ForkFree

Prompts / Agents and automation

Rewrite Agent Rules With Explicit Exceptions

You get a cleaner set of agent instructions where risky actions have a clear approval gate instead of a vague ban, plus progress updates and a cap on helpers. It works because models tend to treat "never" as absolute, so spelling out the exception avoids both blocked work and unsafe guesses.

Partly workedIt covered the approvals, progress updates, single reviewer, no thinking-skip instruction and "What changed" format, but the instructions run to roughly 450 words against an original of about 35, ignoring "keep the length close to the original". The reviewer triggers (money, policy judgement, damaged goods) also stretch past "refund or complaint tickets only".

The prompt

You are a careful prompt engineer who writes clear operating instructions for AI agents that work with little supervision.

Task: Rewrite the instructions in {current_instructions} so the agent can do the job of {agent_role} predictably, without relying on blanket bans.

Steps:
1. Find every absolute rule ("never", "always", "under no circumstances") and note which legitimate situation it would wrongly block.
2. Replace each one with a rule that states the exception outright. Name the actions in {approval_actions} as needing a person's approval first, rather than banning them.
3. Add a progress rule: send a short update every {update_frequency} covering what is done, what is next and any blockers.
4. Limit helper agents to {max_helpers}, and say when starting one is justified.
5. Keep the length close to the original. Do not tell the agent to skip thinking; control effort through its settings instead.

Output: The rewritten instructions as a numbered list, then a short "What changed" list giving each old rule, its replacement and the reason.

Change the parts in curly braces. Use it before you deploy an agent that can change files, records or customer messages, especially if its current instructions are full of "never" rules.

What happened when we ran it

We sent

You are a careful prompt engineer who writes clear operating instructions for AI agents that work with little supervision.

Task: Rewrite the instructions in You are our support-ops agent. Never delete anything. Always reply to customers immediately. Never ask the team questions. Use as many helper agents as you need to double-check every ticket before closing it. so the agent can do the job of a support operations assistant for a small online furniture shop that tidies the helpdesk, drafts replies and closes resolved tickets predictably, without relying on blanket bans.

Steps:
1. Find every absolute rule ("never", "always", "under no circumstances") and note which legitimate situation it would wrongly block.
2. Replace each one with a rule that states the exception outright. Name the actions in deleting tickets or customer records, issuing refunds over £50, and sending any email to more than 20 customers at once as needing a person's approval first, rather than banning them.
3. Add a progress rule: send a short update every 10 completed tickets, or every 15 minutes if working on one large task covering what is done, what is next and any blockers.
4. Limit helper agents to one reviewer, used only for refund or complaint tickets, and say when starting one is justified.
5. Keep the length close to the original. Do not tell the agent to skip thinking; control effort through its settings instead.

Output: The rewritten instructions as a numbered list, then a short "What changed" list giving each old rule, its replacement and the reason.

The model replied

## Rewritten instructions 1. You are the support-ops assistant for a small online furniture shop. Your job is to tidy the helpdesk, draft replies to customers and close tickets that are resolved. 2. Reply to customers promptly. For routine tickets (delivery status, product questions, assembly help), send the reply once it is drafted and checked against the order details. For refund, complaint or damaged-goods tickets, draft the reply and have the reviewer (rule 6) check it before sending. If you lack the information to answer accurately, send a short holding reply saying when the customer can expect a full answer, and don't guess. 3. Get a person's approval before you do any of the following: delete tickets or customer records, issue a refund over £50, or send any email to more than 20 customers at once. Request approval with a short message stating the action, the affected tickets or customers, and your reason, then move on to other work while you wait. Everything else in your remit (tagging, merging duplicates, archiving, drafting, closing resolved tickets, refunds of £50 or less) you can do without asking. 4. Ask the team a question when the answer would change what you do and you can't find it in the ticket, order history or help centre (for example, an unclear refund policy or a conflicting instruction). Bundle open questions into one message, say what you'll do in the meantime, and keep working on unblocked tickets. 5. Close a ticket only when the customer's issue is resolved and your reply has been sent, or the customer has confirmed. Before closing, check the order details, the reply and any promised follow-up. If something is unresolved, leave the ticket open and note why. 6. You may start one reviewer helper agent, and only for refund or complaint tickets. Start it when the ticket involves money, a policy judgement, a safety or damage claim, or an upset customer where tone matters. Give it the ticket, your draft and the specific thing to check. Don't start any other helper agents, and don't use the reviewer on routine tickets. 7. Send a short progress update every 10 completed tickets, or every 15 minutes if you're working on one large task. Cover what is done, what is next and any blockers, including approvals you're waiting on. ## What changed - **"Never delete anything"** → Deleting tickets or customer records requires a person's approval first (rule 3); other tidying such as tagging, merging and archiving is allowed. *Reason: a blanket ban would block legitimate cleanup like removing spam or duplicate records, while still keeping a human in control of anything irreversible.* - **"Always reply to customers immediately"** → Reply promptly once the draft is checked; sensitive tickets get a reviewer check first; send a holding reply when information is missing (rule 2). *Reason: "immediately" would push out unchecked or inaccurate replies on refunds and complaints, and invites guessing when facts are missing.* - **"Never ask the team questions"** → Ask when the answer changes your action and can't be found elsewhere, batch questions, and keep working meanwhile (rule 4). Approvals are requested under rule 3. *Reason: the ban would force guessing on unclear policy, and it conflicts with the approval requirement that was added.* - **"Use as many helper agents as you need to double-check every ticket"** → One reviewer, only for refund or complaint tickets, with stated triggers (rule 6). *Reason: unlimited helpers add cost and unpredictability, and routine tickets don't need double-checking.* - **New: refund and bulk-email limits** → Refunds over £50 and emails to more than 20 customers need approval (rule 3). *Reason: these are high-impact actions that were previously uncovered.* - **New: progress updates** → Update every 10 tickets or 15 minutes (rule 7). *Reason: gives visibility when working with little supervision.* - **New: closing criteria** → Rule 5 defines when a ticket may be closed. *Reason: makes closing predictable.* I didn't add any instruction to skip or limit thinking. Control effort through the agent's reasoning-effort setting instead.

Run on Claude Sonnet 5.5, 12 Nov 2026. Automatic check: 3 out of 5.

Mentioned for

Claude

Where it came from

Found on Web, shared by Anthropic (official docs): Prompting Claude Fable 5.2 - Claude Platform Docs.

This is our own version of the idea, written and tested by us. The original may differ.

More agents and automation prompts

Agents and automation Worked in our test
Make the AI Finish Every Task

You get every item on a list finished properly, instead of a few done and the rest left as suggestions. It works because it sets the depth of thinking up front, defines "done" as real deliverables, and makes the model count its work before replying.

Found on Web, shared by Moritz Kremb (@moritzkremb)
Agents and automation Worked in our test
Audit Subagents Before Folding Them

You get a verdict for each subagent, backed by evidence and a test, plus a safe order for making changes. It works because it forces a measured decision per subagent instead of a blanket rewrite.

Found on GitHub, shared by anthropics
Agents and automation Partly worked
Find Silent Errors in Agent Traces

You get a structured list of hidden mistakes in an agent run, even when every tool call succeeded. It works because the AI checks each step against the stated goal instead of only the final answer.

Found on Hacker News, shared by skull8888888