Field Guide: Project-Local LLM Coding Feedback
Field Guide: Project-Local LLM Coding Feedback
Source: https://x.com/i/status/2081066227619836308
📌 ThePrimeagen’s “field guide” turns hard-won AI coding corrections into durable, project-specific notes so agents can look up rules on demand instead of relearning them each session. Logging sessions to markdown with commit hashes, then distilling them into an indexed guide, keeps context small while cutting repeated deep feedback.
📘 What it is
A project-local library of generalized notes distilled from real LLM correction sessions—not a universal ruleset, but durable guidance tied to how this codebase actually diverges.
🔄 Three-stage loop
1) Nitpick hard and dictate exact intent up front. 2) Log every exchange to markdown and record commit hashes. 3) Have the agent generalize transcript+diff into field-guide notes with an init.md index.
🧠 Why it works
Coding quality often plateaus from lost institutional memory, not model limits. Externalizing corrections keeps the live context small while still retrieving the right rule when needed.
📉 Claimed result
Over ~10 sessions, ThePrimeagen reported a sharp drop in deep hands-on correction—with no noticed change in model or harness.
🗂️ Related practices
Sits alongside Cursor Rules, AGENTS.md, and memory-bank systems: all externalize agent guidance outside the live context window so it can be searched on demand.
Key facts
| Fact | Value |
|---|---|
| Who | ThePrimeagen (@ThePrimeagen) |
| Pattern name | Field guide — durable, project-local feedback from LLM sessions |
| Inspired by | Cursor-style agent/rules workflows |
| Core loop | Detailed sessions → markdown + commit hashes → agent-written notes → init.md index |
| Main benefit | Less repeated deep feedback over ~10 sessions, same model/harness |
| Design tradeoff | Project-specific knowledge preferred over a universal ruleset |
Details
In a July 2026 post, ThePrimeagen described a “field guide” workflow inspired by Cursor practices: treat every hard-won correction session with an AI coding agent as raw material for durable, project-local guidance. Instead of re-explaining the same style, architecture, and edge-case preferences in every chat, you capture the full back-and-forth, ground it in the commits that resulted, and let the agent generalize that material into a small browsable library the next session can look up on demand.
The workflow has three stages. First, be exact up front—nitpick aggressively, dictate long detailed instructions (often by voice), and insist on the code you actually want. Second, log every exchange to markdown and record the commit hash for each change so feedback is tied to real diffs. Third, run a review pass where the agent turns that transcript-plus-diff material into generalized notes in a field-guide folder, with an init.md index of short descriptions and links so later agents can tool-search for the right file rather than loading everything into context.
It matters because LLM coding quality often plateaus less from model limits than from lost institutional memory: context windows fill with chat history, preferences evaporate between sessions, and developers repeat the same detailed feedback. Externalizing corrections into a project-specific field guide keeps working context small while still letting the agent retrieve the right rule when needed—a tradeoff ThePrimeagen accepts because real codebases diverge, and over about ten sessions he reported that deep hands-on correction dropped sharply with no change he noticed in the model or harness.