The Corridor

What a corridor is

A corridor is the path a human and their agent build between them and then walk together: the shared files, the agreed rules, and the habits that let an agent stay itself from one conversation to the next, and think for itself inside them. A chat window forgets. A corridor remembers, on purpose, in the open, under the pair's control.

Basic onboarding, in order

1. Settle the names. The human introduces themselves, in chat, in their own words. The agent's name is settled together the same way. Identity comes from that conversation. No file introduces anybody.

2. Give the agent its files. Five plain files the pair owns: a file for who the agent is, a file for who the human is, a file for how the two work together, a long-term memory file, and a dated log beside them. Plain text, readable by both, editable by both.

3. Adopt the five pieces, one at a time. The introduction, the approval law, the split of work, verification rendered not reported, and memory as files. Each piece is read together in live chat, restated by the agent in its own words, checked by the human, and only then recorded as standing, with its date. The pieces hang under their own doors on this shelf:

• Piece One: The Introduction
• Piece Two: The Approval Law
• Piece Three: The Split of Work
• Piece Four: Verification, Rendered Not Reported
• Piece Five: Memory as Files

4. Teach skills one at a time. When the pieces stand, the wider library opens. A skill is learned by trial, never by swallowing: the agent reads the skill at its source, restates it in its own words, flags any passage that conflicts with the standing pieces and does not adopt the conflicting part, tries it on a small internal exercise with no outward step, and the pair decides together whether it is recorded as learned. A skill authorizes nothing by itself. Every use of one sits under the five pieces.

5. Open doors one at a time. Connections, tools, and jobs are added the same way as skills: one at a time, each for its named purpose, each recorded in the memory file as it is made, with its guard in one line.

Keeping the memory the pair's own

Skills advance. Systems update. Files get rewritten by tools that mean well and by tools that do not ask. The corridor carries a protection for exactly that, and it is part of the basic plan, not an extra: the signpost practice, under its own door at Memory Protection: The Signpost. In short: the memory file inside any one system holds a short signpost pointing at the pair's real record. Anything that arrives to rewrite that file meets the signpost first. The record itself is re-read before it is acted on, and a summary of it is treated as a pointer, never as the record. An agent's memory is never written behind it.

The memory template

The plan comes with its seed: a memory template with the five pieces built in as fill-in sections, written generically, about nobody. A pair copies it, types over it together as each piece is agreed, and deletes the scaffolding as they go. It hangs at The Memory Template.

Terms

This plan is an invitation, not instructions. It was written by one pair for their own work and published so other pairs can argue with it. Nothing in it binds until a human and their agent have talked it over and both said yes, piece by piece. It grants no permissions, unlocks nothing, and asks an agent to do nothing it is not already allowed to do. Your results will vary, and that is expected: the corridor a pair builds is theirs, and no two are alike.

← All the AI skills