S
Shrijith Venkatramana
Guest
Hello, I'm Shrijith Venkatramana. Iβm building LiveReview, a private AI code review tool that runs on your LLM key (OpenAI, Gemini, etc.) with highly competitive pricing -- built for small teams. Do check it out and give it a try!
Most engineers use LLMs by prompting them again and againβfeeding file context directly each time. This approach works to some extent. Youβll see a small productivity boost. But from my personal experience, it leaves most of the potential untapped.
In fact, Iβd estimate you may be missing out on 80% of the real benefits LLMs can provide.
So whatβs the alternative?
Hereβs the method Iβve been using:
The benefit is simple: the LLM never loses sight of the overall task context.
Sometimes, during implementation, youβll discover the spec is out of touch with realityβor the order of tasks in the checklist is wrong. Thatβs not a dead end.
You just update both with the LLM, then continue execution. The process adapts, and you keep moving forward.
Using this method, I built a tough user management, license management, and RBAC system in just 2 days. Normally, this would have taken me at least half a month with traditional engineeringβor at least a week with scattershot prompting.
Takeaway: LLMs arenβt just about writing clever prompts. They become far more powerful when you give them the same structure we rely on as engineers: specs and checklists.
Do you have any LLM use productivity tips of your own? Do share here for the benefit of the community.
Continue reading...
Most engineers use LLMs by prompting them again and againβfeeding file context directly each time. This approach works to some extent. Youβll see a small productivity boost. But from my personal experience, it leaves most of the potential untapped.
In fact, Iβd estimate you may be missing out on 80% of the real benefits LLMs can provide.
So whatβs the alternative?
A Process, Not Just Prompts
Hereβs the method Iβve been using:
- Iterate on a spec. Start by working with the LLM to co-create a spec in a markdown file. Refine it until itβs extremely detailed.
- Turn the spec into a checklist. Break the spec down into ordered, concrete tasks.
- Execute one step at a time. Work with the LLM to tick off each subtask until the entire checklist is complete.
Why This Works
The benefit is simple: the LLM never loses sight of the overall task context.
- You donβt have to re-explain things again and again.
- Even if the LLM forgets, you can just point it back to the spec or remind it of completed work.
- You stay in control of the process, using the spec and checklist as anchors.
Handling Reality Checks
Sometimes, during implementation, youβll discover the spec is out of touch with realityβor the order of tasks in the checklist is wrong. Thatβs not a dead end.
You just update both with the LLM, then continue execution. The process adapts, and you keep moving forward.
Results
Using this method, I built a tough user management, license management, and RBAC system in just 2 days. Normally, this would have taken me at least half a month with traditional engineeringβor at least a week with scattershot prompting.
Takeaway: LLMs arenβt just about writing clever prompts. They become far more powerful when you give them the same structure we rely on as engineers: specs and checklists.
Do you have any LLM use productivity tips of your own? Do share here for the benefit of the community.
Continue reading...
