IdeaWell

Project ideas from Hacker News discussions.

How to code Claude Code in 200 lines of code

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

1. Core Simplicity of Coding Agents

The fundamental agent is just an LLM in a tool-calling loop, implementable in ~200 lines.
"the core of a coding agent is really simple, it's a loop with tool calling." – libraryofbabel
"Agents really are just tools in a loop. It's not rocket science." – nyellin

2. Production Features Enhance Reliability

TODO lists, subagents, and context management prevent early stopping and boost productivity.
"To fix 'early stopping' you need extra features... Claude Code does this with TODOs that are injected back into every prompt." – nyellin
"Subagents are a huge boon to productivity." – prodigycorp

3. Model Improvements Outweigh Harness

Better LLMs drive gains more than complex harnesses; simple agents with strong models compete well.
"model improvements far outweigh any harness or tooling." – aszen
"Simpler agents aren't that far behind in performance if the underlying model is the same." – qsort

4. Practical Alternatives and Open-Source Competition

Users share minimal agents and tools, arguing anyone can rival Claude Code.
"here's my take, in 70 lines of code: [nanoagent.js]" – kirjavascript
"We (the SWE-bench team) have a 100 line of code agent... pretty popular." – ofirpress

🚀 Project Ideas

TraceQuery

Summary

  • A diagnostic and investigative tool for exploring the execution traces of coding agents (Claude Code, Codex).
  • It enables developers to "Total Recall" past sessions, identifying why an agent failed, where it hallucinated, or what subagent sidechains were spawned.
  • Solves the problem of "black box" agent behavior by providing a searchable, queried history of logic and tool calls.

Details

Key Value
Target Audience AI Engineers and Power Users of Claude Code/Codex
Core Feature RAG-based search over local .jsonl session logs with TUI/CLI
Tech Stack Rust, Tantivy (Search Engine), SQLite
Difficulty Medium
Monetization Revenue-ready: $10/mo for advanced analytics/sync.

Notes

  • Inspired by claude-trace and comments from bredren and d4rkp4ttern who found that "it is pretty incredible what you can get when letting a frontier model RAG your complete CLI convo history."
  • HN users love understanding the "paperwork around the loop" to debug why an agent "confidently declares victory three steps into a ten-step task."

DoneTools SDK

Summary

  • A specialized reliability harness designed to stop agentic "early stopping" and "tool-forgetting."
  • It implements specific bookkeeping tools: a persistent TODO injector, a DoneTool for explicit signaling, and a "Nudge" mechanism for when the LLM outputs plain text instead of a tool call.
  • Moves beyond the "200-line loop" to handle production-grade persistence and task verification.

Details

Key Value
Target Audience Developers building custom agentic workflows
Core Feature Deterministic task-completion signaling and TODO bookkeeping
Tech Stack Python, Pydantic, LangChain/Langroid
Difficulty Medium
Monetization Hobby (Open Core)

Notes

  • Directly addresses nyellin’s insight: "To fix 'early stopping' you need... TODOs that are injected back into every prompt."
  • Solves the "Emperor has no clothes" vs. "Production reality" debate by providing the "load-bearing paperwork" as a library.

Plan-as-a-Folder (PaaF)

Summary

  • A workflow tool that forces agents to maintain a plans/ directory in a repository containing markdown files of working history.
  • It bypasses context window limits by treating the filesystem as a "living system prompt" and long-term memory.
  • Allows for "forking" an agent's task, manual intervention in the plan, and auditing billable hours/outcomes.

Details

Key Value
Target Audience Teams using AI for complex, multi-day engineering tasks
Core Feature Filesystem-based state management for agentic loops
Tech Stack Python/Node.js, Markdown-diffing
Difficulty Low
Monetization Hobby

Notes

  • Based on lmeyerov’s successful experiment: "all repos get a (.gitignore'd) folder of plans//plan.md work histories... quite helpful... forking/auditing/retrying."
  • It appeals to the HN preference for "with-the-grain" tools that don't rely on complex database backends.

Headless Checkpoint Editor

Summary

  • A lightweight, terminal-friendly wrapper for Claude Code that fixes terminal "jank" and cursor issues.
  • It replaces the default prompt editor with a more robust readline-compliant interface and adds pre-commit hook integration to automatically audit agent output before execution.
  • Includes a "Belt and Suspenders" mode that automatically appends quality-check instructions (linter, static analysis) to every final step.

Details

Key Value
Target Audience Terminal-heavy developers frustrated by current CLI UX
Core Feature Robust Readline editor and auto-verification hooks
Tech Stack Go (Bubble Tea / Charmbracelet)
Difficulty Medium
Monetization Hobby

Notes

  • Addresses loeg’s frustration: "Claude code terminal has the jankiest text editor... holding down backspace just moves the cursor."
  • Incorporates shnpln’s workflow of manual verification: "as your last todo, go over everything you just did... and use a linter."

Read Later