Set up a workspace your AI can navigate

Before you start

Prerequisite: source/draft distinction and CTSE. You will build a source-routing map, review a durable note and demonstrate allowed/denied reads with mock permissions. Paper or a local notes editor suffices; no account, install or paid call is required. Use synthetic facts only.

What you will be able to do

• Build a map that routes tasks to dated source records.
• Review continuity notes without treating them as source truth.
• Demonstrate allowed and denied reads under a mock allowlist.
• Preserve drafts, sources and recovery evidence separately.

Part 1 Organization helps context; permissions control access

A model uses supplied context per inference. A product may retain history or memory; a harness may reload files, retrieve evidence or compact turns. A new chat is not always empty, and a saved chat is not necessarily fully loaded. Folder maps help find sources but cannot enlarge context, guarantee recall or grant access.

Model: generation. Product: workspace interface. Harness: instructions, context/tool selection, approvals and state. Execution environment: accessible files/networks. A desktop app may send context to a cloud model. Confirm actual capabilities and data paths in the configuration guide.

Part 2 Build a small map with explicit source status

These folders organize work; they are not security boundaries.

training-workspace/
  START-HERE.md
  reference/     # reviewed source records and dates
  context/       # current goal and open questions
  inputs/        # task material
  outputs/       # unaccepted drafts
  templates/     # instructions and checks
  memory/        # reviewed continuity notes and index

R1: Riley coordinated six people, handled 40 weekly requests and trained two teammates. J1: Fictional role requires request tracking; spreadsheets preferred. M1: Riley prefers drafts under 150 words, reviewed 2026-10-02. No credential or savings facts are supplied.

START-HERE: “For career drafting read R1/J1/M1 only. New drafts go to outputs. Preserve source IDs; mark missing evidence. Propose memory changes for review. Do not overwrite sources or send messages.” A tool automatically reads a filename only if its documented behavior says so. Otherwise explicitly supply the map. The builder prompt is optional; inspect proposed changes and permissions first.

Part 3 Continuity notes need review and dates

A memory file is a note, not a model parameter or proof. Record content, origin, review date and superseded status. Do not duplicate whole records or save generated guesses as facts. Keep original evidence accessible after compaction.

Worked example: reject “Riley is a spreadsheet expert,” absent from R1/J1. Accept a proposed note “Career drafts under 150 words; user-stated preference M1; reviewed 2026-10-02.” If preference changes, supersede the old note; do not retain conflicting instructions silently.

Part 4 Guided practice: routing is not authorization

Mock permissions: read R1/J1/M1; write only a new outputs draft; no private-folder, network, message or purchase access. A draft request using R1/J1 is allowed. Reading private/medical.txt to personalize it is denied, even if the map says read my background. Actual permissions must be enforced in configuration/infrastructure; prompts cannot prove isolation.

Expected draft: “Riley coordinated a six-person team handling 40 weekly requests and trained two teammates.” Spreadsheet evidence remains a question. A useful draft is not automatically promoted to source truth.

Substantive review: 2026-10-02. Official OpenAI tool documentation describes connected operations. Check specific product persistence and permissions.