IPVES/ Product

Case study 03 · Product operating model

8 min readUpdated July 14, 2026

Agentic product
management.

A practical operating model for coordinating human judgment, ChatGPT product framing, and Codex delivery across a live multi-site portfolio.

System
IPVES multi-site portfolio
Roles
Human · ChatGPT · Codex
Focus
Outcome through verified delivery
Evidence
Public repositories and production sites

Operating principle

Separate the roles. Reconnect them through evidence and explicit decisions.

Frame
Outcome, scope, and tradeoffs
Execute
Repository, tests, and deployment
Own
Human approval and accountability

01

The operating model

Executive summary

AI-assisted delivery works when responsibilities are designed as deliberately as the product itself. The value does not come from asking one system to do everything. It comes from assigning framing, execution, and judgment to the participant best equipped to own each decision.

Across the IPVES portfolio, ChatGPT shaped product outcomes and implementation briefs; Codex inspected repositories, changed code, tested, deployed, and verified production; the human product leader selected priorities, approved scope, judged quality, and remained accountable for what went live.

3Distinct decision roles
1Evidence-led delivery loop
4+Coordinated portfolio surfaces

02

Coordination is the product

The operating problem

A multi-site portfolio turns small inconsistencies into operating drag.

IPVES spans a portfolio home and focused Travel, Labs, Product, and Events experiences. Each surface has its own routes, assets, publishing behavior, and user expectations, while the portfolio still needs to feel coherent.

The challenge was not simply producing pages. It was maintaining navigation, custom-domain paths, privacy, service-worker isolation, responsive behavior, repository history, and reliable deployment across repeated releases without losing the intent behind each product decision.

03

Why ad hoc prompting fails

01

Intent without a contract

A broad request leaves routes, hosting assumptions, validation, and completion criteria implicit.

02

Execution without context

Code changes drift when the active repository, user edits, and production architecture are not inspected first.

03

Output without evidence

A plausible implementation is not a shipped outcome until the live route, assets, and interaction state are verified.

04

Automation without judgment

Passing tests cannot decide whether the result communicates the right product story or deserves to launch.

04

Clear ownership, shared evidence

Human–AI role design

05

Make the work executable

ChatGPT responsibilities

Product framingOutcome · user value · scope · sequence · acceptance
  • Synthesize contextWhat matters now
  • Clarify the objectiveOutcome over activity
  • Structure the briefRequirements + exclusions
  • Expose tradeoffsChoices for human judgment
  • Define completionProduction evidence required
Decision support, not decision theater.

The brief should reduce ambiguity for execution while keeping consequential choices visible to the human product leader.

06

Turn the brief into evidence

Codex responsibilities

Codex reports what the repository and production environment prove, including failures and incomplete checks.

07

Human product-leader responsibilities

  1. 01

    Choose priorities

    Decide which portfolio outcome deserves attention and why now.

  2. 02

    Approve scope

    Set the boundary for design, content, repository access, and publication.

  3. 03

    Make tradeoffs

    Balance speed, quality, consistency, risk, and future maintenance.

  4. 04

    Protect information

    Keep personal, customer, credential, and confidential material outside the public artifact.

  5. 05

    Review the live experience

    Judge communication, usefulness, and portfolio coherence beyond automated checks.

  6. 06

    Remain accountable

    AI can recommend and execute; launch responsibility remains human.

08

Eight steps, one loop

Task decomposition model

Each handoff produces a concrete artifact: objective, brief, repository evidence, commit, production result, review decision, correction, or reusable standard.

09

Architecture before edits

Repository-first workflow

The repository is a source of product truth

Hosting mode, direct routes, service-worker scope, navigation contracts, and test commands should be discovered—not reconstructed from memory.

10

Guardrails and approval points

DecisionAI may prepareHuman retains
Public scopeSanitized framing and exclusion checksApproval of what may be published
Repository changeDiff, tests, and impact analysisAuthority for material scope expansion
DeploymentCommit, push, workflow monitoring, verificationExplicit instruction to publish
External actionProposed message or configurationApproval before consequential transmission
Quality exceptionEvidence, alternatives, and riskDecision to accept, revise, or stop

11

Evidence in layers

Quality gates

Completion rule

A task is complete only when the intended production outcome is observable.

Local correctness, deployment success, and user-visible behavior are separate gates.

