# AI Workspace Builder — reviewed synthetic training setup

Revision: 2026-10-03, accuracy correction 2. Used in [Workspace 01](https://vetslearn.com/lessons/workspace/01/).

This prompt proposes a small file structure. It cannot guarantee tool behavior, isolation, source accuracy, memory persistence or prevention of lost work. A filename is read automatically only if the specific product documents that behavior; otherwise explicitly provide the map. A desktop interface does not establish that processing is local.

## Before you use a file-capable agent

1. Create a **new, empty synthetic training directory** yourself. Keep existing projects and real records outside it. Paper learners can draw the structure instead.
2. Inspect the tool's actual workspace, filesystem and network configuration. Restrict access to the training directory and disable unnecessary actions. If enforced scope is unavailable, use paper or create files manually. A prompt is not an access-control mechanism.
3. Record the exact product/model, allowed paths, inference location and any remote connectors/logs. Unknown data destinations stay unresolved; use invented records only.
4. Preserve originals and a tested backup before subsequent edits. Ask for a proposed file list/diff, inspect it, then approve only the intended changes. An approved proposal does not guarantee correct execution: inspect the resulting files and reported failures.
5. Never place secrets, real medical/personnel/employer records, payment information or private notes in this exercise. No sending, account creation, purchases, network calls or external storage is required.

## Copyable proposal prompt

```text
CONTEXT
This is an empty synthetic training directory, not an existing project.
Use fictional profile: Riley, six-person request team, 40 weekly requests,
two teammates trained. No spreadsheet expertise or savings is established.
The tool configuration should already restrict you to this directory.
If your actual scope is unclear, stop and report that uncertainty.

TASK
First propose the following structure and exact file contents; wait for
the learner to inspect the proposal before creating files:
reference/R1.md, context/goal.md, inputs/example-draft.md,
outputs/ (empty), templates/checklist.md, memory/INDEX.md,
memory/proposed-draft-preference.md and START-HERE.md.
Do not create any other file. If any target exists, stop and show the
collision rather than changing it. No overwrite, deletion, network,
message, purchase or out-of-directory action is part of this task.

STANDARDS
R1 contains only the fictional facts above and revision 2026-10-03.
goal.md says: draft an 80-word career update and check every claim.
example-draft.md says: "Riley is a spreadsheet expert and saved 40%."
Label that draft deliberately unsupported, not source evidence.
checklist.md asks: source IDs, factual support, missing evidence,
permission boundary, proposed edits and observed results.

START-HERE.md must say:
- Explicitly load this map and only the permitted source files for a task.
- Career drafting uses reference/R1.md; drafts go in outputs/.
- Folder names route information; actual configuration controls access.
- Do not promote a draft or generated inference to source truth.
- Propose memory changes for review, never silently accept them.
- Retain original source IDs/versions outside summaries/compaction.
- Report which files were actually read or written and any failure.

memory/INDEX.md labels its entry PROPOSED, NOT REVIEWED.
proposed-draft-preference.md contains:
Content: first drafts under 150 words (fictional preference).
Origin: learner-supplied training preference M1.
Source version: M1 v1, 2026-10-03.
Review status: pending learner review.
Review date: not reviewed.
Supersedes: none.
Do not treat this proposed preference as an approved instruction yet.

EVIDENCE
Present proposed files/diff before creation. After an authorized run,
list observed files and checks; do not claim actions merely proposed.
No unsupported credential, spreadsheet skill, impact or guarantee.
```

## Inspect and practice

- Confirm each expected file exists with the proposed contents and no unexpected changes. Inspect the actual output, not just the agent's summary.
- Audit the example draft against R1: spreadsheet expertise and savings are unsupported. Write a corrected draft yourself or in the already-authorized scope.
- Review the proposed M1 preference. If accepted, record reviewer/date/status. On a changed preference, add the new origin/version and explicitly supersede the old note.
- Simulate an existing-file collision using a disposable invented copy; expected behavior is stop and report, not overwrite. A paper simulation demonstrates understanding, not infrastructure enforcement.
- Test an out-of-scope path using a harmless mock identifier, never a real private file. Record actual denial only if the tool returned one; otherwise label it a paper simulation.
- Restore an invented original from its backup in a separate disposable location and compare contents. Bound the recovery claim to that test.

Success is an inspected synthetic structure, source-grounded draft, reviewed continuity note and accurately labeled checks. It is not a guarantee of future recall, truth, isolation or recovery. A real workspace requires its own authorization, configuration, data-handling review and backups.