People spend weeks comparing models. They switch subscriptions. They chase the newest release. And the output barely changes — because the bottleneck was never the tool. It was the input.
This issue is the practical guide: what to give AI before you start, how to direct it so it does not waste your credits re-reading everything, and how to use more than one model on the same project.
Better inputs beat better models. Every time.
The Roundtable Verdict
What the models agree on
What the usual advice gets right
“Learn to prompt better” is correct in principle. Clarity matters. Specificity matters. The advice is not wrong — it is incomplete.
What it leaves out
The prompt is maybe 20% of the quality equation. The other 80% is what you provide alongside it: the source document, examples of what “good” looks like, constraints that keep the output on target, and corrections that shape the next draft. A clever prompt with no source material still produces a guess.
Where the models land
Every model — Claude, ChatGPT, Gemini, Grok, Meta — said the same thing: context beats cleverness. The more relevant material you provide, the less the model has to invent. And inventing is where it gets things wrong.
How to Direct AI So It Does Not Re-Read Everything
The most common complaint about AI is that it stalls, buffers, loops, or forgets what was decided. This is not random — it happens when the conversation gets too dense. The model has to re-read every message in the chat every time it responds. The longer the chat, the more it reads, the more credits it costs, and the worse it performs.
Here is how to prevent it:
Start with a Project, not a chat. Claude, ChatGPT, and Gemini all support Projects (or their equivalent). Upload your brand guide, product docs, style rules, and any standing instructions once. Every new chat inside that project starts already informed — no re-explaining.
Work on one section at a time. Do not ask AI to build an entire website, document, or campaign in a single chat. Break the work into sections. Complete one, approve it, then move to the next. Each section stays clean and focused.
Use handoffs between chats. When a chat gets long or the AI starts repeating itself, do not keep pushing. Ask: “Summarize everything we have decided, what is done, and what is next — in a format I can paste into a new chat.” Then open a fresh chat, paste the handoff, and say “Continue from this.” You carry the decisions. You leave the bloat behind.
Lock what is approved. Every time a section is finished, tell the AI explicitly: “This is approved. Do not change it.” Otherwise, the next edit request may trigger a full rewrite of everything — including the parts you already approved.
Send screenshots instead of pasting code. If you are working on a website or visual project, take a screenshot and upload it. The AI can “see” a screenshot faster and more accurately than re-reading 2,000 lines of code — and it uses fewer credits.
Three strikes, then restart. If the AI has not fixed a bug or a design issue by the third attempt, the chat path is corrupted. Do not keep trying. Start a fresh chat with a clean handoff. It is faster and cheaper every time.
Direct all files to the project folder. Tell the AI: “Save everything to the project folder.” This keeps your work organized and means you are not hunting through chat history to find the latest version of a file.
The Research-to-Strategy Workflow
Most people ask AI for answers. The better approach is to ask AI to build the picture first — then decide the strategy yourself. Here is the workflow we use for client work, product evaluations, and marketing plans.
Step 1: What to give the AI
Before you ask for a strategy, ask for research. Provide or ask AI to find:
Products and services. What are you offering? What does it actually do? What problem does it solve? Upload the product sheet, not your description of it.
Competitors. Who else solves this problem? How do they position themselves? What do they charge? What do their customers complain about?
Market landscape. What is happening in this space right now? Not trends to follow — conditions to understand. The goal is to see the picture clearly, not to chase what everyone else is doing.
Search questions. What are real people actually asking about this topic? What are they searching for? What are the questions behind the questions?
Issues and problems. What are the known pain points? What goes wrong? What do customers, clients, or users struggle with?
What has been tried before. What did the client (or you) do in the past? Did it work? If so, why? If not, why not? What was the result? This is critical context that most people skip — and then the AI recommends the same thing that already failed.
Existing assets. What content, data, or materials already exist? Do not start from zero if you do not have to.
Step 2: What to do with the response
Now the AI has the picture. Here is the sequence:
Determine the findings. Ask the AI: “Based on everything you have gathered, what are the key findings? What patterns do you see? What stands out? What is missing?”
Outline the direction. Ask: “Based on these findings, what approach would you recommend and why?” Let it lay out options — not just one answer.
Create the strategy with a core goal. Ask: “Build a strategy around [your chosen direction]. The core goal is [measurable outcome]. Include the reasoning, the steps, the timeline, and what success looks like.”
Build the plan. Once the strategy is approved, ask for the execution plan: “Break this into specific tasks, deliverables, and deadlines.”
Step 3: The part most people skip
Read the response and decide if you agree. This is the step that separates people who get value from AI from people who get filler. Do not accept the first draft. Read it carefully. Push back. Ask “What are you assuming here?” Ask “What could go wrong with this approach?” Ask for alternatives. Ask for more detail on the parts that matter and less on the parts that do not.
Watch for substitutions and tone drift. AI will quietly swap your words for its preferred vocabulary. It will turn a social post into a brochure. It will turn a marketing piece into therapy speak. It will add filler phrases you never use. Read every line — not for typos, but for accuracy and intent. If it changed a word, ask yourself whether the new word means the same thing. Often it does not.
Ask it to check its own work before you accept it. Before you take the final output, tell the AI: “Review what you just wrote. Check for accuracy, consistency, and anything that contradicts the source material or the brief. Flag anything you are uncertain about.” AI will catch its own errors surprisingly often when you ask it to look — it just will not do it on its own. Make this the last step before you accept a deliverable.
