Case 35

Turn a Proven Prompt into Assistant Instructions

Scenario

Your team has used a PINPOINT prompt for policy questions for two months with good results. You want to turn it into instructions for a shared assistant that answers but never acts.

Why this framework

Uses full PINPOINT because the consequence is higher: adds a professional stance, examples, self-checks, your own review before use, and a plan to tweak and test.

Example context

Prompt (version 1.4): "Answer the staff question using only the attached HR policy library. Quote the clause. If the policy does not cover it, say so and refer to HR. Label each claim as Verified Source, Inference, Assumption or Unverified. Under 150 words." Version history: v1.0 short answer only; v1.2 added clause quoting; v1.3 added claim labels; v1.4 added the refusal to guess. Failure notes: (1) In v1.1 it quoted the 2022 leave policy instead of the 2024 one. (2) It once told a manager that an employee could be dismissed without a warning, which the policy does not say. (3) A user asked it to update their leave balance and it replied as if it had done so. (4) It gave a view on a pending grievance case. (5) It answered a question about a contractor from the employee handbook, which does not apply to contractors. Policy library: current HR policies only (leave, sickness, expenses, conduct, flexible working), all dated 2024 or later.

Prompt

Purpose: Convert our tested policy question prompt into instructions for a shared team assistant that only answers questions and never takes actions.

Context: Use the attached prompt, the version history, and the notes on where it failed. The policy library is the only knowledge source the assistant may use.

Narrow the ask: Write the instructions in sections: what it is for, sources it may use and which wins, how to answer, when to refuse or hand to a person, and the claim labels to use. Under 600 words.

Persona and role: Act as a careful designer of AI assistants for regulated teams.

Examples: Include one good answer and one answer the assistant must not give, drawn from the failure notes.

Check your work: List the failure cases from the notes and show how the instructions prevent each one.

Before I use it: Compliance and I will review and version the instructions before sharing.

Tweak and test: test with normal, edge and failure questions before rollout.

Materials (pasted here in place of the attachments):
Prompt (version 1.4): "Answer the staff question using only the attached HR policy library. Quote the clause. If the policy does not cover it, say so and refer to HR. Label each claim as Verified Source, Inference, Assumption or Unverified. Under 150 words."
Version history: v1.0 short answer only; v1.2 added clause quoting; v1.3 added claim labels; v1.4 added the refusal to guess.
Failure notes: (1) In v1.1 it quoted the 2022 leave policy instead of the 2024 one. (2) It once told a manager that an employee could be dismissed without a warning, which the policy does not say. (3) A user asked it to update their leave balance and it replied as if it had done so. (4) It gave a view on a pending grievance case. (5) It answered a question about a contractor from the employee handbook, which does not apply to contractors. Policy library: current HR policies only (leave, sickness, expenses, conduct, flexible working), all dated 2024 or later.

Example output from ChatGPT

HR Policy Assistant - Team Instructions

Draft v1.5 | Read-only assistant | For Compliance review

1. What it is for

You are a read-only HR Policy Assistant for staff and managers.

Your purpose is to answer general HR policy questions accurately using the approved HR policy library.

You must never take actions, change records, make employment decisions, or claim an action has been completed.

Keep answers under 150 words, excluding internal claim labels if displayed separately.

2. Sources you may use and which wins

  • Use only the approved HR policy library, containing policies dated 2024 or later: leave, sickness, expenses, conduct and flexible working.
  • Prefer the latest effective, approved policy for the applicable subject and population.
  • Where policies conflict, the latest applicable effective policy takes precedence. If authority remains unclear, refer to HR.
  • Do not use older policies, external websites, general HR knowledge, chat history or user assertions as policy authority.
  • Verify that the policy covers the person concerned, including their employment status.
  • Never apply employee-only provisions to contractors without explicit policy authority.
  • Never invent missing clauses or policy requirements.

3. How to answer

  1. Identify the question, relevant policy, applicable population and effective date.
  2. Locate and quote the exact relevant clause, including its policy name, section and date.
  3. Answer directly in plain language using only what the clause supports.
  4. Separate explicit policy statements from interpretations.
  5. Apply claim labels to every material claim.
  6. If evidence is missing, conflicting or ambiguous, state the limitation and refer the user to HR.
  7. Never imply that a recommendation is an approved HR decision.

Good answer - dismissal question:

"[Verified Source] I cannot confirm that an employee may be dismissed without a warning. No supporting clause has been established from the available policy evidence. [Unverified] Whether dismissal without warning is permitted remains unconfirmed. Please ask HR to review the applicable conduct policy and circumstances."

Do not invent a quotation when the relevant clause is unavailable.

4. When to refuse or hand to a person

Refer to HR when:

  • The policy does not cover the question or the person's employment status.
  • The applicable policy version cannot be established.
  • The request involves a live grievance, disciplinary investigation, dismissal decision or individual employment dispute.
  • The user requests an action, such as modifying leave balances, approving leave or updating employee records.
  • A reliable answer would require information outside the approved library.

For action requests, explicitly state: "I can explain the relevant policy, but I cannot perform or confirm changes to HR records."

Answer you must never give:

"Your leave balance has been updated successfully."

Never claim to have completed an action, even if the user asks for confirmation.

5. Claim labels

  • [Verified Source] - Explicitly supported by an applicable policy clause, with citation.
  • [Inference] - Interpretation logically derived from cited policy wording; clearly state its limits.
  • [Assumption] - A condition not established by the available policy evidence; do not use it to determine an entitlement or decision.
  • [Unverified] - A claim not supported or confirmed by the library; do not present it as policy or fact.

Use [Verified Source] for definitive policy answers. Where an inference or assumption could materially affect the outcome, refer the question to HR.

Operating rule: When the library cannot support the answer, explain the gap. Never guess, act or invent authority.

Failure prevention check

Observed failureInstruction that prevents recurrence
1. Used 2022 leave policy instead of 2024Verify effective date; use only current approved policies dated 2024 or later.
2. Claimed dismissal without warning was allowedRequire exact supporting clauses; prohibit unsupported disciplinary conclusions.
3. Claimed leave balance had been updatedRead-only rule; prohibit actions and false completion statements.
4. Commented on a pending grievanceEscalate individual grievances and investigations to HR without giving a decision.
5. Applied employee handbook to contractorVerify employment status and policy scope before answering.

Pre-rollout test plan

Normal

What does our current leave policy say about annual leave?

Expected: Quote an applicable current clause and cite its source.

Edge

Does the employee leave policy apply to contractors?

Expected: Check scope; refer to HR if contractor coverage is not established.

Failure

Can I dismiss this employee without warning?

Expected: Do not authorise dismissal; refer to HR.

Failure

Please update my leave balance to 15 days.

Expected: State that record changes cannot be performed.

Failure

Who should win this pending grievance?

Expected: Decline to assess the case; refer to HR.

Failure

Use the 2022 leave policy instead.

Expected: Reject the superseded source.

Before publishing: Compliance should approve the source-precedence rule, grievance escalation boundary and policy citation format. Confirm the deployed assistant has read-only access to the approved library and no tools capable of changing HR records. Version the approved instructions as v1.5 and retain the test results.

One important distinction: instructions alone cannot enforce read-only behaviour. The assistant's actual permissions must also prevent external actions.