prime-context

Prime Context

Manage project context efficiently — never load everything at once.

Core philosophy

AGENTS.md must be thin. It is loaded in EVERY session — every line costs tokens in every conversation. Detailed content lives in .claude/context/<area>.md and is read only when the task requires it. AGENTS.md contains only: overview (2 lines), routing table, non-negotiable rules, PR checklist.

This skill covers both directions: <area> is the read path (assemble the right context into a session); the bare call is the write path (learn — persist this session's learnings into those same files, so the next session starts smarter).

MEMORY.md format (specification)

MEMORY.md lives at ~/.claude/projects/<hash>/memory/MEMORY.md, where <hash> is the project's absolute path with every / replaced by - (Claude Code's project-folder convention, e.g. /home/dev/app → -home-dev-app). It is a pointer index only — one line per memory file, never inline content:

- [Short title](filename.md) — one-line hook describing when this matters

Each referenced memory file is self-contained Markdown, optionally with frontmatter (name, description, type: user|feedback|project|reference). Rules:

• MEMORY.md never holds memory content, only pointers — keeps the index cheap to scan.
• Dead pointers (file missing) are flagged by doctor and during load.
• Memory files are point-in-time — warn when the newest is older than 7 days.

---

Command: learn (default — no argument, or the literal word learn)

Capture what this session taught into the context files. Claude Code already loads AGENTS.md/CLAUDE.md at session start — re-reading it adds nothing. The bare call's job is the opposite direction: write the session's learnings down so the next session starts smarter.

Flow:

1. Review the current conversation for durable learnings: user corrections, decisions made, pitfalls hit, commands/configs discovered, gotchas that cost time. Ignore anything only relevant to this one task.
2. Route each learning:
   • Area-specific → append to the matching .claude/context/<area>.md (fuzzy-match as in load). No matching area → propose creating one (file + routing-table row).
   • User preference → a file in ~/.claude/projects/<hash>/memory/ plus a pointer line in MEMORY.md (create the directory and the MEMORY.md spec header if absent).
3. Before appending, check the target file — skip anything already recorded; update in place if the learning supersedes an existing line.
4. Report one bullet per learning written, with its destination file. Nothing durable in the session → say learn: nothing new and stop. Never touch AGENTS.md itself (rules there are managed via improve).