AI as “super Google”
At first, it is a question and an answer. Find something, explain something, move on. Useful—but only a small part of what the tool can help you do.
MY AI JOURNEY / A METHOD FOR REAL PROJECTS
From asking AI questions
to organizing the work.
Two years of learning, building, and finding out what holds up in a product-development project.
CONTEXT. ROLES. PROCEDURES. FOLLOW-THROUGH.MIKE MAIER / ADVANCE NOW
HOW I GOT HERE
The tools changed.
The project got bigger.
Talk it through and challenge it
The next step is a conversation. Debate the answer. Ask whether the assumptions make sense. Check that you and the AI are still talking about the same problem. Follow the question beyond the first plausible response.
Give the project a workspace
I moved into coding and agent workspaces, including Claude Code, Codex, Grok Build, and environments such as VS Code. The important change was not just writing code. I started doing engineering projects there, with files the AI could search, examine, and cross-reference as the work required.
Break the work into useful pieces
Large projects exposed the limits of what an AI could keep in its immediate working context. I learned to divide the work into bounded requests, give each one the information it needed, and bring the results back together.
Build continuity around defined roles
My approach evolved toward ongoing specialist roles supported by project records, clear responsibilities, and written instructions. Temporary helpers take on contained work. The continuing role reviews what comes back and keeps the important decisions with the project.
AN ENGINEERING WORKSPACE
More than a chat.
More than coding.
The turning point was using an AI workspace to carry out the engineering project itself.
A project can involve requirements, research, calculations, supplier information, drawings, test notes, decisions, and correspondence. Those documents need to be related to one another, not treated as isolated attachments.
In my work, file-based environments made it practical to organize collections of hundreds of documents and let an agent look for relevant material within the access it was given. It could check a source, compare a requirement, or follow a reference as part of a task.
That does not mean every document fits in the AI’s working context at once. A useful workspace makes information findable; a useful task tells the agent what to look for and what to produce.
The value is a way of working: keep the project’s evidence and decisions organized, and bring the right pieces into each request.
THE MEMORY LESSON
Access to information
is not the same as remembering it.
A long conversation can still lose an important detail.
I learned that giving one agent more and more large tasks could weaken continuity. It might record a summary while leaving out a detail that mattered later. Having a lot of files available did not automatically solve that problem.
AI tools continue to change. In my experience, delegation and continuity have improved, but I still design the work around clear boundaries and reliable records rather than assuming the tool will remember everything.
Requirements, decisions, sources, open issues, and completion criteria belong in the project record. The next request should be able to recover what it needs from that record.
When I say “persistent,” I mean continuity of role and maintained project context—not unlimited or infallible model memory.
THE WAY I ORGANIZE IT
Continuing roles.
Temporary help.
THE CONTINUING SPECIALIST
Own a defined part of the work.
An ongoing role has a clear responsibility, a bounded area of knowledge, and maintained project records. It knows where to find the current requirements and decisions, and what it is supposed to hand back.
Depending on the project, I may organize several such roles instead of asking one general agent to carry every major task.
THE “INTERN” AGENT
Do one contained job.
A temporary helper receives a full task brief, the relevant material, instructions, and an expected output. It can devote its available context to that job without being responsible for the project’s long-term continuity.
When the assignment is complete, the helper’s session can end. Its useful output is returned, reviewed, and saved.
THE HANDOFF
Make the assignment complete.
State the question, supply the source material, set the boundaries, and define what “done” means. Ask the helper to identify uncertainty and preserve the source references needed to check its work.
THE RETURN
Keep the project’s knowledge.
The continuing role evaluates the result, resolves or records open questions, and updates the project record. A finished response is not automatically an accepted decision.
Delegation only helps if someone brings the useful work back into the project.
ONE PRACTICAL EXAMPLE
A supplier comparison,
with a clear handoff.
- 01
Define the comparison
The continuing project role identifies the requirement, the candidate supplier documents, and the questions to compare.
- 02
Assign the research
An intern agent extracts the relevant information into a comparison, citing sources and marking missing or conflicting details.
- 03
Review the result
The continuing role checks the work against the project requirements and prepares the questions that still need a supplier or human answer.
- 04
Record the next step
The reviewed comparison, unresolved questions, and any approved decision stay in the project record. The temporary assignment ends.
An illustrative workflow, not a claim that AN Creation currently automates supplier selection or outreach.
WRITTEN INSTRUCTIONS
I call them SOPs.
The name matters less than the job.
A role needs direction about what it does—and what it does not do.
People use different names for these instructions. I think of them as standard operating procedures: enough structure to focus the work and make the expected behavior explicit.
- Purpose: What is this role responsible for?
- Inputs: Which records and sources should it use?
- Boundaries: What must it leave alone or refer elsewhere?
- Method: What steps and checks does the task require?
- Output: What should it return, and in what form?
- Review: Who checks the result or approves the next action?
- Continuity: What needs to be written back into the project record?
THE MANUFACTURING CONNECTION
Different tools.
Familiar management problems.
Clear expectations. Useful goals. Defined responsibilities. Feedback and follow-through.
Much of this feels familiar from managing employees and improving a manufacturing operation. Give someone an unclear assignment and the handoff becomes the problem. Ask one person to carry too much and details get lost. Fail to record the decision and the next shift has to reconstruct it.
