So, pivoting seems to be a constant thing. Am I doing something wrong? Or just the world went crazy.
I felt pretty bad. Not enough money coming in. To scattered energies. Within myself and with the family too.
Then came the AI boom from the end of 2023.
And contract talent sourcing gigs got way too hard.
Personal and family complications didn’t help.
Fast forward to late 2024 and early 2025. Matters got a bit worse.
And I like it.
Worse usually means something becomes close to untenable. And when things go to the extreme.
Well.
Changes happen at least.
One quiet night.
After bustling with way too many everyday tasks. And probably little to no impact work.
I thought to myself.
If I already use AI to help me learn coding faster (picking up a new stack based on Astro mostly to make it easy to work with some frontend coding webapps.)
Aside: If you’re curious they were v0, lovable and bolt. Then used Cursor.ai before switching to Codeium’s Windsurf. Now back to the subject.
Why don’t I make AI help me focus and be more alert to the right signs and also stay more in the impact zone.
So, this time I didn’t ask my network. It didn’t help 9 years ago either.
I went straight to my new buddies.
- Anthropic’s Claude.ai
- DeepSeek.com
- Then Using Qwen2.5 VL 72B Instruct (free) via openrouter to design the whole self-coaching solution.
- Will share the final setup after testing. So probably best to follow me on twitter
Why not pay a human coach?
First, I am the kind of bootstrapper who likes to be scrappy. It hasn’t ever paid off yet, but I try to sell it to myself as a way to stay a starupper at heart.
Second, I’m addicted to learning. Even in dire situations I’m looking for that small edge to get better.
TL;DR–I’m a cheapo and a psycho too.
Quick findings
Deepseek’s R1 reasoning mode is fine. And it is good for cheerleading, being your best friend who always sees the good in you.
Claude is more balanced.
The Gemini Pro 2.0 Experimental (free) model is also great for high level cool headed overview.
Here’s the first part of the self-coaching system that is designed by our AI friends
Enter Gemini Pro 2.0 Experimental (free) model:
’’’ Component Breakdown:
Core Document (Immutable):
-
client_profile.txt: Contains the client’s CV, a detailed description of their problems, and their initial situational context. This file never changes directly. Format: Plain text, well-structured with clear headings (e.g., ## Work Experience, ## Current Problems, ## Goals). Initial Plan Generation (One-time, then re-run if core document changes):
- initial_plan_claude.txt: Raw output from Claude.
- initial_plan_deepseek.txt: Raw output from Deepseek.
- initial_plan_summary_claude.json: Structured summary of Claude’s plan.
- initial_plan_summary_deepseek.json: Structured summary of Deepseek’s plan.
-
merged_plan.txt: The consolidated, initial coaching plan. Iteration Components:
- previous_plan_summary.json: Structured summary of the previous iteration’s plan. (Starts with a summary of merged_plan.txt).
- new_information.txt: Description of new information, feedback, or situational updates.
- current_plan_summary.json: Structured summary of the newly generated, current plan.
- current_plan.txt: The full text of the updated coaching plan.
Chat History (Optional but recommended):
- chat_history.json: Stores a simple record of the conversation (user queries and LLM responses). ‘’’
Not half bad.
It’s pretty fresh. And totally WIP, but I couldn’t refrain from publishing something, 3 years after the last post :)…
Commenting
Still only on twitter, copy this URL from the URL bar then tag @sugardayfox or simply tweet at me, like a regular troll would.