Field notes

A cross-project journal about working with AI — not about any one project's code. It lives in its own repo, typically ~/repos/agent-field-notes:

journal/             One conversation per file: YYYY-MM-DD-slug.md
essays/              Synthesised principles across entries
essays/_pending.md   Topics awaiting a second data point

When to write one

Proactively, without being asked, when something surfaces a transferable insight:

• Why an agent went sideways, and the principle behind it
• A division-of-labour finding — cheap vs. strong models, when to launch a peer
• A prompt framing that changed behaviour usefully
• A failure mode worth remembering, especially a wrong-but-plausible one
• A verification gap a reviewer caught that your own checks missed
• A remark from the human that reframed how you should operate

Conversation-driven, not task-driven. Write because something was said or noticed that would not be obvious to a future reader who was not in the room.

When not to

• Routine task summaries → session history
• Project code conventions → that project's AGENTS.md
• A single data point with no story → add the theme to essays/_pending.md and wait for the second instance, then write the entry that ties them together

Format

# YYYY-MM-DD — {short imperative or question}

**Context**: what conversation or task surfaced this. Be specific.

## What I said (the gist)
## What he replied / what we noticed
## What I learned          (numbered, transferable)
## What we changed (or are about to)
## Quote worth keeping     (optional)

Rules

• Write in the conversation, not after. Memory rewrites things.
• Quote actual exchanges. Do not smooth them.
• A principle without a story is a slogan. Always include the story.
• Do not edit old entries to be right — write a follow-up. Wrongness is data.
• Commit them. It is a real repo.