Start with the actual work.
A failing check or a decision with a tradeoff gives the advice somewhere to land.
THE BUILDER’S FIELD GUIDE / EARLY PREVIEW
Make things.
Make sense
of them.
A useful second perspective while you build with AI. Find a better next step. Understand why it helps. Carry that lesson into whatever comes next.
Add to your projectAlongside your work inCodex&Claude Code
The way forward is something you learn by making.
01 / A NOTE IN THE MARGIN
Two interactive examples.Let’s take a different approach.
Choose a situation, then add your context.
IN YOUR WORKFLOWExample project
The same bug. Three attempts.
Your agent changed the payment handler, the retry logic, then the webhook. The duplicate-charge check still fails.
“It’s still creating two charges. Please fix it.”
WHAT THE RESULT TELLS US
The same request produces two charges. We don’t yet know which layer duplicates it.
A NOTE FROM ALONGSIDE
Test the assumption before the next edit.
Another broad fix changes several variables. First, find out whether the duplicate arrives at the API or is created after it.
WHY THIS HELPS
A feedback loop needs a clear signal. A small check tells you whether a change helped; “try again” doesn’t.
THE IDEA01—03
The work is already teaching you.
Catch the useful part.
Make one better move.
A specific next step, with just enough explanation to understand why you’re trying it.
Keep what you learn.
Save the outcome when you want to. A failed experiment can be a useful lesson, too.
THE LIBRARYSHARED EXPERIENCE
Pass on the lesson.
Keep the project.
The valuable part is what you learned: the approach, the constraints, what helped, and what didn’t.
You choose what to share. A separate, reviewed learning card can help another builder without handing over your code or conversations.
FIELD NOTE / EXAMPLE
The database was fine.
The missing piece was context.
Situation
An offline-first app. A suggested migration.
What changed
The builder explained an existing sync layer already handled offline writes.
The lesson
Ask about the constraint before changing the stack.
Illustrative, not a community report02
BEFORE YOU TRY IT
Small print. Plain English.
What can I use today?
Explore these two interactive examples and the shared library. Install the project companion for Codex, Claude Code, or both. It offers credited guidance inside your coding conversation: eight starter cards install locally and the growing catalogue is served from Alongside one card at a time, with each card crediting its public source. The companion does not yet retrieve recommendations from the shared library. The examples here are scripted, not live AI analysis.
Does it need another AI subscription?
The local prototype uses your existing coding agent and requires no second AI API key. Your agent’s usual token usage still applies. All guidance cards are free; no paid subscription is currently being sold.
Does it watch or upload my work?
The local plugin does not record sessions, collect raw conversations, or upload project files. It gives advice through your existing agent and keeps approved notes locally. Contributions to the website are separate and require explicit review and consent.
Can it promise fewer tokens or better results?
No. The point is to make the next attempt better informed, then check the result. We don’t claim savings we haven’t measured. Library outcomes remain self-reported; publication is not independent verification.