loremax

Your best people know things your documentation doesn't. LoreMax runs an adaptive expert interview that gets it out of them, then compiles it into a playbook a new hire can read, a SKILL.md an agent can load, and a gap report showing what the interview surfaced that you would not have written down on your own.

The interview runs inside your own agent, on your own model. No API key, no account, nothing sent to a LoreMax server.

Install

npx loremax install

It detects which agents the project uses and writes the right file for each:

| File | Harness | | --- | --- | | .claude/skills/loremax/SKILL.md | Claude Code — run it with /loremax | | .cursor/rules/loremax.mdc | Cursor | | .windsurf/rules/loremax.md | Windsurf | | AGENTS.md | Codex, and anything else that reads the convention |

AGENTS.md is appended between markers, never overwritten. Re-running is idempotent, and it's how you upgrade after a new release.

npx loremax install --all   # write every adapter regardless of what's detected
npx loremax install --dry   # show what would change, write nothing
npx loremax where           # list the adapters and their paths

Use

In a new session:

/loremax

In Claude Code it is slash-only, deliberately. The skill carries disable-model-invocation, so Claude will never start a 20-question interview on its own because the conversation drifted near documentation. In agents that don't support that field — Cursor, and anything installed via skills.sh — asking for it by name works too:

Run a loremax interview on how I triage a production incident.

Pick a decision you make repeatedly and other people ask you about. Answer concretely: name the actual case, and when it asks where a threshold came from, don't invent one. "Honestly, it's a feel" is the correct answer, and it's usually the most valuable moment in the interview.

What you get

• .loremax/playbook.md — narrative and rationale, for a human taking over
• .claude/skills/<topic>/SKILL.md — directives only, for an agent

It also offers an optional gap-report.md — what the interview surfaced that you would not have written down on your own. That one costs two minutes up front writing your current version first, so it is off by default. Worth taking on a first run, or when someone else is paying for the time.

How it works

An interview is a router over named elicitation methods — critical decision method, novice-error probes, repertory grids, laddering, contrastive and boundary probes, teachback — chosen each turn against whichever of ten knowledge types is still empty. Every rule you state gets challenged before the end, and no threshold is accepted without asking where the number came from.

Try it without installing anything at https://loremax.vercel.app.

MIT.