"Vibe" is an unfortunate term that implies an emotional response or feeling; however, I want you to maintain reason as you "vibe" a plan.
In the early days of "vibe coding" people were just looking for an idea machine; "ChatGPT: build me a web page to make a million dollars". Instead we will be looking to steer things to a useful end with planning. As good as AI is, "Artificial Intelligence" is not quite accurate. Instead think of AI as "Amplified Intelligence". Where you are weak AI can help you learn and where you need help implementing, AI can be the hands that follow your directives. In short we will not be mentally checking out, instead we will be guiding and molding the work.
To start we want a "good" plan, not a great one, not a perfect one, we are simply looking for good at this point and can refine as needed to make a great one. Your mileage may vary as a great plan will be mostly dependent on your vision.
I have had several clients request work from us with plans that have been undoubtedly created with ChatGPT or Claude. Some are good, especially when the person sending it has an understanding of the topic. However others are filled with information that the sender doesn't understand and sometimes leaves us scratching our heads as we try to ascertain their intent. The planning or brainstorming process can help you learn what's needed to achieve your goal.
Let's start with an idea.
- cut waste from feature X.
- speed up deployment time.
- deploy new infrastructure for feature Y.
- develop new marketing campaign to draw attention to new service.
- build new inventory application for mobile devices.
Regardless of where you start, be it a large grand plan or a distilled idea, our goal will be to develop action items. Don't worry about perfection at this point. Just realize that the larger your goal the more action items you may need and the more time needed to plan out everything. You can follow an individual line of thinking to conclusion or try to plan out each layer. If you want to go full Waterfall planning you can, but I recommend taking smaller chunks. If you come from an Agile background then this is not much different than breaking down work into Epics, Features, and Stories or Tasks. Once we get to the Story or Task layer we can set about to discovering how much effort will be needed to bring your idea to life.
Additionally, this is where I tend to prefer an LLM harness like Claude Code or Codex vs using the desktop apps as I can develop a library of documents or artifacts that the LLM can consume as needed instead of filling up its context window with items that may not be relevant for the current planning. Context management becomes more important the more complex and larger your plan gets. By using the LLM harness you will develop artifacts you can return to and refine. However you can still use the Desktop client, just realize it will eventually hold you back with larger planning/brainstorming.
PROMPT: I need help brainstorming the implementation of a new feature. Act as a product manager and interview me one question at a time to help me develop a plan for what is needed to implement this feature. Record the questions and answers in a markdown file.
Key items:
- Set Intent: "help brainstorming the implementation of a new feature"
- Define Role: "act as product manager"
- Set Interaction Method: "interview me one question at a time"
- Define Outcome: "develop a plan"
- Define Memory: "record the questions and answers in a markdown file"
Different prompt engineering guides will say similar things so call these what you want, but understand we are leaving some room open for the LLM to work while still providing some guidelines around your expectations. You can add more to this like maybe you already have gathered context and you want the LLM to be aware of it. It is possible to overload the context so keep it to what is necessary for planning purposes. I am assuming you have a goal, but don't necessarily know all of what you need to get to a solution. This work is to help bridge that gap.
At this point the LLM should act like a Product Manager and it should start asking you questions. We left things vague so that the LLM will need to probe and interview you. If you wanted you can change the Interaction method to have it write to a file all the questions and you could fill it out and come back. Or maybe you are a product manager already and need the LLM to act as a Cloud Engineer to help plan out the work needed by the team. Regardless, these five key items can be shifted as you need.
While going through this process the LLM will create a markdown file to record the questions and answers. Think of this as the "save often" step. This is important if you need to come back on a different day or you need to do some added work to find the information the LLM is looking for. Also when you run into something you don't understand this will let you either find the information you need, ask the LLM to clarify, or flag an area to be addressed later.
Before I go further and you think "this can replace a Project Manager?" the answer is no. Depending on the complexity you will likely still want someone in that role. The LLM is an amplifier. If you understand enough about any process then it can help you get better, but in most cases it still isn't at mastery and can be led astray by your own context. Even an experienced person can be led down a bad path because your understanding of something didn't match with the experienced person's understanding. That said, the plan can give you a great starting point to work with a skilled person and in some cases the LLM can help guide you so that you learn and can fill that role sufficiently.