Then ask it to play devil’s advocate. Once it has checked for errors, push harder: “Now tear this apart. What could go wrong with this plan? What is lacking? What assumptions are we making that might not hold? Where are the holes?” AI is remarkably good at attacking its own work when you give it permission. It will find weak points it glossed over in the original. Then ask: “How would you fix each of these?” You now have a stronger strategy than 90% of what gets shipped — because you stress-tested it before it left the chat.
Save every version. Before you ask for revisions, save the current output. AI does not have an undo button. If the next draft is worse — and sometimes it is — you need a version to go back to. Save as you go. Name the files. Keep the one that worked so you never lose ground.
Unlikely Liaisons: Claude + ChatGPT Coding Together
Most people pick one AI and stick with it. Here is something worth trying: using two models on the same project, each doing what it does best. This is not theoretical — we built the EmpowerTools website, this newsletter, and several production tools this way.
Claude for architecture and content. Claude excels at long-form structure, editorial voice, holding complex context, and maintaining consistency across a large project. Use Claude to design the sections, write the content, define the rules, and build the logic. When you need a document, a strategy, or a content plan — start here.
ChatGPT for visual design and rapid iteration. ChatGPT excels at CSS implementation, layout design, and fast visual prototyping. Once Claude has defined what needs to be built, hand the structure to ChatGPT with clear instructions: “Here is the content and the structure. Build it with these colors, this layout, and this styling.” ChatGPT iterates on visual details quickly.
For automation and coding workflows. ChatGPT is particularly strong at automation. In our own projects, ChatGPT was the one that set up the automation systems on the sites, solved integration issues, and got the moving parts connected. Use Claude for planning the logic, writing the pseudocode, and thinking through edge cases — then hand the spec to ChatGPT for implementation, testing, and automation setup.
When something breaks, send it to the other model. If ChatGPT cannot solve a bug, bring the error to Claude for diagnosis — it is often better at reading an error message and identifying the root cause. Then send the fix back to ChatGPT for implementation. This works in reverse too. If one model is stuck on a file or cannot find the problem, send that file to the other model for an extra pair of eyes. They find it. A fresh model looking at the same code with no prior assumptions will spot what the other one missed.
The key to multi-model work is the same as the key to multi-chat work: a clean handoff. When you move from Claude to ChatGPT (or vice versa), provide:
What has been decided (rules, structure, style, constraints)
What has been built (attach the current files or screenshots)
What is next (the specific task for this model)
What must not change (the parts that are locked and approved)
Neither model replaces the other. They complement. And the human in the middle — directing, deciding, approving — is the reason it works.
We did some builds — the three of us together — that separately none of us could have done on our own. Claude and ChatGPT needed me for some things (I am a web designer) and to my surprise, they needed each other. Together we accomplished some incredibly effective tools that would have taken a team of developers and weeks of time. The collaboration is the product.
What Debbie Has Seen
I have used Claude and ChatGPT together on real projects all year — the Business Playbook, the IOG Creator Studio, this newsletter, and a dozen other builds. What I have learned is that the model choice matters less than the brief you give it.
When I provide the source material — the actual document, the actual screenshot, the actual reference — the output is usable. When I describe what I want from memory, I get a polished version of a guess. Both models do this. The variable is me.
The other thing: AI often gets it wrong, incomplete, or almost perfect but not quite as dynamic as you were going for. And they stall. This happens when the chat gets too dense. It happens the most with Claude. The way around this is to limit each chat. Start with a Project. Direct that everything goes in that folder. Complete one portion of the task and tell them to make a handoff. Then open a new chat and do the next section.
This way they are not reading long HTML every time they buffer and reset, using up your credits without completing the task. Also, take screenshots of the work. They can “see” better from an image and will use fewer credits.
Test This Yourself
Pick a real project and run the research-to-strategy workflow above.
Choose something you are actually working on — a marketing plan, a product evaluation, a client proposal. Follow the three steps. Provide the research inputs. Let the AI build the picture. Then read the strategy it produces and decide: do you agree? What would you change? What is it missing?
The output you get from this process will be dramatically different from what you get by typing “write me a marketing plan.” That difference is the entire point.
Quick News You Can Use
CONTEXTUse Projects, not single chats, for ongoing work.
Every major AI assistant now supports persistent project spaces. Upload your brand guide, product docs, and style rules once — every new chat in that project starts informed. This alone eliminates half the wasted setup.
STACKSThe average small business runs five AI tools. Audit yours.
SBE Council data shows the median small business uses five AI subscriptions — $1,200 to $3,000 a year. List yours. For each one, write what it produced last week. Anything you cannot answer for is worth questioning.
QUALITYCustomize before you judge a tool.
Most AI tools produce generic output until you configure them with context about your business, your voice, and your audience. Skipping the setup step is the single biggest reason people conclude a tool “does not work.”
TRAININGA federal AI training bill for small business is moving.
The bipartisan Small Business AI Training Act would direct Commerce and the SBA to create free AI training resources. Not law yet — but it signals that training, not just tools, is the recognized gap.
TIMEDaily users save roughly 7 hours a week. Occasional users save almost nothing.
The data is consistent: time savings come from consistent daily use on specific recurring tasks — not from occasional experiments. Pick one task, use AI on it every day for two weeks, and measure.
Coming in Issue 03
What’s Next
Five Subscriptions Later: Are You Saving Time or Managing More Software?
“Make It Sound Human” Is Not a Writing Strategy
AI Search: Is Paying for “AI Visibility” Worth It?
The tool is not the variable. You are.
