Since November 2024, the share of code I commit to GitHub that AI wrote has climbed from 15% to 75%. Even so, the time—and more importantly the energy—I spend building software has not dropped. Here is why I feel blocked: ### Can you please keep asking me questions until the plan is exactly clear to you? - I keep having to choose between writing an oversized prompt that tries to cover every edge case or a concise prompt that is too thin to give proper context.
- When I write the oversized prompt, it bloats the context window with planning files, TODO lists, and walkthroughs, creating context sprawl and causing the model to ignore crucial details. When I write the concise prompt, I get an under-specified solution and fall into endless correction loops.
- The model rarely asks me the precise follow-up questions that would remove the need for repeated planning. **(I probably need to design a research experiment to measure this)** ### Why aren't you using environment variables? - My fundamental issue is that LLMs write code exactly like an intern when being prompted by an intern and an employee on their notice period when being prompted by a Staff Engineer. - Because the agents run out of context before they consider the unfinished parts of my roadmap, they ignore modularization, method composability, separation of concerns, and sensible aggregation of responsibilities. - The worst of these types of offenses is failing to properly use configmaps and environment variables which has of course led to and make shift RBAC solutions in vibecoding infra like row-level security, which leads to very weird workarounds like assigning db rows to customer tokens directly. ### No, I did not ask you for stubbed code; I want a functional integration - Doing service integrations in software is like performing surgery, a precise endeavor. A lot of it involves, just preparing the surface yet models spend less than <50% of their output tokens dedicated to planning standard integrations. - When I ask an LLM to connect to Stripe or post to Slack on an event, I expect working code. I usually get TODO comments, missing functions, or pretend endpoints instead. ### Can you please put all the imports at the top of the file? - "Claude actually writes nested functions and inline imports in python because it must be way more efficient" is an actual line an ex-engineer. The request to move imports to the top of the file is a proxy for a broader issue: the LLM's tendency to produce code that is technically functional but stylistically poor and difficult to maintain. This often stems from the model's focus on local, immediate task completion rather than global code quality and adherence to project standards. - The core problem driving this poor style is Context Sprawl. As the conversation with the LLM progresses, the models have a tendency to generate files of code and try to skunkwork the integrations for them by looking backwards in its generation stack.
![Screenshot 2025-12-14 at 1.06.31 AM](https://hackmd.io/_uploads/rJ2HeriMWl.png)
- For example, you can see here the LLM just tried to regenerate the entire agent with BAML even though it could just compose 2 existing functions and deliver the functionality I requested. ### No I really meant "Claude 4.5 Haiku", can you please use it? - Literally every single vibe coding platform begins with this type of a chatbox, which psychologically drives the user to focus on typing in a short prompt for quick generation as opposed to thinking through and defining what needs to be built. Attaching a file like a pitch deck doesnt work very well either because, extracting software specifications accurately from the file is expensive at scale.
![Screenshot 2025-12-13 at 11.31.59 PM](https://hackmd.io/_uploads/By8Q97sfZl.png)
- As a result, you keep defining additional details/requirements in smaller chunks in followup conversations and the chat history stagnates at an exponential rate. Since the model is obviously biased by its own knowledge cutoff when it gets overwhelmed by the information in its context, it leads to pretty funny behavior by models with the intelligence of PhD interns.
![Screenshot 2025-12-13 at 11.28.56 PM](https://hackmd.io/_uploads/Bk3DtXjG-g.png)
This is how much context I need to provide the model when I am 3 chats deep in a conversation to get a simple task done. Thank God, for Antigravity; Cursor would have bankrupted me for this simple model setting change. ### Can you please try again? Its still broken - I think the premise of vibecoding has been most valuable to non-technical or lightly technical users to build personal software until their lovable or v0 chat thread is stuck in a loop of "This isnt fixed. Can you try once again?" replied in increasingly irriated variations. And this isnt fixed by model quality as you would expect. ![Screenshot 2025-12-13 at 11.15.14 PM](https://hackmd.io/_uploads/Hkv4IXiGWe.png)
Thankfully, Vercel keeps **fix with v0** free. That still does not compensate me for the time lost in the ordeal. ---- > **Inferred Insight:** Codebases are a very "token inefficient" and machine interpretable representation of large scale software systems.

Blog

Prism: Query Intent Governed ETL Pipelines

05. Helix: Data Models That Evolve With Your Usage

06. Conduit: Infrastructure That Simplifies Itself

07. Membrane: Service Boundaries That Actually Make Sense

08. Synapse: Lean Real-time Generated User Interfaces

09. Nexus: Continuously Improving Agents

10. Crucible: Continuously Evolving RL Environments

03. Eureka: Search That Learns From Every Query

02. Software as Accumulated Design Decisions

01. The Issue with Autogenerated Code

04. Prism: Analytics That Anticipate Your Questions