Why Your AI Prompt Didn’t Work and How to Fix It

This post will be a little different from my usual content because I wanted to include a recent, relevant experience I had that inspired this topic. It serves as a real-world example of what can go wrong when prompting AI and how some slight modifications can significantly improve AI output.

If you’d prefer to just skip to the solution portion of this post, click here.

---

My Recent AI Failure

I recently asked AI to assist me in creating a PowerPoint template. I had a list of content to include and I requested that the template use clean layouts, consistent formatting, and placeholder text so it could easily be customized for any industry.

Copilot was my choice – I figured with it being a Microsoft product, it might be a good choice for this task, since I was requesting a PowerPoint output. My goal was to save time – I could provide the content and leave the design work up to AI.

The model produced an initial draft, I asked for four more slides, Copilot updated my deck, and I decided to start working with the file.

I was deeply disappointed.

Upon opening the file, I was informed it needed repair. That’s not something I’ve encountered much with Microsoft products, so I ran through the automated repair process. That seemed to make the file workable, so I opened it up.

At first glance, the presentation looked great. The colors were attractive, the layout was logical, using tables, grids, charts, and bullet points where appropriate, and it looked really polished.

Then I tried to edit it.

That’s when I realized this presentation was a functional mess:

• What I thought were tables and grids were actually independent components – box shapes that would have to be strategically moved to accommodate any changes with text boxes overlaid.
• The charts were not functional – they were boxes placed over other boxes, so users couldn’t use the chart Excel integration to add data, which is a really helpful feature in PowerPoint.
• The bullet points weren’t actually a bulleted list – the bullets were icon images and each line of text was in individual text boxes, so everything would need to be rearranged any time the content changed on that page.
• Images were used where it would have been more functional to use special characters, or emoji-type icons that could be copied and pasted as a line of text so that it would move and function easily when being edited.
• There was no slide master, so if I wanted to add a new slide, all I would get was a blank slide with no formatting.
• Slide formatting lived on each individual slide, and some had layers upon layers of shapes in varying levels of transparency that prevented the content from being copied and pasted – only some of the content would transfer.
• The initial repair issue seemingly caused some of the formatting to break, so some slides were missing portions of the content, both text, images, and background.

I did like the overall look of the presentation, so I spent the day building my own slide master, converting the pages of individual components into tables and grids, changing superficial bullet-point lists into functional bullet-point lists, and thinking about how I got to this place.

It wasn’t Copilot’s fault I hadn’t been more explicit in my request.

While AI is steadily growing in capabilities – I know it is, because I’ve helped train and test it on a professional level – it’s still working on its ability to read between the lines. It struggles with nuance and translating things that are obvious to humans and the ways humans think, but these things aren’t literal and they’re often not expressed in a way that AI comprehends (yet).

I decided not to have Copilot fix its mess, since I’d already done that myself and I was close to my monthly tasks limit, so instead I decided to provide it constructive feedback. I started by stating I did not want it to redo this task, that I was simply providing feedback (although I still half expected it to rebuild the presentation anyway). I then started with a compliment – the content was excellent, it was well-organized, and attractive in appearance.

Then I let it know the formatting was a nightmare to deal with. I gave it the same information I listed above, with several specific examples for each, including slide number and specific issues caused by the way the presentation was formatted.

Copilot analyzed my feedback, thanking me and producing a list of “key learnings” summarizing my feedback for future reference. It seemed to understand my feedback – that documents, especially those meant to be templates, need to be easily editable, not just aesthetically pleasing.

Then it got confused and said it didn’t see a message from me, even though it had just responded to my message and I hadn’t replied. But that’s another issue for another day.

My Second Attempt

Although I’ve spent hundreds, maybe thousands of hours professionally training AI, I hadn’t applied my knowledge to this situation. I know that if I want a specific result I need to make a specific request.

I decided to test out Claude’s ability to perform a similar task, but this time, I used what I learned from my experience with Copilot and applied it to my request to Claude:

• I stated that the purpose of this PowerPoint presentation was to be a template that should be easily customizable.
• I stated that the presentation must include a slide master with a variety of layouts so additional content can be added.
• Bulleted lists should actually use bullet points and function as a bullet point list, not a disjointed combination of images and individual text boxes.
• Tables need to be real, fully functional tables, not a combination of components that just looks like a table.
• Every component should be easy to modify, like adding bulleted content or adding/removing table content.
• The focus should be on functionality and reusability, not appearance, although I opted to have Claude apply a simple, clean design to make it look appealing.

And wouldn’t you know it – the result was much more aligned with what I wanted.

I still had to make some tweaks, because I’m very particular about how my PowerPoint presentations are formatted, and I rely heavily on a solid and resuable slide master, but the end result was so much closer to what I wanted because I gave very explicit instructions.

What Actually Went Wrong

In this situation, the problem wasn’t that AI failed or that I failed; it was a communication issue:

I thought I’d explained myself well enough. The AI model thought it understood my request.

Ultimately, AI is not a mind reader and should be instructed similarly to how you’d make a request from a coworker. The better quality your prompts are, the better the AI results will be.

---

Applying the “Key Learnings” From This Experience

So now that you’ve read (or skipped) the inspiration of this post, I want to provide some practical guidance on how to get better results from AI.

Just remember to always check the output for accuracy whenever using AI. It can be confidently wrong, hallucinate information, or I’ve even had it slip some Hindi words into what should be an English response. So make sure you are the final reviewer and editor of AI-generated content, even if it looks great at first glance.

