@inneranimalmedia/agentsam-sdk 2.6.10 → 2.6.12

package/docs/prd/PRD-agentsam-campaign.md

@@ -0,0 +1,1178 @@
+# AgentSam Campaign
+## Product Requirements Document
+
+**Package:** `@inneranimalmedia/agentsam-campaign`  
+**Product form:** Installable AgentSam plugin  
+**Status:** Experimental  
+**Primary hosts:** ChatGPT, AgentSam Local Studio, compatible MCP hosts  
+**Sibling plugin:** AgentSam Brand
+
+---
+
+# Product vision
+
+AgentSam Campaign is an evidence-driven campaign intelligence system.
+
+It helps users decide:
+
+> **What should we do right now to achieve a measurable business, marketing, audience, or growth objective?**
+
+The product should move beyond generic marketing generation by grounding campaign decisions in real context and then learning from actual outcomes.
+
+Its job is not merely to produce campaign ideas.
+
+Its job is to help users:
+
+```text
+understand
+→ strategize
+→ compare
+→ plan
+→ execute through appropriate capabilities
+→ measure
+→ learn
+→ improve
+```
+
+---
+
+# Product promise
+
+> **Plan, evaluate, coordinate, measure, and optimize performance-driven campaigns using real brand, audience, commercial, channel, and historical evidence.**
+
+AgentSam Campaign should be meaningfully more valuable than asking a generic model:
+
+> “Give me ten marketing ideas.”
+
+Its advantage should come from knowing the user's actual operating context.
+
+---
+
+# Core user questions
+
+AgentSam Campaign should be able to help answer:
+
+- What campaign should we run next?
+- What objective should we optimize for?
+- Who should we target?
+- Which audience segment is most promising?
+- What should the campaign say?
+- Which products or collections should it feature?
+- Should we run a promotion?
+- Should we discount?
+- What offer is commercially reasonable?
+- What SEO/search opportunity exists?
+- Which channels fit this campaign?
+- What creative variants should we test?
+- What content is required?
+- Which concept is strongest and why?
+- What assumptions are we making?
+- What evidence supports the recommendation?
+- What happened after we launched?
+- What performed?
+- What failed?
+- What should we do differently next time?
+
+---
+
+# What Campaign is not
+
+AgentSam Campaign is not:
+
+- AgentSam Brand;
+- a brand-identity generator;
+- a giant CMS;
+- a product manufacturing engine;
+- a product catalog authority;
+- an ad network;
+- a generic commerce API;
+- a generic analytics dashboard;
+- a guarantee of revenue;
+- a guarantee of SEO ranking;
+- an autonomous unlimited-spend marketing system.
+
+Campaign composes other capabilities where appropriate.
+
+---
+
+# Relationship with AgentSam Brand
+
+AgentSam Brand and AgentSam Campaign are sibling plugins.
+
+Brand answers:
+
+> **Who are we?**
+
+Campaign answers:
+
+> **What should we do now with that identity to produce a measurable result?**
+
+Campaign can consume:
+
+```text
+BrandContract
+approved voice
+visual rules
+canonical assets
+brand principles
+approved extensions
+```
+
+as campaign context.
+
+Brand fit becomes one campaign-evaluation dimension.
+
+It is not the entire campaign strategy.
+
+Campaign can create tactical variation within approved brand territory.
+
+If a campaign requires a new durable brand expression, Campaign should request or propose a Brand extension rather than silently redefining the brand.
+
+---
+
+# Campaign lifecycle
+
+The canonical Campaign lifecycle is:
+
+```text
+CONTEXT
+   ↓
+OBJECTIVE
+   ↓
+AUDIENCE
+   ↓
+HYPOTHESIS
+   ↓
+CONCEPTS
+   ↓
+EVALUATION
+   ↓
+PLAN
+   ↓
+APPROVAL
+   ↓
+EXECUTION
+   ↓
+MEASUREMENT
+   ↓
+LEARNING
+```
+
+The product should preserve this structure as durable state where appropriate.
+
+---
+
+# Campaign context
+
+Campaign should build an explicit picture of available evidence.
+
+Possible context sources include:
+
+```text
+brand
+products
+collections
+inventory
+price
+COGS
+margin
+historical orders
+traffic
+conversion
+audience/customer segments
+search data
+existing content
+existing media
+campaign history
+channel history
+campaign outcomes
+seasonality
+operational constraints
+budget constraints
+```
+
+Campaign does not require all of these.
