Back to Blog

What I Actually Say to Start a Coding Session

The first 5 messages that set up everything (for bugs AND new features)

·Read on Substack →

People ask me how vibe coding works. They want the secret prompt. The magic words.

Here’s the truth: it’s not about one perfect prompt. It’s about how you start the conversation.

Thanks for reading Speak Human! Subscribe for free to receive new posts and support my work.

The first five messages set the tone for everything that follows. Get them right, and AI becomes a genuine collaborator. Get them wrong, and you’re fighting it the whole time.

But here’s the thing: bugs and new features need different openings.

Let me show you both.

Pattern 1: Fixing a Bug

When something’s broken, I need AI to help me diagnose and fix. Here’s the pattern:

Message 1: Context + Situation

I tell AI what we’re working on and what’s broken. I always check the console first and include any errors I see.

Real example from this week:

“I’m working on Clarineffable, Flask backend with PostgreSQL. The family billing feature generates PDF invoices. It works perfectly on localhost but in production I click the button, see a flash, and nothing downloads. I checked the console and I’m getting ‘Error downloading PDF: No module named reportlab’”

Note ~ That error message saved us twenty minutes of guessing.

Message 2: What I’ve Tried

If I’ve poked at it already, I say so.

“I tried the generate button a few times. Same behavior. The modal closes like it worked but no file appears in downloads.”

Message 3: My Hypothesis (If I Have One)

Sometimes I have a guess. I share it so AI can confirm or redirect.

“I think maybe the PDF is being saved to a server path that doesn’t exist in production? Locally it saves to ~/Downloads but on the server that would be /root/Downloads.”

Message 4: The Ask

Now I ask for specific help.

“Can you help me figure out what’s happening and how to fix it so it works in both environments?”

Message 5: Gut Check

Before we start changing things:

“Does streaming the PDF directly to the browser instead of saving to disk make sense? Or is there a simpler fix?”

What happened: Turned out to be two issues, a missing Python package on the server, plus the file path problem. We fixed both in about 15 minutes because AI had all the context upfront.

Pattern 2: Building Something New

New features are different. There’s no error message. You’re starting from “I want users to be able to do X.”

Here’s how that conversation starts:

Message 1: Context + The Why

I tell AI what the app does and what I want to add focused on the user outcome, not the implementation.

Real example:

“I’m working on Clarineffable, my task management app. Right now if someone creates a one-time task and later realizes it should repeat, they have to delete it and create a new recurring task from scratch. I want them to be able to convert an existing task to recurring without losing the task history.”

Notice: no code talk yet. Just the problem I’m solving.

Message 2: What I’m Imagining

I share my rough vision so AI knows what direction I’m thinking.

“I’m imagining a button on the task edit page that says ‘Convert to Recurring.’ They click it, pick a pattern (daily, weekly, monthly), and it creates the recurring template while keeping the original task linked to it.”

Message 3: Ask for the Approach

Before any code, I want AI to help me think it through.

“Does this approach make sense? What would need to change in the database and UI? Are there edge cases I’m not thinking about?”

Message 4: Confirm the Plan

AI usually comes back with considerations I missed. We align on the approach.

“Okay, so we’ll add a recurring_task_id column to link them, create a collapsible UI section for the recurrence settings, and handle the case where they cancel mid-setup. Let’s do it that way.”

Message 5: Start Small

Now we build it in pieces.

“Let’s start with the database change and the API endpoint. We’ll do the UI after that’s working.”

What happened: Built the whole feature in one session. Because we planned before coding, there was no backtracking.

The Difference

Bugs: You have information (error messages, behavior). Share it all upfront. Ask for diagnosis.

Features: You have a vision. Share the “why” and the rough shape. Ask for thinking partner first, coder second.

Both patterns front-load the context. That’s the real secret.

Try It

Next time you sit down with AI whether you’re fixing something broken or building something new try these patterns.

See if starting the conversation differently changes how the whole session goes.

What’s your starting ritual look like? Hit reply ~ I read everything.

Until next time ~ Dacia

Thanks for reading Speak Human! Subscribe for free to receive new posts and support my work.

Want more?

Subscribe to Speak Human for real guidance, no jargon, no hype.

Subscribe Free