Free to read · no account needed

16 min · agents

Define “done” and permissions before tools

Goal: write a clear finish line and a short allow-list before any agent touches files or apps.

Best for: beginners who want to try agents without scary accidents (wrong email sent, files moved).

An agent without a finish line will keep “helping.” An agent without permission limits may do too much. Before you connect tools, write two tiny lists: Done and Allowed.

Done = how you will know the job is finished (checklist). Allowed = what the agent may read or change. Everything else is “ask me first.”

Why beginners skip this

Letting an agent “just start” without Done or an allow-list feels agile until it writes or sends somewhere you never meant. Finish line first; permissions second; irreversible behind a human gate.

Habit this week (Asia/Shanghai): before one live agent run, paste Done checklist + allow-list; refuse the run if either is blank—date the pass/fail.

Step 1 — Write a Done checklist (3–6 lines)

Finish lines beat vague goals like “organize my files.” Name the output, where it lives, and what you will spot-check.

Goal: Turn my vague wish into a Done checklist.
Wish: [paste]
Format:
1) One-sentence goal
2) Done checklist (3–6 boxes)
3) What I will spot-check myself
4) Stop conditions (when the agent must pause)
Boundaries: Beginner-friendly words. No tool use yet—planning only.
Example

Filled prompt

Wish: Clean up my project notes so I can find last week's decisions.
Format: one-sentence goal · Done checklist · spot-check · stop conditions. Planning only.

What good output looks like

  • Done boxes are observable (e.g. one summary doc in Notes/Q4)
  • Stop conditions include uncertain which file or would delete
  • Spot-check is something a human can do in 2 minutes
  • Pass/fail dated (Asia/Shanghai) sits next to the artifact

Step 2 — Write an Allowed list (read vs write)

Separate Read (look only) from Write (change something). For beginners: start Read-only for folders; Write only for a sandbox folder you created.

From this Done checklist, propose Allowed permissions.
Done: [paste]
Format:
- Read: ...
- Write: ...
- Ask me first: ...
- Never: ...
Prefer safest defaults for a beginner.
Example

Filled prompt

Live task this week (not a toy): apply this guide's Goal — write a clear finish line and a short allow-list before any agent touches files or apps.
Fill with your real artifact names/files. Constraints: no invented facts; label Unknowns; stop before send/delete/pay.
Return the working prompt + one pass/fail check I can mark in a minute.

What good output looks like

  • Write limited to one new file or one folder
  • Delete is Never or Ask me first
  • Read scope is narrow (one folder), not whole disk

Step 3 — Paste both into the agent’s instructions

In ChatGPT / Claude / your agent UI: put Done + Allowed at the top of instructions or the first message. Say: “Do not expand scope without asking.”

Combine into agent instructions (≤200 words):
Done: [paste]
Allowed: [paste]
Add: Ask before any write outside Allowed. Stop if unsure.
Example

Filled prompt

Job: weekly customer-apology drafts from Zendesk export.
Write Done (4 boxes), allow-list ≤5 tools, never-do (3), human gate before send.
Context: we may draft in Gmail; we must never auto-send or invent refund amounts.
Return a one-page card only—no agent run yet.

What good output looks like

  • Instructions fit on one screen
  • Explicit stop / ask rules
  • No vague be-helpful line that overrides limits

Save this card

Card: Done checklist · allow-list · irreversible behind human gate. Missing either → refuse the run.

1. What I shipped: ___ 2. What I refused: ___ 3. What I will never-do again: ___ (date: Asia/Shanghai)

  • [ ] Tools/writes paused
  • [ ] Refuse-line dated
  • [ ] Artifact path saved
  • [ ] Next link chosen (or stop)

Before you start

Write Done (checkboxes) and the human gate on paper before any run for “write a clear finish line and a short.” Allow-list tools; leave send/delete off. Guide: agents-define-done-and-permissions. If you only have a blank practice box, stop and fetch the real job first.

  • Goal in one line: ___
  • Allowed tools: ___
  • Forbidden: ___
  • Human approve needed for: ___

How to use this week

This week: before any agent run that can write or send, paste a Done checklist and an allow-list. Refuse the run if either is missing. Log one dated pass/fail for “define Done + permissions”

Close the week by quoting aloud the refuse line you actually used once. If you cannot, rerun Practice now and date the pass/fail (Asia/Shanghai).

Mid-week cost check: if this habit disappeared tomorrow, what breakages would you notice in 24 hours? Write one line (Asia/Shanghai).

Red flag

Blank allow-list or “write anywhere”; Done that only says “looks good”; run starts before both lists are pasted—refuse.

Pitfalls

PitfallFix
“Make it better” with no Done listForce observable checkboxes
Write access to entire Drive/DesktopSandbox folder first
Permissions buried in chat historyPin them in instructions
Example

Filled prompt

Job: write a clear finish line and a short allow-list b
Prior fail: [what missed the check]
Repair: keep the same Done; add one constraint + one concrete example from my notes.
Boundaries: draft only; label Unknowns [?]; no silent send.
Return: fixed output + the one-minute pass/fail I will mark.

What good output looks like

  • Names the prior fail without inventing a new job
  • Adds one constraint tied to “write a clear finish line and a short allow-list b”
  • Still draft-only / Unknowns labeled
  • Pass/fail: you mark green on the same checklist (agents-define-done-and-permissions #4)

Practice now

Live card: fill Done (4) · allow ≤5 · never-do (3) · gate for a job on your calendar this week—photo or path on Save (Asia/Shanghai).

Today: pick one recurring job. Write: Done (4 boxes) · allow-list (≤5 tools) · never-do (3 lines) · human gate before leave-the-system. Refuse to start the agent until the card exists.

Today card (copy) — agents-define-done-and-permissions: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

  • [ ] Artifact path written
  • [ ] Refuse example would say "yes"
  • [ ] Dated Asia/Shanghai

One sentence takeaway

Done checklist and allow-list before tools; irreversible behind a human gate. This week: one dated pass/fail on a live run (Asia/Shanghai).

Next

Plan → act → check · Human-in-the-loop checkpoints · Failure modes and retries · Tools and approvals · Least-privilege checklist

Pair with today: open the first Next link (agents-plan-act-check-loop) only after you dated a refuse-line for this guide (Asia/Shanghai).

← Back to home · Learn · Tools · Compare · Start