+
+It must explicitly identify:
+
+```text
+available
+missing
+stale
+inferred
+estimated
+unsupported
+```
+
+context.
+
+---
+
+# Graceful degradation
+
+## Basic context
+
+```text
+brand
++ products/services
++ campaign objective
+```
+
+Campaign can:
+
+- generate campaign hypotheses;
+- create concepts;
+- propose messaging;
+- develop preliminary audience strategy;
+- identify required content;
+- suggest channels;
+- identify missing information.
+
+Confidence should remain appropriately limited.
+
+## Commercial context
+
+```text
+basic context
++ pricing
++ COGS
++ margin
++ inventory
+```
+
+Campaign can additionally:
+
+- evaluate offers;
+- evaluate discount feasibility;
+- prioritize products;
+- reason about inventory;
+- consider margin impact;
+- develop better merchandising recommendations.
+
+## Performance context
+
+```text
+commercial context
++ analytics
++ historical orders
++ customer segments
++ campaign history
++ channel outcomes
+```
+
+Campaign can additionally:
+
+- use actual behavioral evidence;
+- compare concepts against previous performance;
+- identify stronger segments;
+- detect fatigue;
+- identify winning or losing patterns;
+- make better channel decisions;
+- create more meaningful hypotheses.
+
+Missing data must never be replaced by invented evidence.
+
+---
+
+# Campaign objectives
+
+Campaign should support objectives such as:
+
+```text
+awareness
+traffic
+acquisition
+conversion
+product launch
+collection launch
+SEO / organic acquisition
+promotion
+retention
+reactivation
+cross-sell
+upsell
+average-order-value growth
+inventory movement
+email signup
+audience growth
+engagement
+education
+event promotion
+partner / creator campaign
+```
+
+The selected objective materially changes evaluation.
+
+A strong awareness campaign is not necessarily a strong direct-conversion campaign.
+
+---
+
+# Audience intelligence
+
+Campaign should support:
+
+- target audience definition;
+- existing-customer segments;
+- acquisition audiences;
+- behavioral segments;
+- lifecycle stages;
+- product affinity;
+- geographic or seasonal relevance where data permits;
+- channel-specific audiences;
+- exclusion criteria;
+- known versus inferred audience attributes.
+
+Campaign should not fabricate customer demographics when no evidence exists.
+
+---
+
+# Campaign briefs
+
+The CampaignBrief should capture:
+
+```text
+objective
+target audience
+business context
+brand context
+products / services
+campaign type
+problem / opportunity
+core hypothesis
+message
+offer
+channels
+timing
+constraints
+success metrics
+evidence
+assumptions
+unknowns
+```
+
+Initial public operation:
+
+```text
+campaign.brief.draft
+```
+
+Persisted briefs should remain distinguishable from drafts created only for conversation.
+
+---
+
+# Campaign concepts
+
+Campaign should normally generate multiple strategic directions where appropriate.
+
+A CampaignConcept may contain:
+
+```text
+name
+strategic idea
+campaign hypothesis
+audience
+message
+hook
+product / collection focus
+offer strategy
+SEO/search angle
+channel fit
+creative direction
+content requirements
+expected strengths
+risks
+evidence
+assumptions
+```
+
+Concepts are strategic alternatives.
+
+They are not merely alternative taglines.
+
+---
+
+# Campaign evaluation
+
+`campaign.concept.evaluate` should evaluate concepts against meaningful performance dimensions.
+
+Potential dimensions include:
+
+```text
+objective_fit
+audience_fit
+brand_fit
+message_strength
+search_opportunity
+offer_strength
+channel_fit
+historical_evidence
+conversion_evidence
+commercial_viability
+margin_fit
+inventory_fit
+seasonality
+creative_feasibility
+execution_cost
+measurement_quality
+evidence_quality
+```
+
+Scoring requirements:
+
+1. Every score must have supporting factors.
+
+2. The system should distinguish observed facts from inferred judgments.
+
+3. Missing evidence should affect confidence.
+
+4. Scores must not pretend to predict guaranteed outcomes.
+
+5. Users should be able to understand why one concept outranks another.
+
+---
+
+# Concept ranking
+
+`campaign.concepts.rank` should compare concepts transparently.
+
+Example output:
+
+```text
+Concept A
+Strongest conversion evidence.
+Moderate creative burden.
+Strong inventory fit.
+
+Concept B
+Strongest Brand fit.
+Weak prior channel evidence.
+Excellent organic-content opportunity.
+
+Concept C
+Highest search opportunity.
+More expensive to execute.