---

7 Common Issues & Resolutions

Below are some common issues that result in poor AI outputs and how to resolve them.

---

1. Missing Context

AI can only work with the information you give it. If important background is missing, the response may be technically reasonable but not useful for your actual situation.

Bad prompt:
“Create a project update presentation.”

Better prompt:
“Create a 6-slide project update presentation for senior leadership. The project is a website redesign that is currently on schedule. Include progress to date, completed milestones, current risks, upcoming work, and decisions needed from leadership.”

How to avoid it:

• Explain who the audience is.
• Include the purpose of the request.
• Add relevant background information.
• Clarify what has already happened and what still needs to happen.
• Mention any important limitations or constraints.

---

2. Vague Goals

A prompt may include plenty of information but still fail if the desired outcome is unclear. AI needs to understand what success looks like.

Bad prompt:
“Make this presentation better.”

Better prompt:
“Revise this presentation so it is easier for executives to scan quickly. Shorten long paragraphs, make key findings more prominent, and remove repetitive information without changing the meaning.”

How to avoid it:

• State what you want to improve.
• Define the intended result.
• Use concrete verbs such as summarize, compare, simplify, reorganize, or prioritize.
• Avoid words like “better,” “professional,” or “impressive” unless you explain what they mean in context.

---

3. Conflicting Instructions

AI can struggle when different parts of a prompt ask for incompatible things. When instructions conflict, the model has to decide which one to prioritize.

Bad prompt:
“Create a detailed presentation covering every finding, but keep it extremely concise and limited to three slides.”

Better prompt:
“Create a three-slide executive summary covering the most important findings. Prioritize the information senior leaders need to make a decision, and omit lower-priority details.”

How to avoid it:

• Review the prompt for instructions that compete with each other.
• Decide which requirement matters most.
• Clearly identify priorities when tradeoffs are unavoidable.
• Break a complicated task into multiple prompts if necessary.

---

4. Insufficient Source Information

Sometimes the problem is not the prompt itself; the AI simply does not have enough factual material to produce the requested result.

This is different from missing context:

• Missing context explains the situation.
• Insufficient source information means the facts needed to create the answer are not available.

Bad prompt:
“Write three accomplishments from my last job.”

If the user has only supplied a job title and a few duties, the AI does not have real accomplishments to work from.

Better prompt:
“Using only the information below, identify any accomplishments that are directly supported by my experience. If there is not enough information to identify an accomplishment, tell me what additional details would be helpful instead of inventing one.”

Then provide the actual responsibilities, projects, results, and metrics.

How to avoid it:

• Provide the source material needed to complete the task.
• Include results, examples, metrics, or other evidence when relevant.
• Tell the AI what to do when information is missing.
• Do not expect the model to know private details about your work, organization, or project.

---

5. No Output Format

Even when AI understands the task, it may return the information in a structure that is difficult to use if you do not specify how you want the response organized.

Bad prompt:
“Summarize these survey findings.”

Better prompt:
“Summarize these survey findings using the following format: a two-sentence overview, three key findings with bullet points, and a final section listing recommended next steps.”

How to avoid it:

• Specify paragraphs, bullets, tables, headings, or another structure.
• Provide a desired length or character limit.
• State how many items you want.
• Mention formatting you do not want, such as emojis or excessive headings.
• Give an example structure when consistency matters.

By providing formatting expectations with the initial request, the response is easier to immediately use, since the AI knows what is expected.

---

6. No Protection Against Invented Details

Generative AI is designed to produce plausible responses. If you do not clearly limit it to known information, it may fill gaps with assumptions or details that sound believable but were never provided.

Bad prompt:
“Turn this job history into a strong LinkedIn description.”

Better prompt:
“Using only the information I provide, write a LinkedIn experience description. Do not invent or assume accomplishments, responsibilities, metrics, software, certifications, or results. If important information is missing, omit it rather than filling in the gap.”

How to avoid it:

• Explicitly say when the model should use only supplied information.
• Tell it not to invent metrics, accomplishments, quotes, sources, or other facts.
• Ask it to identify missing information instead of guessing.
• Fact-check the completed response.
• Treat polished wording as a draft until you verify it.

---

7. Too Much in One Prompt

A prompt can be individually clear but still ask the model to solve too many problems at once. The result may satisfy some instructions while ignoring others.

Bad prompt:
“Create a complete 20-slide presentation, choose the visual structure, write all the copy, create five layout variations, make it reusable for any company, include instructions for every slide, and keep everything concise.”

Better approach:
Break the work into stages:

1. Define the presentation structure.
2. Review and approve the slide types.
3. Draft the content.
4. Create layout alternatives.
5. Add customization instructions.

How to avoid it:

• Break complex tasks into stages.
• Approve the direction before asking for detailed execution.
• Separate content decisions from formatting decisions.
• Use follow-up prompts to refine one problem at a time.

---

Before You Rewrite the Prompt, Ask:

• Does the AI understand the situation?
• Have I clearly stated what I want?
• Did I provide the information needed to do it?
• Are any of my instructions competing with each other?
• Did I explain what the response should look like?
• Did I tell it when not to make assumptions?
• Am I asking it to do too much at once?

---

---

Stay Updated

There are two easy ways to get notified when new tutorials are published:

Subscribe on YouTube for new video tutorials: @SimpleSoftwareTutorials

or enter your email below to receive new blog posts directly in your inbox: