Company identification for this article; no endorsement implied. OpenAI logo belongs to OpenAI.I use the same AI ecosystem in two businesses that could not look more different if they tried.
In one corner, I am building Shopping Guru PhD: products, prices, referral rules, WordPress, databases, TikTok videos, website design, and the occasional cache that insists the old version of a page is still having its moment.
In the other, I am developing curriculum and tutoring resources for Lifetime Tutors: lesson goals, standards, explanations, practice, teacher guidance, student-facing materials, and the very important question of whether an activity is actually useful for a learner.
ChatGPT and Codex help me with both. They do not run either business for me.
I make the business decisions. I make the editorial conclusions. I decide what belongs in a lesson, whether a product opinion reflects my real experience, and whether technical work is ready for production. AI helps me think, organize, build, test, and document at a scale that would be much harder to manage alone.
AI assists my work. It does not inherit my judgment, my experience, or my responsibility.
I Use AI in Two Completely Different Businesses
My background makes this crossover feel less strange than it might sound. I have a PhD in Engineering and Computing Education, an MS in Software Engineering, and years of work across education, curriculum, research, and technology.
Apparently those interests were never going to stay politely separated.
For Shopping Guru, I move between consumer research and technical systems. A product story may begin with something I bought, a TikTok I made, or a price I noticed. It may eventually involve editorial planning, affiliate-link verification, a database record, a WordPress article, responsive design, and production QA.
For Lifetime Tutors, the subject changes but the need for structure does not. A lesson can begin with a standard, a learner need, or an instructional goal. It may grow into explanations, guided practice, a student-facing digital lesson, teacher guidance, printables, game-question banks, and a validation checklist.
The content is different. The underlying habit is familiar: establish the source, define the goal, build the right thing, and verify what actually exists at the end.
What ChatGPT Does for Me
In my workflow, ChatGPT is where I do much of the thinking before implementation.
I use it to organize complicated requirements, compare possible approaches, develop article or lesson structure, refine language, identify missing questions, and turn a very large collection of ideas into a directive another person—or Codex—can actually follow.
That includes work such as:
- planning Shopping Guru articles and content hubs;
- working through product, merchant, referral, and disclosure questions;
- developing comparison methods and SEO direction;
- planning curriculum sequences, examples, practice, and teacher guidance;
- refining explanations for different learners; and
- reviewing technical reports and deciding what should happen next.
This is how I divide the work. It is not a declaration that ChatGPT can only plan or that Codex can only implement. The products overlap. My distinction is practical: ChatGPT is often where I shape the problem; Codex is often where the problem meets files, code, tests, and a real environment.
What Codex Does for Me
Codex changed the workflow because it can work inside the project instead of merely describing what a finished solution might look like.
For Shopping Guru, that has meant inspecting repositories, finding the authoritative template, editing WordPress theme code, working with the Finds product architecture, reconciling database-backed Products and Offers, testing responsive layouts, creating rollback checkpoints, verifying file hashes, deploying bounded changes, checking public canonical pages, and documenting the final state.
For curriculum work, it has meant building structured lesson packages, producing single-file HTML lessons and supporting resources, maintaining student/teacher boundaries, generating validation reports, checking links and file inventories, enforcing project frameworks, and packaging approved deliverables.
Current OpenAI documentation describes Codex as able to work with repository environments, run tasks in parallel, and return summaries and diffs for review. OpenAI also documents code-review workflows that inspect selected changes without automatically changing the working tree. Those capabilities match the part of my process I value most: implementation can be delegated, but the result remains inspectable. You can read the current official overviews for Codex cloud and code review.
Codex is not a magic “make website” button. It is much more useful when the task has an exact target, defined authority, constraints, sources, tests, and a clear finish line.
Business #1: Shopping Guru PhD
Shopping Guru is where consumer experience and technical operations keep introducing themselves to each other.
ChatGPT helps me develop the editorial and governance layer. I can work through what a reader needs to know, which claims are supported, how an affiliate relationship differs from a personally used service, what must remain private, and what evidence is still missing.
