... ],
  "audience_warnings":   [ ... ],
  "send_time_suggestions": [ { "suggested_at": "...", "reason": "..." } ],
  "content_notes":       [ ... ]
}

The advisory is advisory only. Zyphr never blocks a send based on AI output — it's a second pair of eyes, not a gate.

Usage and budget

curl https://api.zyphr.dev/v1/muse/usage \
  -H "X-API-Key: zy_live_your_key"

Returns month-to-date spend, remaining budget, projected month-end spend, and breakdowns by action and model tier.

Limits and errors

Error code	HTTP	Meaning
`feature_not_enabled`	403	Your plan doesn't include Muse — upgrade required
`muse_quota_exceeded`	402	Per-action monthly quota reached for your plan. Response includes `action`, `used`, `limit`
`muse_budget_exceeded`	402	Internal safety ceiling reached (should never happen before quota under normal use)
`muse_rate_limited`	429	Per-user generation cap hit (default 60/hour); retry after `Retry-After` seconds
`bedrock_throttled`	503	AWS Bedrock returned `ThrottlingException`; back off with jitter
`muse_workflow_invalid`	502	Generated workflow failed schema validation; the dashboard will surface the error
`muse_advisory_invalid`	502	Generated broadcast advisory failed schema validation

Rate limiting is keyed per user-per-action, so a buggy script can't burn through your whole account's budget by spawning many "users".

Eval harness and quality

Every Muse prompt has a committed eval baseline in source control. Each prompt is judged across 5–6 rubric dimensions by Claude Sonnet against 20+ fixtures covering edge cases. CI fails on baseline regressions. Current overall scores (higher is better, 1.0 max):

Prompt	Overall baseline
`suggest-subjects.v1`	~0.93
`draft-template.v1`	~0.91
`rewrite-block.v1`	~0.92
`generate-workflow.v1`	0.94
`analyze-broadcast.v1`	0.90

If a prompt regresses by more than 5% in production traffic, we ship a patched prompt version (e.g. v1 → v1.1) without changing the public contract.

What Muse can do​

Plan tiers and monthly quotas​

What Muse can see — grounding​

Privacy and PII​

Invoking from the dashboard​

Invoking from the API​

Entitlement check​

Suggest subject lines​

Stream a generated template​

Stream a generated workflow​

Analyze a broadcast before sending​

Usage and budget​

Limits and errors​

Eval harness and quality​