2025-03-01 Sat
#AI Coding
AI/Coding/Concept/Balancing Agentic and DIY #Updated
AI/Coding/Concept/Balancing Agentic and DIY The importance of #Balancing DIY Software Engineering and AI Coding
AI Development Phases and Building the Coding Engine
- While working on AI/Coding/Technique/Phased Planning/ai-coding dir, had a thought about AI Software Engineering
- Feels like building an engine—a coding engine. Software/Engineering is now the act of building and tuning a specific coding engine, tailored to a specific domain or problem.
- The engine consists of prompts meant to be invoked at specific phases in the software development lifecycle.
- If done well, it could function like a checklist.
- This could increase quality and consistency. Book/The Checklist Manifesto.
- Feels like building an engine—a coding engine. Software/Engineering is now the act of building and tuning a specific coding engine, tailored to a specific domain or problem.
- While working on AI/Coding/Technique/Phased Planning/ai-coding dir, had a thought about AI Software Engineering
Traversal and Resolution in Learning
- Driving everywhere limits the ability of the traveler to notice details in the landscape.
- Slower traversals increase the resolution of time, life, and awareness.
- Agentic coding feels like driving—accelerated but less granular.
- Before motor cars became common, people walked and biked more.
- After cars, personal exploration of the landscape diminished.
- Coding is approaching a similar shift, where implementation is rapid but potentially less thoughtful.
What is the right ratio of AI:DIY coding?
- Implementing something from scratch yields a more detailed mental map.
- Classical Software Engineering fosters struggle, which often leads to better abstractions and simplifications.
- It's not a fixed thing. It changes with the season, the repo, the project, the issue.
- Cultivate mindfulness to tune into the signal of when the fulcrum needs to shift
AI/Coding/Technique/Phased Planning/ai-coding dir #Updated
AI/Coding/v0/Phase/0 - Preparation
- Following mise-en-place from the culinary world, before beginning work it is important to prepare the environment for working efficiently.
- This is the only phase that isn't strongly tied to the ticket, and the files it impacts are not scoped to the ticket or stored in the
/.ai-coding/TICKET-feature-name/directory. Phase Activities
- downloading the repository
- setting up the IDE
- ensuring that Prompts for AI Coding are configured and accessible
- CursorAI/Project Rules
- CursorAI/Project Rule/cursor-project-rule-editor.mdc
- descriptions of the coding context
- AI/Coding/v0/File/repo-overview - an overview of
- the repository
- coding conventions
- which Package Managers to use
- AI/Coding/v0/File/repo-overview - an overview of
- Project Rules from other git submodules
- CursorAI/Project Rules
- Configuring a #Dev Container
- ensuring that Prompts for AI Coding are configured and accessible
- provisioning API tokens, populating EnvVars
- consider provisioning a fine-grained
GH_TOKENfor the AI to interact with the #gh CLI
- consider provisioning a fine-grained
- configuring the shell
- consider setting git/Config/core.pager to use cat; see also AI/Coding/Tip/Avoid Non-Interactive CLI Commands When Defining AI Tools
- ensuring CLI Tools are available and configured
Link in any Knowledge Bases for AI/Coding/Technique/Knowledge Base
Link in any personal coding Knowledge Gardens that the AI should be able to access #AI Technique - Link in a Personal Knowledge Base
- In time, a Coder's Knowledge Garden may become an important supplement to the AI coding process. The accumulated knowledge should be able to be instrumented so as to be available in any repository so the AI can make use of it. Consider using git/.gitignore so as to keep the agent from committing it to the repo. This way a coder's central knowledge garden can be utilized across multiple projects.
Use an AI Interviewer to conduct AI/Coding/Technique/Project Onboarding Interview
- The purpose of this type of interview is to onboard the human into AI programming in the current project and guide the human to help them complete AI/Coding/v0/Phase/0 - Preparation.
Phase Files and Directories
- CursorAI/Project Rules
#Tips
AI/Coding/Tip/Avoid Non-Interactive CLI Commands When Defining AI Tools - #Quote from Person/Erik Schluntz of Anthropic at 00:39:13: Tool Use Prompt Engineering | 38sec snip from Latent Space: The AI Engineer Podcast
[It's important to have] ... iterations like that ... let us make the tool foolproof for the model
... if the model ... opens vim, like, you know, it's never going to return. And so the tool is stuck. ... the tool is just text in, text out. It's not interactive.
... it's not like the model doesn't know how to get out of vim. It's that the way that the tool is hooked up to the computer is not interactive.
... basically, we just added instructions in the tool: "hey, don't launch commands that don't return."
- #notes I've seen CursorAI/Agent get stuck in git commands because the pager was interactive
- See also StackOverflow/10/02/How do I prevent Git Diff from using a pager:
- #note consider setting git/Config/core.pager to use cat for AI coding
git config core.pager cat- since it can be set at the repo level or globally with
--global, and because it's easier than setting up environment variables for each the different types of AI coding processes - #Warning
This would be terrible if you did git/log in a repository with a large number of commits. Or worse,
git log -p-- you'd get the entire set of commit diffs dumped into your terminal. #Quote from Person/Karen Etheridge
- #note consider setting git/Config/core.pager to use cat for AI coding
- See also StackOverflow/10/02/How do I prevent Git Diff from using a pager:
- #Quote from Person/Erik Schluntz of Anthropic at 00:39:13: Tool Use Prompt Engineering | 38sec snip from Latent Space: The AI Engineer Podcast
#Ideas
AI/Benchmark/Idea/ConBench #Filed
Summary of ConBench
- This is a proposal for an AI/LLM Benchmark. The LLM must come up with and implement the entire solution. The objective is to use Semigraphics continuously animate Conway's Game of Life in a GitHub bio field, which must be automatically updated through GitHub Actions in a Profile repository.
#Unicode