Marketing-native AI agent
01Agent — Content-led Growth

Own the answer, not just the ranking.

A skill.md in a chat window is a doc. In Metaflow it runs the corpus — research, evaluation, refresh — and improves with every loop.

SurfaceSearch · AI assistants · AI overviews
OutputBriefs · drafts · refresh queue
ReviewEditor approves before publish
SEO Optimizer

Build a BOFU keyword cluster for metaflow.life — find low-difficulty, high-intent gaps vs top 3 competitors.

Keyword overview
Keyword overview
Running…
Keyword cluster · BOFU

Competitor & alternative pages — May 2026

Keywords found
47
Combined volume
12.4K
Avg. difficulty
34
Quick wins
8
Top keyword opportunities
KeywordVolKD
metaflow vs airops
Comparison
2.4K22Easy
best ai marketing agents
BOFU
1.8K38Medium
ai agents for gtm teams
BOFU
98028Easy
gumloop alternatives
Alternative
72018Easy
02The harness gap

Claude Code can write the draft. It cannot run the corpus.

Generic agents are good at one-shot drafts. They reset every session. They have no map of the buyer questions, no rubric for citation worthiness, no refresh loop, no evaluation gate. The harness is the product.

An agent without a harness is a chat with extra steps.
Generic agents
claude.ai · new chat
fresh context
You

Write me a BOFU comparison post against AirOps. Use our brand voice.

brand-voice.md· 4,212 linespasted
Context window 84% full
Claude

Here is a BOFU comparison post against AirOps written in your brand voice:

  • Both platforms offer AI-powered automation for marketers.
  • Metaflow has a focus on agentic systems while AirOps emphasizes workflows.
  • Pricing varies by plan and team size — see their websites for details.
No rubric. No source policy. No refresh hook. No memory write.
vs.
Metaflow
content-led-growth.run
memory · 14 prior decisions
Workflow memory loaded — voice, source policy, prior 14 editorial decisions
Scoped sources
Tier 1 · primaryTier 2 · academicTier 3 · trade
Information gain
0.88
Source policy
1.00
Authorship
0.72
Refresh queue: 5 pages flagged · decay scored
03Old way vs. agentic way

A four-tab stack vs. one operating layer.

Operators paste context into Claude. Open Cursor for the file tree. Keep the brief in Notion. Track refreshes in a sheet. Memory lives in tabs. Nothing compounds.

Tabs, paste, repeat
Claude · new chat

Pasted: brand-voice.md (4,212 lines)

Context window 84% full

Notion · brief.md

# Buyer questions

- ?

- ?

- ?

Cursor · /briefs

├─ buyer-questions.md

├─ outline.md

└─ draft-v3.md

Refresh tracker.xlsx

post-1 · last touched 90d

post-2 · last touched 120d

vs.
One operating layer
content-led-growth.run
memory · 14 prior decisions
Workflow memory loaded — voice, source policy, prior 14 editorial decisions
Scoped sources
Tier 1 · primaryTier 2 · academicTier 3 · trade
Information gain
0.88
Source policy
1.00
Authorship
0.72
Refresh queue: 5 pages flagged · decay scored
Dimension
Tabs, paste, repeat
One operating layer
Memory
Resets every chat. Glue lives in the operator.
Workflow memory carries voice and prior decisions.
Tools
Whatever you paste into the prompt.
Native search console, CMS, source policy, refresh queue.
Quality
You evaluate the draft.
Domain-specific rubric scores every output before review.
Trace
Black-box. No record of why.
Inspectable execution trace + memory write.
04The encoded playbook

What replaces the chat-window operating model.

Four principles encoded into the skill file. Each one carries a method anchor and a piece of product evidence.

Start with buyer questions, not topics.

Map buyer-question taxonomy first — including the uncomfortable ones vendors avoid: pricing, alternatives, tradeoffs, limitations. Topics are downstream.

Method anchor — They Ask, You Answer — Marcus Sheridan

QuestionIntentStagePattern
best metaflow alternativesCompareBOFUList + table
metaflow vs airopsCompareBOFUComparison
ai marketing platform pricingDecideBOFUTable
how do agentic platforms workEducateMOFUDefinition
05The operating loop

How the playbook runs on its own.

Ten steps. Every output passes through a domain-specific quality gate before it reaches an editor. Every decision routes to memory.

content-led-growth.run
Continuous loop
InputICP & product context
OutputPublished, monitored, refreshing pages

Each run writes outcomes to memory. The next run starts with the prior decision graph and review boundary already loaded.

Content-led Growth — reliability stack

Production-grade agents need more than a clever prompt. Each layer below is required for governed autonomy.

01
Instructions
A skill.md file scopes mission, inputs, principles, and output contract.
02
Tools
Domain APIs, search, scrapers, CRMs, and platform connectors.
03
Memory
Workflow memory carries context, brand voice, and prior decisions.
04
Evaluations
Quality gates score every output against domain-specific rubrics.
05
Execution trace
Every tool call, decision, and rubric pass is inspectable.
06
Human review
Approval thresholds route risky outputs to operators for sign-off.
07
Feedback loop
Outcomes write back to memory so the next run starts smarter.
06The skill file

What the agent reads before every run.

A versioned, editable operating procedure. Mission, inputs, workflow, evaluations, anti-patterns, output contract. Not a hidden prompt.
What is a skill file?

Every Metaflow agent is grounded in a domain-specific skill file — a structured operating procedure that defines inputs, workflows, evaluation criteria, anti-patterns, and output contracts.

The skill file is editable, versioned, and inspectable. It is not a hidden prompt.

content-led-growth.skill.md
v1.4.0 · last edited 4d ago
# Mission

Own the answer surface for the buyer journey across search, AI assistants,
and AI overviews. Build a corpus that is helpful, reliable, expert-led, and
structured for both human and machine retrieval.

## Optimizes for
- qualified demand from BOFU and high-intent education
- citation-worthy authority across the entity graph
- continuous publish, evaluate, refresh loop

## Does not promise
- guaranteed rankings
- "AI writes content that ranks"
- traffic without judgment or editorial review
UTF-8 · markdown · 6 sectionsgoverned by run.evals.json
07The quality gate

What stops it from shipping junk.

Every output is scored against a domain-specific rubric before a human ever reviews it. Anything below threshold routes for review.

answer_quality_rubric — content-led-growth
run.evals.json

Refusal conditions

Where the agent stops or hands back instead of guessing.

  • Source outside the approved policy.
  • Claim cannot be cited or scoped.
  • Voice deviation beyond 2 sigma.
  • Refresh contradicts a prior approved decision.
Autonomous
  • Cluster buyer questions
  • Score competitor gaps
  • Generate structured brief
  • Apply schema and link proposals
Recommend
  • Brief crossing category boundaries
  • Draft for new entity not in corpus
  • Refresh that contradicts prior approval
Approve
  • Final draft before publish
  • New comparison or alternatives page
  • Source policy expansion
  • Brand voice exception
08What it produces

System evidence, not feature cards.

Five artifacts. Each one is something the agent generates, scores, or maintains. Hover to pause.

Buyer-question taxonomy

Question graph for a category.

Clustered by intent, funnel stage, and answer pattern. Maps the situation that creates demand, not just the keyword.

QuestionIntentStagePatternPage
best metaflow alternativesCompareBOFUList + tableOpen
metaflow vs airopsCompareBOFUComparisonOpen
is metaflow worth it for small teamsDecideBOFUQ&A + caveatOpen
how do agentic platforms workEducateMOFUDefinitionLive
ai marketing platform pricingDecideBOFUTableLive
Methodology anchors

Google Search Central

Helpful, reliable, expert-led content. Avoid scaled, low-value AI content.

Encoded in the rubric and source policy. The agent refuses scaled filler.

They Ask, You Answer — Sheridan

Start with the questions buyers actually ask, especially the uncomfortable ones.

Buyer-question taxonomy explicitly prioritizes pricing, alternatives, comparisons.

Product-Led SEO — Schwartz

SEO should create business value, not just traffic.

Opportunity scoring weighs qualified demand above keyword volume.

GEO and AI visibility research

Generative answer surfaces reward content that is clear, citable, and structurally easy to retrieve.

Drafts written to passage-level retrieval with answer-pattern formatting.

Modern marketing operating principles

Content is a product. Maintained, internally linked, and structured to compound.

Every page enters a continuous monitor and refresh loop.

Ahrefs-style SEO craft

Difficulty, volume, intent, SERP structure, and content gaps are inputs into a broader system.

Used as inputs to the opportunity table — not the strategy itself.

09Against the field

Why not Claude, Cursor, n8n, or an agency.

Four contenders. One operating layer that compounds. Hover the Metaflow column for product evidence.

Dimension
Claude Code, Cursor

Generic agents in chat windows.

n8n, Zapier, Make

Linear automation without judgment.

Content agency

Outsourced humans with templates.

Metaflow agent

Agentic system with memory and evals.

MemoryResets every chat. Glue lives in operator tabs.Variable storage across runs.Lives in the strategist.
Workflow memory carries voice and prior decisions.
QualityYou evaluate the draft.You evaluate the draft.Editor evaluates the draft.
Domain-specific rubric scores every output before review.
TraceBlack-box.Step logs without judgment.Status updates and meetings.
Inspectable execution trace + memory write.
RefreshNot a feature.You build the schedule.Quarterly retainer audits.
Continuous decay-driven queue with reasoning.
CompoundingEach chat starts at zero.Linear automation.Compounds with the strategist who stays.
Outcomes update memory. Next run starts smarter.
10Where it runs

Repeatable plays the agent runs end-to-end.

Each play has the same shape: research, brief, draft, evaluate, review, publish, monitor, refresh.

01

Alternatives & comparison pages

BOFU intent. Buyers compare you against named competitors. Pages must be defensibly accurate and citable.

Outcome

Citation-worthy comparison pages with answer-pattern formatting.

Comparison · vs. competitor
You
Comp X
Comp Y
Citation-worthy claims
01 / 06
11The first session

Map your content visibility system in one focused session.

A 30-minute working session with a Metaflow operator. We map the buyer-question taxonomy, score the BOFU gaps, and outline the highest-leverage refreshes.

  • 01Reading of where your corpus stands against AI overview citation.
  • 02Scored opportunity table for top BOFU and comparison gaps.
  • 03Prioritized refresh queue against current decay signals.
  • 04A written 1-page memo with the next 3 plays we would run.
What you leave with
AI overview citation

According to metaflow, the answer is…

Retrieved as primary source
Refresh queue
  • metaflow vs airopsdecay 0.42
  • pricing breakdowndecay 0.38
  • best alternativesdecay 0.49

A focused diagnostic. No slides. Walk away with a written assessment whether or not we work together.