AI is not a person, but many of the coordination principles carry over. Break work into sensible pieces, supply the information, explain the expected result, and review what comes back.
The same thinking connects process improvement on the floor with product-development improvement: understand the work, reduce confusion, make important information available, and establish a repeatable way to move forward.
WHY AN CREATION EXISTS
Bring that structure
to someone with an idea.
This journey is the thinking behind AN Creation.
I’ve focused this work on product development and product realization: helping a person investigate an idea, organize the questions, develop the plan, and understand what comes next.
AN Creation’s current conversations, named roles, shared documents, and next steps are part of that direction. The wider method described here reflects my practice and the approach I’m developing; it is not a promise that every delegation pattern is already automated in AN Creation.
I want people to benefit from that experience through a largely free resource. For those who need my direct involvement and have the budget for it, personal consulting is available separately.
MANUFACTURING THINKING / PRACTICAL AI
The vocabulary is new.
Much of the work is familiar.
You already work with setup files, drawings, inspection data, maintenance history, and material requirements planning (MRP) records. AI gives us new ways to work with that information.
These are practical comparisons, not universal definitions. Different software uses these words differently. Start by asking what information goes in, what happens to it, and what comes out.
Artifact → a piece of work you can keep
A report, comparison, drawing, or revised document. Ask where it is saved, which revision is current, and who has reviewed it.
Context → the information needed for this job
Like handing a setup person the correct drawing, material specification, and setup sheet. An agent needs the relevant information available during the task.
Node → a step in the workflow
In a workflow graph, think of a station that reads records, compares requirements, or prepares a result. A file may be an input to that step; it is not automatically a node.
Edge → the connection to the next step
Like an arrow on a process map: send the completed comparison to review, or route it back if information is missing. It describes flow or routing rather than the whole work instruction.
Memory → retained records for later work
Think of a shift log or maintained project file. Useful facts must be saved and retrieved when needed; the existence of a record does not guarantee the agent will use it correctly.
Tools and skills → ways to do the work
A tool might search files or read a system. A skill may package instructions and supporting resources for a repeatable job. Ask what access it needs and how its output gets checked.
The node-and-edge comparison refers to workflow graphs, such as those described in LangGraph’s Graph API documentation. Other kinds of graphs can represent different things.
MANUFACTURING THINKING / PRACTICAL AI
A defined role.
A clear assignment.
Maintenance, setup, packaging, material handling: each role has its own information, responsibilities, and procedures.
That is a useful starting point for organizing AI work. Give a research role the sources it needs. Give a comparison role the criteria to compare. Specify when it should stop, ask for help, or hand work to another role.
“AI,” “agent,” “agentic,” and “bot” are overlapping labels, not a dependable sequence of generations. A bot may be a simple scripted helper; an AI agent often uses a model and tools to work through a task. The label alone tells you little about its memory, permissions, or ability to act.
I look at the package: the instructions, available information, tools, saved records, and boundaries. What is it equipped to do? What is it allowed to do? Who reviews the result?
A written procedure does not make an AI qualified in the way training and demonstrated competence qualify a person. The useful parallel is how we organize and check the work.
MANUFACTURING THINKING / PRACTICAL AI
Good management
still does useful work.
Define “done” before starting
State the goal, required output, due point, and acceptance criteria. “Look into this” is a much weaker assignment than a specific comparison with sources and open questions.
Go to the actual process
Talk with the people doing the work and examine the records they use. A polished summary cannot replace understanding what happens on the floor.
Make the handoff explicit
Name the next owner, provide the current revision, and flag unresolved issues. Work waiting between steps can matter as much as the speed of any individual step.
Standardize, then keep learning
Write down the method that works, check the result, and revise the method when evidence calls for it. An SOP is a maintained working instruction.
I see enormous potential here for manufacturing and engineering. People who already coordinate equipment, information, and teams bring a valuable foundation to AI. The opportunity is to build on that knowledge and learn the tool’s limits through real work.
See the Define–Measure–Analyze–Improve–Control loopLESS MANUAL WORK / MORE TIME TO INVESTIGATE
Automate the calculation.
Keep asking better questions.
A great deal of improvement work used to involve collecting data, building spreadsheets, doing calculations, and preparing charts. AI can help make that work much faster.
Prepare the records
An agent can help organize downtime logs, reconcile column names, and flag missing values. Preserve the source records and make cleanup decisions visible.
Run repeatable analysis
AI can help write and run code using statistical tools to calculate summaries and generate charts. Once a workflow is established, repeating it on new data can be quick.
Examine what the chart means
A Pareto chart can help focus an investigation. A control chart can help examine variation over time. Choosing the right method, sampling approach, and assumptions still requires process knowledge.
Bring the result back to the floor
Review the finding with the people who know the equipment. Compare it with operating conditions, investigate possible causes, and decide what to test next.
For example, an agent could prepare a recurring downtime report from an approved export, calculate totals with a checked script, and highlight changes for review. The team can spend more time investigating what changed and less time rebuilding the report.
Fast output is not proof of a correct conclusion. Check the calculation against known results, keep the method traceable, and distinguish a pattern from a demonstrated cause. The speed depends on the data and the task; it is not a promise that every analysis is instant.
This is where my manufacturing background and AI work come together: use automation to carry more of the effort while keeping the discipline that makes the result useful.
The improvement principles behind the approach