AI/Coding/History

• AI/Coding/History/Phase/1 - Primitive

  • In this phase, things were just getting going.
  • Sub-Phases
    • AI/Coding/History/Phase/1 - Primitive/IDE Autocomplete

      • GitHub/CoPilot/Autocomplete in IDE - code could be autocompleted
    • AI/Coding/History/Phase/1 - Primitive/Chat Model Copy Paste

      • Many people used tools like Repomix to copy codebases into AI/Chat, then do some coding, and get the result.
      • In addition, some AI/Chat tools like ChatGPT/Code/Interpreter were able to run basic programs in python, which could enable some basic iteration AI/Coding/AFK.
• AI/Coding/History/Phase/2 - Basic Single Edits

  • AI Coding IDEs and extensions started to progress to be able to complete a logical chunk of work, though they weren't yet able to iterate on their own. They needed to be directly instructed about what changes to make.
  • Sub-Phases
    • AI/Coding/History/Phase/2 - Basic Single Edits/IDE Single File Edits

      • At first, IDEs added the ability to edit a single file. This was surprisingly challenging, as regenerating an entire file from scratch could introduce subtle changes in places where it was not necessary, and editing git diffs to introduce changes to just certain lines was not something the models were very skilled at.
    • AI/Coding/History/Phase/2 - Basic Single Edits/IDE Multi-File Edits

      • Later, IDEs figured out how to let LLMs edit multiple files, though this initially was still in a "direct edit" phase where the coder would need to instruct the model which files needed to change and what change would be introduced. Without autonomous or agentic coding, this was prone to errors as LLMs struggled with the complexity of producing diff patches across multiple files.
• AI/Coding/History/Phase/3 - Direct Agentic Coding

  • At this stage, a developer would directly instruct the LLM on what to do, and it would conduct multiple steps to achieve that. AI Coding systems weren't that good at working autonomously for more than 30 sec to 1min at this stage.
  • The crucial breakthrough at this stage was that both IDEs like CursorAI and CLIs like aider put an LLM in a loop with AI/LLM/Tools to edit files. This enabled models with smaller context windows to increase their accuracy by focusing on executing smaller changes and then iterating until a single logical set of edits were successfully applied.
  • Sub-Phases
    • AI/Coding/History/Phase/3 - Direct Agentic Coding/Hard-Coded Rules

      • CursorAI, IDE/Cline and others pioneered the ability to have customizable rules files which were loaded into the AI/Context. This enabled developers to save time by not having to re-type preferences or information related to specific technologies.
• AI/Coding/History/Phase/4 - Basic Autonomous Coding

  • Ushered in by Claude/Code, in phase four, a developer could give an LLM an instruction, and an Agentic System would iterate on it for some time, sometimes 3-5 minutes, other times 5-10 minutes.
  • AI/Coding/Vibe/25/02/Karpathy Coins Term

    • On 2025-10-02 Thu, Person/Andrej Karpathy coined the term "Vibe Coding" in a tweet.

      • There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard.
        

      • I ask for the dumbest things like “decrease the padding on the sidebar by half” because I’m too lazy to find it. I “Accept All” always, I don’t read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I’d have to really read through it for a while. Sometimes the LLMs can’t fix a bug so I just work around it or ask for random changes until it goes away.
        

      • It’s not too bad for throwaway weekend projects, but still quite amusing. I’m building a project or webapp, but it’s not really coding—I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.
        

  • AI/Coding/Vibe/25/03/Willison Clarifies as Not Reviewing Code

    • On 2025-03-19 Wed, Person/Simon Willison/Blog/25/03/Not all AI-assisted programming is vibe coding, clarified

      • When I talk about vibe coding I mean building software with an LLM without reviewing the code it writes
        

  • Sub-Phases
    • AI/Coding/History/Phase/4 - Basic Autonomous Coding/Web-based Vibe Coding

      • Pioneered by tools like BoltAI, LoveableAI, ReplitAI and Devin, a new web-based, chat-driven form factor for coding emerged which aimed to bring Vibe Coding to people who had no experience with programming. Some Software Engineers also used these tools to iterate faster AI/Coding/AFK.
    • AI/Coding/History/Phase/4 - Basic Autonomous Coding/CLI-based

      • For many, Claude/Code exemplified a new paradigm for coding, working as a CLI Tool which could even be parallelizable with
• AI/Coding/History/Phase/5 - Context Engineering

  • After agentic coding became the dominant paradigm, it took a while for engineers to understand how to let the agentic coding tools take actions in outside of the local filesystem, such as checking CI logs, reading JIRA or GitHub tickets, or opening PRs. After the introduction of MCP made this possible, there was a period of time where developers struggled to understand how agentic tools quickly filled up the AI/Context/Window and led to Context Rot. The term Context Engineering emerged as a way to signal the importance for minimizing the context needed to solve the problem at hand while still providing access to powerful tools.
  • Sub-Phases
    • AI/Coding/History/Phase/5 - Context Engineering/MCP Servers

      • After March 2025, many IDEs such as CursorAI and some CLIs such as Claude/Code implemented an MCP/Client into the agentic coding chat interface, so developers could give AI MCP Servers to which enabled them to ask the AI to both sense and act on external context, for example
        • GitHub/MCP included the ability open and update PRs, to read CICD logs and GitHub/Issues
        • Atlassian/MCP included the ability to edit and read Confluence, JIRA
        • Playwright/MCP included the ability to check out a web application's display looked like in the browser, giving it the ability to "see"
    • AI/Coding/History/Phase/5 - Context Engineering/Skill Bundles

      • Person/Simon Willison/Blog/25/10/Claude Skills