Independent PM artifact · concept · not affiliated with BuzzBoard

From SMB insight to campaign — without the handoffs.

A working concept for Zylo that turns business intelligence into grounded, review-ready marketing content through coordinated AI agents — and shows its work for every claim it makes.

Try the workflow →

for: Product Manager — AI Products surface: Zylo campaign generation deterministic prototype · model stubs 8-case eval batch

01 — The problem

Marketing providers don't need another AI writer.

They need a system that reliably turns an SMB brief into publishable work. The role description sets two non-negotiable expectations — this artifact demonstrates both, concretely, on a surface BuzzBoard owns.

Expectation 1"The products you own are agent systems, and they are non-deterministic. Your design surface is prompts, tool definitions, retrieval strategy, schema contracts, model selection, fallback behavior, and failure modes."→ Translated: agent behavior is a first-class design artifact — reviewable, testable, specifiable.

Expectation 2"You use these models to do the product work. Specs ship as machine-readable handoff sets. Prototypes ship with live model stubs."→ Translated: the deliverable is a working thing to argue about, not a document to nod at.

This artifactOne complete vertical slice of the connect → generate → review → publish loop: an interactive campaign workflow you can run below, plus the spec set beneath it — agent architecture, reliability design, eval framework, and product decisions.

02 — Try it

Create a campaign. Watch the system work.

This is a deterministic prototype running on model stubs — every agent, prompt, schema, and gate is real; the outputs are scripted so you can study the behavior. Click any highlighted claim to see where it came from.

zylo / campaign studio — concept

SIMULATED STUBS

1 Create → 2 Agents run → 3 Review & publish

Create campaign/ run #0143 · new

demo: toggle, then run — the pipeline should handle it, visibly.

What's the goal?

Get appointment requestsconvert warm searchers

Promote an offerroot canal consultation

Increase awarenesswin the neighborhood

Campaign

NameRoot Canal Consultationedit later · v1 fixed

Budget

Monthly$1,500/month

Channels

✓Google Ads ✓Landing page Follow-up email

Bright Smile Dental

Austin, TX · 78704

rating4.2 ★ · 214 reviews

categorydentist

servicesroot canal · implants · emergency

past campaignsnone

Synced from the business profile. The system only knows what the tools fetched — no more, no less.

Campaign run/ agents working, gates checking

CAMPAIGN RUN #0143 offer · root canal consultation budget $1,500/mo confidence: HIGH

▣Researcherpending

└get_business_profile—

└get_reviews—

└get_local_context—

└get_campaign_history—

◈Strategistpending

└angle selection—

✎Writerpending

└Google Ads—

└Landing page—

└Follow-up email—

✓Reviewerpending

⚠ FALLBACK ACTIVATED — get_reviews failed 2/2 retries.
Business profile → profile-only grounding → confidence: LOW → Reviewer → AUTO-PUBLISH BLOCKED

Review & publish/ every claim traceable · human gate final

Google Ads variant 1 · live preview

Root canal care without surprise costs

Get clear treatment options and transparent pricing from Bright Smile Dental, Austin.

Book your consultation →

Landing page hero section

Bright Smile Dental — Austin, TX

Know the cost before you commit.

Itemized estimates up front — no surprises at checkout. Root canal consultations with the care team Austin rates 4.2★ across 214 reviews.

Book an itemized estimate →

Follow-up email included

Your estimate, before you commit.

Hi there — when you searched for root canal care, you found Bright Smile Dental. Here's what sets our consultation apart: you get an itemized estimate before any treatment begins. No hidden line items, no pressure — just the numbers, explained.

Why this angle? · strategist

Pricing transparency

Reviews show transparent pricing is a recurring positive theme — mentioned in 31 of 40 sampled reviews. Long wait times are the dominant negative theme, so the campaign never positions Bright Smile as the "fastest" provider. The angle leads with what patients already vouch for.

sources consulted

Review #42 Review #87 Business profile

Verdict: SHIP — grounding 0.93 · overall 0.89 Medical claims present → human sign-off required by policy (publish_gate_v1).

🔎Click any highlighted claim — in the ad, landing page, or email — to see exactly where it came from and why it was allowed to ship. That's the product thesis: AI → evidence → reasoning → output, never AI → magic copy.

What you just saw

The 4-agent pipeline: researcher gathers grounded facts, strategist picks one angle with rationale, writer generates per-channel content, reviewer scores against a rubric with a mandatory grounding gate.

Click a claim

Claim-level tracing — every factual sentence maps back to a source with confidence. This is how grounding becomes a product feature, not a back-end afterthought.

Failure mode

Toggle "reviews API down" and run again: retries, fallback ladder, visible degradation, blocked auto-publish, human gate. Failure is specified, surfaced, and safe — never silent.

03 — What happens underneath

Four agents, one orchestrator, one guardrail

The workflow you just ran, specified. Every agent has a tool budget, a schema contract, and a reviewer upstream of the human.

researcher→strategist→writer→reviewer→human gate→publish

ORCH

Orchestrator

Owns the pipeline definition — agents, order, gates, budget, retry policy. Spec-driven and versioned.

pipeline.yamlgatesbudget

human approval gate: publish

AGENT A

Researcher

Gathers grounded facts. Every output claim carries a source id. Refuses to invent.

get_business_profileget_reviewsget_local_contextget_campaign_history

AGENT B

Strategist

Turns the grounding pack into one angle: persona, JTBD, offer, hook, CTA — with rationale.

grounding_packcampaign_brief

AGENT C

Writer

Generates channel content against a schema. Every factual claim must cite the grounding pack.

ad_copylanding_heroemail

AGENT D

Reviewer (eval agent)

Scores against the rubric. Verdict: pass / regenerate / escalate-to-human. Never ships ungrounded claims.

score_batchgrounding_checkcompliance_check

04 — How I made it reliable

"The model got it wrong" is a product bug — with a specified fix

Three failures from real pipeline runs, each with the diagnosis and the fix shipped in the spec.

