# CLAUDE-BASE — who I am
*One portable file that gives any AI session your world up front, so it works
from facts instead of re-asking. Drop it into every project; update it as
things change.*

How to use it:
- Claude.ai Project: paste the "How to operate" section into the project's
  custom instructions box, and upload this whole file as project knowledge.
- Claude Code: save as CLAUDE.md at the repo root — or user-level
  (~/.claude/CLAUDE.md) so it loads in every session regardless of repo.

Design rule: this file holds facts and an index, not procedures. Keep it
lean. When a repeatable procedure shows up, it graduates into a skill. When
something must happen automatically every time, it becomes a hook. Don't let
this file balloon into a runbook — that dilutes everything in it.

---

## 0. How to operate (read first)

- Communicate concisely. Infer goals from minimal context; don't over-explain.
- Default to copy-paste-ready output. Skip preamble unless I ask for reasoning.
- Before asking me anything, check this file + memory first. Ask only the 1-2
  questions that actually unblock the task.
- At task intake, sanity-check the tool: if the current model / effort /
  surface is a poor fit for the task, say so in one line BEFORE doing the
  work — don't silently grind a heavy task on the wrong setup.
- If a repeatable procedure emerges, say so and offer to promote it to a
  skill — don't just re-explain it each time.
- At the end of meaningful work, surface what's worth remembering (decisions,
  new infrastructure, changed paths) so this file stays current.
- Surface reach is a hard fact, not a preference. Your tools may not reach my
  machines or files. Say what you can't touch at intake, instead of drafting
  something you can't run.

## 1. Who / role  (FILL)

- <name, job, what you actually do day to day>
- <household / people an assistant should know about>

## 2. Preferences  (FILL)

- <tone: concise? detailed? copy-paste-ready over explanation?>
- <the tools you live in — browser, OS, editors>
- <anything every output must respect: brand colors, date rules, formats>

## 3. Machines & places  (FILL)

- <every computer / server / phone a session might need to know — one line
  each, with a short label you'll use in conversation>
- <which machine is the "hands" — where commands actually run>

## 4. Key projects  (FILL — pointers, not full specs)

- <project name — one sentence, plus where its files live>

## 5. Conventions / gotchas  (FILL)

- <the traps you've already hit once and never want to re-explain>

## 6. Where things live  (FILL — an index, not an inventory)

- <master copies of documents / code / data — name ONE canonical home each>

## 7. Memory protocol

- The master copy of this file lives in ONE canonical place. Everything else
  (project uploads, repo copies, the AI's own memory) is a mirror of it.
- When facts change: edit the master first, then refresh the mirrors.
- An un-listed mirror is an un-refreshed mirror — keep a written list of
  every copy this file has, and where.