Codex works closer to the implementation layer. It can inspect the current Main and Finds systems, identify the existing architecture, update the appropriate source rather than creating a duplicate, run tests, and show me what changed.
That matters because Shopping Guru is not only a collection of articles. The public system includes WordPress editorial content, a custom Finds experience, product and offer records, media, SEO, disclosures, consent behavior, cache layers, and governance that prevents a tracked link from quietly becoming public before it is verified.
Real Example: From TikTok Review to a Live Product Page
The EPOMAKER Glyph keyboard is a good example of how the roles fit together.
I supplied the real-world ingredients: my TikTok review, the product context, merchant links, and additional evidence as questions appeared.
ChatGPT helped organize the requirements and separate facts that are easy to blur together. What I paid is a historical purchase fact. A marketplace price is a current, changeable observation. An affiliate link needs its own authority. Two links that look similar may lead to different sellers, variants, warehouses, or fulfillment choices.
Codex then worked through the technical reconciliation: matching the product identity, inspecting the existing Product and Offer structure, checking the public destinations, updating the proper records, implementing the review presentation, running responsive and canonical-page QA, and documenting what was published or withheld.
I stayed responsible for the choices. When an affiliate relationship changed, when purchase evidence was incomplete, or when a merchant image lacked confirmed reuse rights, the correct answer was not to make the article look finished anyway. It was to publish the useful, supported content and withhold the unsupported commerce element.
That is the part of AI-assisted work I wish received more attention: sometimes progress is not “generate more.” Sometimes progress is knowing exactly what not to manufacture.
Business #2: Lifetime Tutors
Education uses many of the same tools, but the responsibility feels different because the work ultimately serves learners.
ChatGPT can help me reason through a lesson: how to sequence an explanation, which examples may clarify a concept, how to create supported practice, how to adapt language, and where a learner may need a different route into the material.
Codex can turn an approved instructional plan into an actual package: a responsive HTML lesson, teacher guidance, practice resources, printable materials, game-question files, standards metadata, file manifests, and automated checks.
Lifetime Tutors records include UFLI-based phonics work, coding lessons, Spanish and ESL resources, and structured supporting materials. Those projects are not evidence that AI decided what to teach. They are evidence that a carefully defined instructional plan can be turned into a more complete, testable deliverable.
Real Example: From Instructional Goal to Lesson Package
Consider a phonics lesson built from an established instructional source.
I begin with the instructional purpose: what skill the lesson addresses, what the learner should practice, how the session should flow, and what source material or standard controls the content.
ChatGPT can help me examine the sequence, refine an explanation, propose examples, and think through guided practice or differentiation. Those suggestions are working material—not automatic curriculum authority.
Codex can then assemble the technical package: the student-facing lesson, teacher guidance, linked resources, practice or game files when they are instructionally appropriate, and checks that confirm required files exist and private teacher material has not wandered into a public student folder.
I review the instructional result. I decide whether the examples are accurate, whether the pacing is realistic, whether the activity supports the goal, and whether the material is appropriate for the learner.
A beautifully functioning lesson can still teach the wrong thing. Passing a technical test does not pass the pedagogical test.
AI Does Not Replace the Teacher—or the Reviewer
AI can help me build faster, explore more variations, and keep large projects organized. It does not decide what an individual student needs, whether a lesson is appropriate, whether a learner mastered a skill, or how I should respond professionally.
It also does not decide whether I like a keyboard, handbag, beauty product, bank, host, or shopping service.
My product opinions must come from my experience. AI can help organize evidence, compare current prices, reconcile merchant links, structure the story, and present the result. It cannot retroactively use a product on my behalf.
The same principle protects both businesses: assistance can be powerful without becoming authority.
Why Project Context Matters
Neither Shopping Guru nor Lifetime Tutors fits inside one conversation.
OpenAI’s current Projects documentation describes projects as spaces that keep related chats, files, instructions, and sources together. Project instructions can apply across chats, and local projects can connect work to folders on a computer. That is useful for ongoing work where the same standards, files, terminology, and decisions keep returning. See the official Projects and chats guide.