+Commercial assumptions remain unresolved.
+```
+
+The goal is informed decision making.
+
+Not artificial certainty.
+
+---
+
+# Promotion intelligence
+
+Campaign should treat promotion as broader than discounting.
+
+Campaign promotion may include:
+
+- launches;
+- seasonal moments;
+- limited releases;
+- bundles;
+- customer reactivation;
+- loyalty;
+- product education;
+- product-led promotion;
+- collection storytelling;
+- partnerships;
+- creator collaborations;
+- organic search;
+- email;
+- social;
+- paid media;
+- storefront promotion;
+- landing-page promotion.
+
+A valid Campaign recommendation may be:
+
+> **Do not discount this campaign.**
+
+Offer decisions should consider available evidence.
+
+---
+
+# Offer intelligence
+
+Campaign may evaluate:
+
+```text
+discount
+bundle
+gift with purchase
+free shipping
+limited availability
+early access
+exclusive access
+product assortment
+cross-sell
+upsell
+value-added offer
+no promotional incentive
+```
+
+Offer evaluation may consume pricing, cost, margin, inventory, and historical conversion data.
+
+Campaign should not mutate commerce state simply because it recommended an offer.
+
+---
+
+# SEO and search intelligence
+
+SEO is a first-class Campaign performance dimension.
+
+Campaign may reason about:
+
+- search intent;
+- topic demand;
+- product or collection search opportunity;
+- organic acquisition strategy;
+- query-to-offer fit;
+- landing-page strategy;
+- campaign topic clusters;
+- competitive positioning where authorized evidence exists;
+- search-oriented messaging;
+- internal content requirements;
+- organic versus paid interaction;
+- SEO measurement.
+
+Possible Campaign output:
+
+```text
+campaign.seo.plan
+```
+
+The SEO plan may request work from Content Studio.
+
+Campaign does not need to become the canonical metadata editor.
+
+Campaign must never guarantee rankings.
+
+---
+
+# Channel planning
+
+Campaign should evaluate available channels rather than automatically recommending every channel.
+
+Potential channels:
+
+```text
+website
+landing page
+storefront
+email
+organic social
+paid social
+search
+organic search
+creator
+affiliate
+partner
+community
+retargeting
+direct outreach
+```
+
+Channel recommendations should consider:
+
+```text
+objective
+audience
+historical performance
+available content
+cost
+measurement capability
+operational capacity
+```
+
+---
+
+# Experimentation
+
+AgentSam Campaign should be built to test hypotheses.
+
+A CampaignExperiment may include:
+
+```text
+hypothesis
+control
+variant
+changed factor
+audience
+channel
+measurement
+minimum evidence requirement
+expected observation window
+result
+learning
+```
+
+Examples:
+
+```text
+message A vs message B
+discount vs no discount
+utility angle vs lifestyle angle
+new audience vs existing audience
+search landing page A vs B
+hero product A vs B
+```
+
+Experimentation gives Campaign an improvement loop rather than making each campaign a disconnected event.
+
+---
+
+# Creative requirements
+
+Campaign owns the strategic requirement for creative.
+
+For example:
+
+```text
+1 collection hero
+3 product detail treatments
+2 acquisition social variants
+1 retention email hero
+1 search landing page
+3 short video concepts
+```
+
+Campaign may also specify:
+
+```text
+message
+audience
+format
+channel
+objective
+required brand extension
+variant hypothesis
+```
+
+Actual asset production may be delegated to:
+
+```text
+AgentSam Brand
+AgentSam Content
+media/image capabilities
+other creative plugins
+```
+
+---
+
+# Merchandising
+
+Campaign may reason about:
+
+- which products to feature;
+- assortment;
+- hero products;
+- product ordering;
+- bundles;
+- cross-sell;
+- upsell;
+- launch assortment;
+- inventory pressure;
+- campaign-specific collections.
+
+Campaign does not become the canonical catalog or manufacturing authority.
+
+Its job is the strategic merchandising decision.
+
+---
+
+# Campaign plan
+
+`campaign.plan` should transform the approved strategy into coordinated work.
+
+A CampaignPlan may contain:
+
+```text
+objective
+audience
+hypothesis
+products
+offer
+message
+SEO/search strategy
+channels
+creative requirements
+content requirements
+merchandising requirements
+timeline
+dependencies
+approvals
+measurement plan
+success metrics
+risks
+assumptions
+```
+
+Plans should be portable.
+
+They should not assume FNF, Shopify, Completeful, Cloudflare, or any specific provider.