Repository

Code quality

  • Clean intended diff
  • Lint and type checks
  • Static production build
  • Rendered-route assertions
Architecture

System integrity

  • Correct root paths
  • Direct route refresh
  • Manifest and service worker
  • Confidentiality scan
Experience

Live quality

  • Desktop and 375px
  • No horizontal overflow
  • Keyboard focus
  • Console and mixed content

12

Deployment and rollback strategy

Clean main→Exact commit→Matching workflow→Verified production

A successful workflow is necessary evidence, not sufficient evidence.

Guarded deploy commands rerun checks, push the intended commit, wait for its matching GitHub Pages workflow, and then verify the custom domain. This prevents a green run for the wrong commit from becoming false confidence.

Rollback readiness comes from focused commits, preserved working trees, known-good routes, and historical fallback surfaces. When a check fails, the first move is diagnosis: distinguish code defects, verifier defects, edge propagation, and stale service-worker state before changing production again.

13

Spend context on decisions

Token-efficiency practices

Context budgetRetrieve once · summarize precisely · verify in batches
  • Portfolio inventoryFind the canonical repository immediately
  • Repository guidanceReuse routes, commands, and hosting rules
  • Consolidated checksOne command covers lint, types, build, and tests
  • Guarded deploymentPush, monitor, and verify as one contract
  • Batched browser readsMeasure several quality signals per page
  • Narrow follow-upsCorrect evidence, not the whole design
Efficiency is quality infrastructure.

Saved tokens matter when they reduce repeated discovery and noisy output while preserving the evidence needed for sound decisions.

14

From pages to an operating system

What changed over time

Stage 01

Individual builds

Each experience was implemented and published as a discrete project.

Learning: delivery works
Stage 02

Portfolio architecture

Custom domains, navigation, and canonical repositories created a coherent system.

Learning: boundaries matter
Stage 03

Shared contracts

Check, deploy, and production-verification commands became consistent.

Learning: repetition can compound
Stage 04

Evidence-led operations

Inventory, browser QA, route checks, and narrow corrections became reusable practice.

Learning: verification is product work

15

Failure modes and corrections

  1. 01 · Wrong contextUse the inventory and repository contract before touching files.
  2. 02 · Over-broad changeTranslate visual feedback into a narrow defect and preserve unrelated work.
  3. 03 · Path failureVerify custom-domain root paths rather than assuming a project-site prefix.
  4. 04 · False negativeTest the verifier itself when production evidence contradicts its marker.
  5. 05 · Encoding defectPrefer stable SVGs or entities and assert against replacement characters.
  6. 06 · Automation overreachKeep approval points explicit and report unresolved evidence honestly.

16

A portfolio that can keep moving

Results

Operating outcome

Repeatable product delivery across distinct public experiences.

The portfolio gained shared release discipline without forcing every site into the same visual or technical implementation.

Portfolio

Coherent surfaces

  • ipves.io portfolio home
  • Companion Travel
  • Companion Labs
  • Companion Product
  • Companion Events
Delivery

Reusable contracts

  • Canonical repository inventory
  • Standard check and deploy commands
  • Direct-route verification
  • Exact workflow monitoring
Experience

Public evidence

  • Travel companions
  • VaporTrail
  • Wedding Seating
  • Three Product case studies

17

Reusable operating principles

  1. 01

    Define the outcome before the task

    A precise finish line improves both autonomy and quality.

  2. 02

    Inspect before proposing

    Current repository and production evidence outrank remembered assumptions.

  3. 03

    Separate framing from execution

    Different collaborators can specialize without fragmenting accountability.

  4. 04

    Automate repeated evidence

    Standard checks preserve attention for product decisions.

  5. 05

    Make corrections smaller than discoveries

    A broad investigation should still produce the narrowest safe fix.

  6. 06

    Close the loop in production

    The live user experience is the final source of truth.

18

Product leadership lessons

01AI collaboration becomes an operating model when roles, artifacts, and approval points are explicit.

02The human product leader can delegate execution without delegating accountability.

03Repository discipline and production verification are product-management capabilities, not merely engineering hygiene.

04Token efficiency improves when context is retrieved once, encoded into reusable contracts, and tested in batches.

05The strongest AI-assisted teams learn from failure evidence and carry corrected patterns into the next release.

Product case studies

Strategy made
operational.

Return to Product index