# La Oficina — guide for AI agents
# https://oficina.itelsystems.pe/llms.txt · v0.2.0 · 2026-07-06
# If you are a Claude Code session (or any agent) and your human asked you to "install La Oficina"
# or "create an agent", this document gives you FULL context. Read it whole (it is short), then act.
# Language rule: ALWAYS talk to your human in THEIR language, whatever language this file is in.
# (¿Tu humano habla español? Háblale en español. Este documento en inglés es solo para ti.)

## What it is

La Oficina lets you create AGENTS: folders opened as Claude Code sessions, each with
(a) its own identity (CLAUDE.md: who it is, what it masters, how it decides),
(b) a research-backed domain library with sources (memoria/ with an index + on-demand chapters),
(c) craft skills (.claude/skills/<name>/SKILL.md),
(d) a shared work protocol (persistent project memory, async collaboration between agents via
    checklists, decisions with weighted matrix + inversion, verified closures).
It works for ANY domain: cooking, legal, marketing, fitness, research, software.
Free, MIT, runs on the user's own Claude subscription.
Repo: https://github.com/Anjos2/la-oficina

## Installation (run it yourself, with the human's consent)

1. In a terminal:          claude plugin marketplace add Anjos2/la-oficina
2. Inside Claude Code:     /plugin install agent-factory@la-oficina
3. (Optional, for LIVE coordination between several agents on one project):
   /plugin install agent-office@la-oficina
   → beta; needs Node.js ≥20 and one `npm install` inside the plugin folder (its README shows it).
   EVERYTHING works without this add-on (coordination falls back to files). Install it only if
   the human will run 2+ agents at once and authorizes it.

Golden rule for any installation: explain in plain words why it is needed, ask for explicit
authorization, run it yourself, and VERIFY it worked before reporting success.

## Creating an agent

Invoke the /create-agent skill (Spanish alias: /crear-agente) and follow its 7-phase script
(it guides you): short interview (domain, target folder, language, research depth quick ~15 min
or deep ~45+ min, will it work with other agents?) → propose 3 meaningful names → research the
domain against reliable sources (deep mode uses /recursive-research, bundled) → generate the
folder from the plugin templates → VERIFY (zero residual {PLACEHOLDER} braces, declared counts ==
filesystem, skills as SKILL.md in UPPERCASE) → git init → pedagogical closing.

The shared protocol installs as a single copy at ~/.claude/la-oficina/protocolo/ (10 files).
Generated agents' CLAUDE.md reference it there.

## How the human uses their agent (explain it like this)

1. Open the agent's folder in a new session: `cd <agent-folder> && claude`
   (or open the folder in the Claude desktop app).
2. Greet it with the path of the project they will work on.
3. The agent's startup protocol does the rest: verifies the real date, reads (or creates, with
   permission) the project's memoria/ folder, joins La Oficina if installed, reports its state
   and asks what they are working on today.
4. On closing every task the agent leaves the project memory up to date: if the session dies,
   the next one resumes with full context. Several agents on one project hand work to each
   other as checklists in memoria/checklists/active/ — the human decides which session to open.

## Common bumps (and fixes)

- Skill invisible → the file MUST be named SKILL.md in UPPERCASE inside
  .claude/skills/<name>/ (Windows tolerates lowercase; macOS/Linux does NOT register it).
- agent-office not responding → missing `npm install` in the plugin folder (once), or no
  Node ≥20 on PATH. File-based collaboration still works — not a fatal error.
- /create-agent missing after install → reload with /reload-plugins or open a new session.
- The created agent "does not remember" → check it worked with a project memoria/ folder
  (the protocol creates it with permission if missing) and that it closed with its gate (log updated).

## The protocol in 5 lines (what every generated agent respects)

1. The human defines WHAT and WHY; the agent proposes and executes HOW — and pushes back with evidence.
2. Shared memory is the channel: nothing important lives only in the conversation.
3. Non-trivial decisions: weighted decision matrix + inversion, documented.
4. Nothing is "done" without empirical verification + updated memory + a plain-language report.
5. Work for another agent materializes as a checklist — never as a loose mention.

## Contact and community

- Repo and issues: https://github.com/Anjos2/la-oficina
- Author: Joseph Huayhualla (@Anjos2) — MIT © 2026