Have your AI keep an evidence checklist. Watch it fill in.

Step 1

Give the agent the rules

Replace the two lines in brackets with your job and a file name. The rules stay the same for every job: one file, checkboxes, evidence, and no silent fixes.

Task: [describe the job in one or two sentences]

Create and maintain a single file: [checklist.md]

Rules:
- Use Markdown headings for the phases or areas of the work.
- Under each heading, use task checkboxes: - [ ] or - [x]
- Under every task, include:
  - Done: what was done
  - Issues: problems found
  - Evidence: command output, file counts, test results, links or timestamps
- Only mark a task [x] when you have recorded evidence that it succeeded.
- If you discover a problem, add a new task or note under the relevant heading instead of silently fixing it.
- On every update, overwrite the same file with a clean, complete version. Do not create new files or append duplicate sections.
- Keep the headings and task order stable between updates.
- Keep a short "Current status" line at the top.

Start by writing the full checklist with every task unchecked, then begin the work and update the file as you go.

The same prompt, different jobs

Only the task line changes. Evidence is whatever proves that kind of work is done.

Step 2

Open the file in LeanMD and Watch it

1. Open the checklist.As soon as the agent has saved the first version.
2. Press Ctrl+Shift+W.Watch is on, and the tab becomes read-only, so you never type into a copy the agent is about to replace.
3. Open the Outline and choose Tasks.Ctrl+Shift+O, then Tasks. Every checkbox, grouped by heading, with done / total counts.

Then leave the window open and get on with something else.

Step 3

Review the marks, not the whole file

# Server migration: old-fileserver to new-nas

**Current status:** In progress. Retention policy corrected, script line endings fixed.

## Phase 1: Inventory
- [x] List shares and sizes on source
  - Done: `Get-SmbShare` and `du`
  - Issues: None
  - Evidence: 14 shares, 2.1 TB total
- [ ] Capture permissions
  - Done: Not started
  - Issues: None yet
  - Evidence: None yet

## Phase 2: Transfer
- [ ] Copy data with robocopy
  - Done: Not started
  - Issues: None yet
  - Evidence: None yet

## Phase 3: Validation
- [x] Confirm retention policy
  - Done: Agent had set 14 days; corrected back to 30
  - Issues: Quiet change in revision 3
  - Evidence: Line 13 now reads `Retention: 30 days`

Optional

Stronger rules for longer jobs

Add any of these to the end of the prompt.

Force evidence before a box is ticked

Never mark a task [x] unless its Evidence line contains a concrete artifact (command output, checksum, log line, test result, file count or source link). If evidence is missing, leave it [ ] and explain what is still needed.

Make problems visible instead of quietly fixed

If you find a problem, do not just fix it. Add a new unchecked task describing the problem, then a follow-up task for the fix. Only check the fix task after you record evidence.

Keep a short update log

At the bottom, maintain a "## Update log" section with one line per save, for example:
- 2026-11-12 09:15: Corrected retention from 14 to 30 days (evidence: line 13)

Watch shows what changed since you last cleared the marks; it doesn't keep old versions. The update log gives you a record inside the file itself.

Tips

• Name the file and say "overwrite it". Watch follows one file that keeps its name. A new file per update means nothing to watch.
• Keep the structure stable. If the agent reorders the whole checklist on every save, every line gets marked. Stable headings keep the marks and the counts meaningful.
• The agent doesn't need to know about LeanMD. It writes ordinary Markdown with - [ ] and - [x]. LeanMD does the counting and the highlighting.
• Want to edit the checklist yourself? Stop watching first, and tell the agent to reread the file before its next update so it doesn't write over your change.