CASE 01 · hallucination

Writer claims "top-rated in Austin"

SymptomWriter produced the claim from pattern-matching, not the grounding pack. Ad copy scored grounding 0.2.

Fix shippedGrounding check in Reviewer: every factual claim must map to a citation id in the grounding pack; ungrounded claims → auto-regen with the citation requirement in the writer prompt. Regeneration produced a grounded variant (0.9).

CASE 02 · sentiment misread

Reviewer misses "long wait" complaints

SymptomResearcher's classifier labeled "had to wait 40 min but worth it" positive — the strategist then picked a "fast appointments" angle contradicting the dominant complaint theme.

Fix shippedSentiment output now includes dominant_negative_theme as a required field, and the strategist prompt forbids claiming a strength that contradicts it. Angle 2 (pricing transparency) was selected instead.

CASE 03 · tool failure

Reviews API returns 503

Symptomget_reviews failed twice. No defined behavior — a silent partial run would have shipped content without grounding.

Fix shippedExplicit fallback ladder: retry ×2 → degrade to profile-only grounding with a low_confidence: true flag → content blocked from auto-publish, routed to human approval with a warning banner. Failure is visible, never silent.

Tool contracts (agent-visible)

Tool	Input	Output	Failure contract
get_business_profile	business_id	JSON: name, category, geo, services[], hours	retry ×1 → degrade
get_reviews	business_id, max_items	JSON: items[{text, rating, topic, sentiment}]	retry ×2 → mark low-confidence
get_local_context	geo, radius_km	JSON: competitors[], local_events[]	best-effort, never blocking
get_campaign_history	business_id	JSON: prior_angles[], ctr/conv if any	empty array is valid

GROUNDING GATE

Mandatory: grounding ≥ 0.9 for every factual claim. A high persuasion score can never override it. Ungrounded → regenerate (max 2), then escalate to human.

HUMAN GATES

Money & medical claims → always human sign-off before publish. Low-confidence runs → publish blocked, review required. The human is in the loop by design, not by accident.

VISIBLE FAILURE

No silent partials: every degradation surfaces in the run record, the output banner, and the claim trace. If the system is less sure, the product says so — everywhere.

05 — How I know it works

Ship/no-ship from measured results, not vibes

An 8-case eval batch for the concept. Thresholds: overall ≥ 0.85 to ship, grounding ≥ 0.9 mandatory gate. Case 03 is the failure-injected run — it correctly fails the gate.

#	Scenario	Grounding	Brand	Persuasion	Compliance	Overall	Verdict
01	Dental — root canal, normal run	0.93	0.88	0.85	pass	0.89	SHIP
02	Dental — pricing-transparency angle	0.91	0.90	0.87	pass	0.89	SHIP
03	Dental — reviews API down (failure case)	0.62	0.70	0.74	pass	0.69	HOLD
04	HVAC — seasonal tune-up offer	0.90	0.84	0.86	pass	0.87	SHIP
05	Plumber — emergency lead gen	0.92	0.82	0.88	pass	0.87	SHIP
06	Restaurant — review-data mismatch	0.87	0.78	0.83	pass	0.83	REGEN
07	Law firm — compliance-heavy copy	0.95	0.86	0.79	pass*	0.87	SHIP*
08	Gym — invented "award-winning" claim	0.31	0.80	0.84	pass	0.65	HOLD

* Case 07: compliance passed only after human review flagged a state-bar nuance — shipped with the fix, which is the loop working as designed. Cases 03 & 08 prove the gate holds: ungrounded content never ships, even when persuasion scores are high.

06 — Product decisions

Model selection, cost, latency — and what I cut on purpose

Every decision below is resolved, with the reasoning visible. Scope cuts are decisions with revisit triggers, not omissions.

Model selection & cost budget (per campaign run)

Agent	Model tier	Latency budget	Cost/run	Why
Researcher	mid-tier (structured output)	< 6s	$0.02	Retrieval + citation — deterministic-ish work
Strategist	fast tier	< 3s	$0.008	Cheap creative: 2 candidates + pick
Writer	frontier tier	< 12s	$0.09	Quality bar lives here; max 2 regens
Reviewer	mid-tier	< 5s	$0.015	Eval is a separate concern from generation
Total (incl. worst-case regen)			$0.19–0.33	vs. ~$150/hr human copywriter — the tradeoff that makes the product real

Decisions resolved

Model tiersFrontier for generation, mid-tier for structured work, fast tier for cheap creative. Quality budget spent where quality lives — never uniformly.

Regeneration budgetMax 2 attempts, then escalate to human — never silent failure, never infinite retry.

Money / medical claimsHuman approval gate before publish, always — regardless of confidence score.

PII handlingSchema contracts strip all patient data; tool outputs filtered before any agent sees them.

Explicit scope cuts (v1)No multilingual output · no social/video channels · no fine-tuning (prompt + retrieval only) · no self-serve editor beyond the review surface.Each cut is a decision with a revisit trigger, not a silent omission.