Another facet of the planning process is that you don't need to get it all complete. While going through the process if there is an item you don't want to walk through now then tell the LLM to remind you later and leave this as an open item, then direct it to focus on the area you want to concentrate on. At some point you will come back or maybe you recognize you need more help in which case you use this planning document with your teammate to work through the items.
Once we have the plan to a point that you can start implementing I want you to pause a moment. Start another LLM session and ask that LLM to challenge your plan.
PROMPT: review myplan.md. Focus on the part that is "ready for work". I want you to challenge this plan. Assume the role of the implementor and push back on the technical items of the plan. Add your challenges to the plan under "Implementer Challenge"
You can also do this from other vantage points. "Assume the role of an accountant and challenge the cost of the plan.". In short use the LLM to function in the roles you would need to implement the plan. If you don't know ask "What roles are needed in order to implement this plan? Add section to the document with those roles and their function." This is also a great place to build a library of personas so that they can be referenced as needed.
After you go through the challenge process we should hopefully be ready for implementation. Maybe you had multiple sessions to move from grand idea to targeted features and finally to a list of action items. Maybe you were quick to get to action items as you only have a singular feature. Either way you need to go back and review the work. Ask yourself "Do I understand the contents of the plan?" and "Can I walk someone through the plan?"
I am not asking if you can do all the work, or if you are well versed in what the plan entails, but I am asking if you understand the terms that are used and can you speak to them? If it says "Setup Entra ID" do you know what "Entra ID" is and do you understand the setup steps enough to ask someone to do it? If you are working with software and it says "Update bun" or "Use bun to update react" do you understand what "bun" and "react" are and how they relate? This process of making a plan can help you understand your own knowledge gaps be they self imposed or due to lack of capability. Either way this may help you understand when you need help or it can help you in your discussion with a partner.
If you understand it all and you think between the LLM and yourself you can make it work, then go for it. Start with a proof of concept and build. In your building process you may realize you need help or you realize you can make this great idea happen on your own. Alternatively you may look at the plan and all the work and realize "There is no way I can test all this and understand the results I'm gonna need to get help." If that's the case then bring your plan to your teammates, company architects, or look for a group like 27Global who can help. Bring your plan and a willingness to have it challenged and clarified. AI tools and planning may get you 80% of the way there or it may only help you realize the level of effort needed to make it happen.
Speaking of effort...
PROMPT: Review these action items and give me a table with the amount of time to implement each of these tasks. Call out which tasks need manual boot strapping vs tasks that can be complete by way of API calls. Note any task that will need money upfront to implement vs pay as you go. Additionally note all items with a documented implementation time with an "*" and anything with a speculative time noted with "?". If the item has a documented implementation time, then cite where this implementation time was found in the reference section at the end of the document.
This will give you an estimate that tends to use rough human time frames. Sometimes it's pretty good and sometimes it's off. Software development is not well defined for time and thus it's almost always a very rough estimate. Some organizations may have processes that slow things down on purpose whereas others move at the speed of lightning. I often use Claude and Codex to challenge each other and challenging time estimates is no different and can be helpful to find gaps in estimation. Always assume it is the "best case scenario" and be prepared to have people speak to how quickly it can be done. Note that there is a difference between POC and Enterprise level software. Code is cheap, Software is expensive, and time estimates are more abstract art than science.
In conclusion, I want you to try to build a plan with your "Amplified Intelligence". Remember our key items for the initial prompt; "Set Intent" so the AI has a defined mission, "Define Role" the AI should assume, "Set Interaction Method" for how the AI should work with you, "Define Outcome" so that the AI has something to build towards, and "Define Memory" so you know where context is stored and your artifacts can be worked with. Leave room in your statements so the AI has some flexibility but still operates in a well-defined topic so as to minimize hallucinations.
With these items and the planning process you can use the AI to distill the plan down to a working set of action items and then determine how well you understand what you are given. If you and your AI agents can implement it in the time frame you wish to work then best wishes. If you go through the process and realize you need more help, then contact your teammates or contact 27Global and we can help implement or help you refine your plan.
Happy planning