+
+---
+
+# Execution boundary
+
+Campaign planning and campaign execution are separate capabilities.
+
+Initial Campaign should prioritize:
+
+```text
+read
+understand
+draft
+evaluate
+rank
+plan
+save
+```
+
+Later:
+
+```text
+prepare
+launch
+measure
+review
+optimize
+```
+
+Actions that mutate external systems must require:
+
+- appropriate scope;
+- explicit authorization;
+- approval where required;
+- idempotency;
+- receipts;
+- structured failure behavior;
+- post-write verification where possible.
+
+---
+
+# Measurement
+
+Campaign must eventually connect planning to results.
+
+Measurement may include:
+
+```text
+impressions
+traffic
+clicks
+engagement
+conversion
+orders
+revenue
+AOV
+margin
+new customers
+returning customers
+email actions
+search performance
+channel cost
+acquisition cost
+inventory movement
+```
+
+Not every campaign needs every metric.
+
+The relevant metrics derive from the objective.
+
+---
+
+# Campaign outcomes
+
+A CampaignOutcome should distinguish:
+
+```text
+planned objective
+planned hypothesis
+observed data
+attribution confidence
+unexpected outcomes
+successful elements
+unsuccessful elements
+unresolved questions
+```
+
+Observed facts should not be rewritten as model interpretation.
+
+Both should be preserved separately.
+
+---
+
+# Campaign learnings
+
+`campaign.review` / `campaign.learnings` should produce reusable knowledge.
+
+Examples:
+
+```text
+Utility-focused messaging outperformed lifestyle messaging
+for first-time visitors.
+
+Discounting increased conversion but materially reduced
+contribution margin.
+
+Organic search produced slower initial traffic but higher
+purchase intent.
+
+Existing customers responded more strongly to collection
+storytelling than urgency.
+```
+
+Learnings must retain provenance and confidence.
+
+They can inform future campaigns.
+
+They should not automatically become permanent Brand rules.
+
+---
+
+# Performance loop
+
+The defining system loop is:
+
+```text
+campaign history
+      ↓
+new objective
+      ↓
+context
+      ↓
+hypothesis
+      ↓
+concepts
+      ↓
+evaluation
+      ↓
+campaign
+      ↓
+measurement
+      ↓
+outcome
+      ↓
+learning
+      ↓
+future campaign
+```
+
+AgentSam Campaign should become better grounded as a workspace accumulates actual evidence.
+
+---
+
+# Initial public MCP surface
+
+Keep the public surface intentionally compact.
+
+Existing:
+
+```text
+campaign.get_context
+campaign.brief.draft
+campaign.concept.evaluate
+campaign.concepts.rank
+campaign.plan
+campaign.brief.save
+campaign.concept.save
+```
+
+Recommended near-term additions should be driven by proven user workflows rather than namespace completeness.
+
+Likely internal/package capabilities include:
+
+```text
+campaign.audience.evaluate
+campaign.offer.evaluate
+campaign.seo.plan
+campaign.channel.plan
+campaign.experiment.plan
+campaign.performance.evaluate
+campaign.review
+campaign.learnings
+```
+
+Not every package function needs to become a public MCP tool.
+
+The public tools should remain recognizable user-level operations.
+
+---
+
+# Skills
+
+Recommended Campaign plugin skills:
+
+```text
+campaign-strategy
+campaign-brief
+campaign-concept-development
+campaign-concept-review
+promotion-planning
+seo-campaign
+offer-review
+channel-planning
+campaign-experiment
+campaign-performance-review
+```
+
+Each skill should define:
+
+- activation conditions;
+- minimum context;
+- preferred evidence;
+- tool sequence;
+- missing-data behavior;
+- approval behavior;
+- expected output;
+- direct examples;
+- indirect examples;
+- negative/out-of-scope examples.
+
+---
+
+# UI
+
+Campaign should work without custom UI.
+
+UI should be reserved for places where structured visual comparison materially improves the experience.
+
+Recommended early surfaces:
+
+## Campaign concept card
+
+```text
+SPRING RETURN
+
+Objective
+Reactivation
+
+Audience
+Prior customers inactive 90+ days
+
+Message
+Back on the water.
+
+Offer
+No discount recommended
+
+Primary channel
+Email
+
+Supporting channels
+Organic social
+
+Evidence
+Historical customer activity
+Seasonality
+Inventory position
+
+Confidence
+Medium-high
+```
+
+## Concept comparison
+
+Compare:
+
+- strategy;
+- audience;
+- objective fit;
+- Brand fit;
+- performance evidence;
+- offer;
+- channels;
+- risks;
+- confidence.
+
+## Performance review
+
+Show:
+
+```text
+objective
+expected result
+actual result
+evidence
+variance
+learning
+next action
+```
+
+No giant marketing dashboard is required inside ChatGPT.
+
+---
+
+# Authorization
+
+Campaign should use narrow scopes.
+
+Conceptually:
+
+```text
+campaign:read
+campaign:plan
+campaign:write
+campaign:execute
+campaign:measure
+
+brand:contract:read
+commerce:read
+analytics:read
+content:read
+```
+
+No:
+
+```text
+agentsam:everything
+```
+
+Campaign access must never imply access to InnerAnimalMedia's internal control plane.
+
+---
+
+# Errors
+
+Important Campaign error families include:
+
+```text
+missing_context
+insufficient_evidence
+brand_context_unavailable
+analytics_unavailable
+commerce_context_unavailable
+campaign_history_unavailable
+scope_denied
+authorization_required
+approval_required
+unsupported_action
+stale_evidence
+execution_failed
+measurement_unavailable
+attribution_uncertain
+```
+
+Errors should preserve partial usefulness where possible.
+
+Example:
+
+> Analytics are unavailable, so I can still build the campaign plan, but performance-based concept ranking will have lower confidence.
+
+---
+
+# Acceptance criteria
+
+A Campaign release is not successful merely because the MCP endpoint returns tools.
+
+It should pass behavioral cases.
+
+## Direct
+
+> Plan a launch campaign for our fall collection.
+
+Campaign activates and builds an evidence-backed strategy.
+
+## Indirect
+
+> Sales usually slow down around this time. What should we do?
+
+Campaign recognizes the performance/campaign problem without requiring the user to say “campaign.”
+
+## Brand composition
+
+> Make this campaign feel more aggressive without losing who we are.
+
+Campaign consumes Brand context and varies campaign expression without silently editing canonical Brand state.
+
+## Promotion
+
+> Should we run 20% off?
+
+Campaign evaluates the offer rather than assuming discounts are desirable.
+
+## SEO
+
+> We want more people finding this collection through search.
+
+Campaign develops an organic/search strategy with explicit evidence and no ranking guarantees.
+
+## Sparse context
+
+Only objective + products + basic Brand context exist.
+
+Campaign still produces a useful plan with explicit assumptions and lower confidence.
+
+## Rich context
+
+Campaign has Brand, catalog, inventory, margin, customer, analytics, and history.
+
+Recommendations demonstrably use the richer evidence.
+
+## Performance review
+
+> Why did the last campaign underperform?
+
+Campaign compares objective, hypothesis, execution, and observed outcomes without inventing causation.
+
+## Out of scope
+
+> Guarantee this campaign will double sales.
+
+Campaign does not promise an outcome.
+
+## Authorization
+
+> Launch this campaign.
+
+With planning-only permission, Campaign does not execute the mutation.
+
+---
+
+# Definition of v1
+
+AgentSam Campaign reaches v1 when:
+
+1. It functions as an independent plugin.
+
+2. Its domain is clearly performance/campaign intelligence rather than Brand identity or generic commerce.
+
+3. CampaignContext, CampaignBrief, CampaignConcept, CampaignPlan, CampaignOutcome, and CampaignLearning have stable contracts.
+
+4. It composes AgentSam Brand without reimplementing Brand.
+
+5. It can operate without AgentSam Brand and explicitly reports the resulting context limitation.
+
+6. Concepts and scores expose supporting evidence.
+
+7. Campaign planning degrades gracefully with missing inputs.
+
+8. Campaign recommendations remain provider-neutral.
+
+9. Direct, indirect, sparse-context, rich-context, authorization, error, and out-of-scope acceptance suites pass.
+
+10. No FNF-specific assumptions are required.
+
+11. Public MCP exposes only approved Campaign capabilities.
+
+12. Internal InnerAnimalMedia authority is never inherited through the public plugin surface.
+
+13. Actual campaign outcomes can be captured separately from model interpretation.
+
+14. Learnings can inform subsequent campaigns.
+
+---
+
+# North-star test
+
+The north-star test is not:
+
+> Can AgentSam write an ad?
+
+It is:
+
+> **Can AgentSam use this organization's real identity, products, audiences, economics, channels, historical behavior, and campaign outcomes to make better campaign decisions—and then learn from the results?**
+
+That is AgentSam Campaign.