In practice, continuity saves me from starting every request with the history of civilization, WordPress, and rabbit ownership.
But project context is not permission to stop verifying. A historical instruction can become obsolete. A current production page can differ from an old report. A retailer link can change. A standard can be misidentified. Durable context is valuable precisely because it makes conflicts visible—not because everything stored there remains true forever.
The Guardrails I Built Around AI
I did not begin with an elaborate governance system. It grew because the projects grew.
For Shopping Guru, a production directive now identifies the authoritative environment, exact scope, prohibited actions, rollback requirements, QA, cache verification, and completion evidence. Before a live change, the affected state is captured and hashed. Afterward, I want evidence from the real anonymous public page—not only a local file that looks correct.
For curriculum work, the controls include authoritative source material, standards verification, student/privacy boundaries, teacher and student separation, accessibility, file inventories, validation, and explicit publication gates.
A small sanitized version of the way I now structure a request looks like this:
Objective: update the approved page. Authority: production is current. Constraints: preserve unrelated content and private data. Before writing: create a verified rollback. QA: test the real page at mobile, tablet, and desktop. Complete only when the canonical result is verified.
That may not be the world’s most romantic prompt. It is much more useful than “make it better.”
OpenAI also documents reusable project instructions and tool connections through its current plugin architecture. I treat those capabilities the same way I treat any operational access: the tool should have the minimum authority it needs, a clear job, and an output I can inspect.
Where AI Gets Things Wrong
AI can lose part of a long instruction. It can follow an older directive after a newer decision has replaced it. It can make a plausible assumption sound like an established fact. It can stop when evidence is unavailable—or, without strong boundaries, try to fill the gap.
I have also learned that one successful technical signal rarely proves the whole outcome:
- An HTTP 200 response does not prove an affiliate link preserved attribution.
- A matching repository hash does not prove the public cache is serving the new page.
- A generated lesson file does not prove the lesson is instructionally sound.
- A polished paragraph does not prove the experience happened.
These are not arguments against using AI. They are arguments for using it with evidence, checkpoints, and review.
How My Prompts Changed
My earlier prompts were often requests. My current prompts are closer to operating instructions.
I usually define the objective, the source of truth, the facts already approved, the actions that are allowed, the actions that are prohibited, the required QA, and the exact condition for calling the work complete.
That structure helps ChatGPT reason about the real problem and helps Codex act within a bounded implementation task. It also gives me something concrete to audit when part of a request is missed.
Good prompting, for me, is not about finding one magical sentence. It is about making the decision system visible.
What I Still Do Myself
I still decide whether a Shopping Guru article sounds like me. I still supply the firsthand experience. I still decide whether a product belongs on the site and whether a referral relationship should be used.
I still determine the instructional goal, evaluate the pedagogy, respond to learners, and protect student privacy.
I review the work, correct the direction, provide missing evidence, authorize production, and own the outcome.
The desktop experience makes that review loop easier because current OpenAI documentation describes a workspace where projects, files, tools, and long-running work can stay visible together. That is helpful. It is not the same as transferring responsibility to the workspace. The current official overview is available in the ChatGPT desktop app guide.
Would I Run My Businesses Without It Now?
I could. I would also spend much more time moving between planning, drafting, implementation, checking, and documentation—and I would complete fewer ambitious projects at once.
ChatGPT helps me turn complicated thinking into an organized plan. Codex helps me turn an approved plan into files, systems, tests, and reviewable technical work.
The value is not that either tool replaces what I know. The value is that they let me apply what I know across more of the work.
One business asks whether a keyboard, cashback service, or hosting setup is genuinely useful. The other asks how to help a learner understand a sound, an equation, a story, or a line of code.
Very different businesses. Same rule:
AI can help me do the work. I still have to know what good work looks like.
Transparency: This article is not sponsored by OpenAI and contains no ChatGPT or Codex affiliate links. Product capabilities can change; the linked official OpenAI documentation is the current source for feature details.
Editorial promise: Recommendations remain independent. Affiliate links may earn Shopping Guru PhD a commission at no added cost to you.