Pair-program with LLMs on personal projects
Not since the advent of LLM based AI. Task: Not since the advent of LLM based AI. I’ve used the web incarnation of Claude in the past and lately, Jules (Gemini with a GitHub integration).These tools require a certain approach to do pair-programming, but I’m OK with not actually sitting at the same computer alongside someone, or sharing some video conference. More a case of being a navigator with occasional driving where I do the coding locally and submit it back to the LLM.If you can pay, you can use the locally-running tools to get a more interactive session going.The usual caveats about using AI apply, of course, but that’s the fun part of working with others, be they human or not. You never know quite how it will go!; tools: Claude. Human role: The original post does not say how a person stayed involved.; limitations: The primary constraint is the need for a specific interaction style where the human drives locally and the AI navigates, which may not suit all coding tasks. The author mentions that locally-running tools require payment for a more interactive session. Standard AI caveats apply, including potential hallucinations or suboptimal code suggestions that require human verification..
Tools
Published on Agent Work Atlas:
Overview
Not since the advent of LLM based AI.
Workflow
Not since the advent of LLM based AI. I’ve used the web incarnation of Claude in the past and lately, Jules (Gemini with a GitHub integration).These tools require a certain approach to do pair-programming, but I’m OK with not actually sitting at the same computer alongside someone, or sharing some video conference. More a case of being a navigator with occasional driving where I do the coding locally and submit it back to the LLM.If you can pay, you can use the locally-running tools to get a more interactive session going.The usual caveats about using AI apply, of course, but that’s the fun part of working with others, be they human or not. You never know quite how it will go!
Reported result
The source does not report specific measured outcomes, such as lines of code saved or bug reduction rates. The author notes that the experience is unpredictable, stating, 'You never know quite how it will go,' implying variability in output quality. No quantitative success metrics are provided to validate the efficiency of this method.
Human role
The original post does not say how a person stayed involved.
Limitations
The primary constraint is the need for a specific interaction style where the human drives locally and the AI navigates, which may not suit all coding tasks. The author mentions that locally-running tools require payment for a more interactive session. Standard AI caveats apply, including potential hallucinations or suboptimal code suggestions that require human verification.
Try it
Act as a pair-programming assistant for my personal project. I will provide the context: [replaceable: project description and tech stack]. I will submit code snippets or problem descriptions to you. Your role is to act as the navigator: review my local code, suggest improvements, or generate the next logical block of code based on my input. Do not assume you are driving the entire session; wait for my local implementation updates. Before finalizing any code suggestion, ask me to confirm the logic aligns with my current local state. Start by reviewing this sample snippet: [replaceable: code snippet]. Provide feedback and a suggested next step. I will implement it locally and return the result for your verification. Ensure all outputs are clearly marked as suggestions for my review, not final commits.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.