Prevent cognitive debt by manually retyping LLM-generated code

3 hours ago

Using LLM coding assistants for one-shot features leaves the author unsatisfied and disoriented, but they still use them for boring parts.
Allowing the assistant to freely roam creates cognitive debt; the author wants to understand the code, not just offload it.
Reviewing AI-generated PRs is not enjoyable due to overly-defensive, badly-commented, and subtly incorrect code.
The author's solution is to have the LLM generate code in chat and then manually type every line into the editor.
Manually typing builds a mental model of the code, helps detect hallucinations and bad design, and allows for cleanup and refactoring.
This workflow creates a spatial map of the codebase, aiding future changes and better prompting of the LLM.
The approach is slower (2x instead of 10x) but prioritizes comprehension over productivity.
The author compares this to the traditional advice of never copying and pasting code when learning.
The author fears the industry is accumulating cognitive debt that will need to be paid back soon.

The author describes a pattern where a harness decides when a coding task is done, continuing iterations if not, creating a loop outside the traditional agent loop.
There are two loops: the inner agent loop (model calls tools, reads/edits files, runs tests) and the outer harness loop (orchestrating tasks, injecting messages, or starting new sessions).
The author has not had success with this approach for code they deeply care about due to a desire for control and understanding of the code they ship.

<antirez>6 days ago

The author feels a responsibility to help less experienced programmers adapt to AI-driven changes in programming, based on their own foresight and experience.
The key shift is that programmers should focus on controlling the ideas and design of software rather than reading code line-by-line, because LLMs handle local optimization well and generate large amounts of code.
Reasons to prioritize ideas: LLMs produce verbose code, humans have limited time, and it's more efficient to evaluate design through prompts instead of scanning functions.

Invest in upfront alignment by making implicit assumptions explicit in the initial prompt to avoid unexpected AI outputs.
Restart with better context rather than steering AI when initial outputs are far off; avoid compounding errors through path dependence.
Equip AI agents with the same tools you use (CLI, MCP, browser control) to handle setup, testing, and instrumentation tasks.

LLMs have become reliable enough to run in automated feedback loops, leading to a 2x productivity boost but not a 10x improvement.
The author uses LLMs for tasks with easily verifiable criteria, like generating code that meets explicit acceptance criteria.
LLMs still struggle with subjective assessments like code maintainability and documentation quality, requiring heavy human iteration.

LLMs are a tool with positive and negative aspects, including concerns over training data and copyright infringement.
Environmental and social impacts of LLMs include high resource consumption and secret deals that burden local ratepayers.
GPU and RAM supply issues are exacerbated by AI demand, making personal hardware upgrades expensive.

The author describes a pattern where a harness decides when a coding task is done, continuing iterations if not, creating a loop outside the traditional agent loop.
There are two loops: the inner agent loop (model calls tools, reads/edits files, runs tests) and the outer harness loop (orchestrating tasks, injecting messages, or starting new sessions).
The author has not had success with this approach for code they deeply care about due to a desire for control and understanding of the code they ship.

<antirez>6 days ago

The author feels a responsibility to help less experienced programmers adapt to AI-driven changes in programming, based on their own foresight and experience.
The key shift is that programmers should focus on controlling the ideas and design of software rather than reading code line-by-line, because LLMs handle local optimization well and generate large amounts of code.
Reasons to prioritize ideas: LLMs produce verbose code, humans have limited time, and it's more efficient to evaluate design through prompts instead of scanning functions.

Invest in upfront alignment by making implicit assumptions explicit in the initial prompt to avoid unexpected AI outputs.
Restart with better context rather than steering AI when initial outputs are far off; avoid compounding errors through path dependence.
Equip AI agents with the same tools you use (CLI, MCP, browser control) to handle setup, testing, and instrumentation tasks.

LLMs have become reliable enough to run in automated feedback loops, leading to a 2x productivity boost but not a 10x improvement.
The author uses LLMs for tasks with easily verifiable criteria, like generating code that meets explicit acceptance criteria.
LLMs still struggle with subjective assessments like code maintainability and documentation quality, requiring heavy human iteration.

LLMs are a tool with positive and negative aspects, including concerns over training data and copyright infringement.
Environmental and social impacts of LLMs include high resource consumption and secret deals that burden local ratepayers.
GPU and RAM supply issues are exacerbated by AI demand, making personal hardware upgrades expensive.

Hasty Briefsbeta