AI Assistant Playbook (Starter Prompts, Handoffs & Guardrails)

Purpose and how to use this playbook

This playbook helps teams get reliable, consistent results from AI assistants while preserving accountability, quality, and appropriate human oversight. It contains starter prompt templates you can copy, rules for when to hand outputs to humans, simple sensitivity tags, suggested attribution language, and a few example flows you can adapt to your context.

Quick principles

• Design prompts with clear roles and expected outputs (what, format, tone, constraints).
• Define when a human must review—don’t let AI be the final decision-maker on high-risk items.
• Tag sensitivity explicitly so downstream agents and people handle outputs correctly.
• Log who requested, which model/version was used, and who reviewed any high-risk outputs.

Common sensitivity tags (use these canonical labels)

• PUBLIC — safe to share externally without restriction.
• INTERNAL — for team/org use; not for public distribution.
• SENSITIVE — contains personal data, strategic plans, or market-sensitive info; requires human review before sharing.
• RESTRICTED — legal, regulatory, or safety-critical content; review by designated approver required.

Metadata to capture for each AI interaction

• Requester name/role
• Purpose / Hunger being solved
• Model / temperature / system prompt used
• Sensitivity tag
• Reviewer name (if applicable) and review outcome
• Timestamp and version or prompt template ID

Starter prompt templates (copy & adapt)

Meeting summary (concise, action-focused)

System / Context: "You are an executive summaries assistant. Produce concise, accurate outputs suitable for busy leaders."

Prompt:

"Summarize the meeting transcript below. Produce: 1) A one-paragraph objective summary (2–3 sentences); 2) A short list of decisions made with decision owners; 3) Action items with owners, due dates (if mentioned) or suggested due dates; 4) Open questions. Output as clear bullet sections labeled: Summary, Decisions, Action Items, Open Questions. If you detect sensitive personal or legal content, tag the result as SENSITIVE and mark for human review."

Input placeholder: {{meeting_transcript}}

Decision extraction (structured)

Prompt:

"From the text below, extract explicit decisions and inferred decisions. For each decision include: decision statement, rationale (one sentence), stakeholders, effective date (if any), and confidence level (High/Medium/Low). If a decision involves regulatory, legal, safety, or financial impacts, tag as RESTRICTED and stop — flag for human review."

Input: {{discussion_notes}}

Research brief (evidence-first, cited)

Prompt:

"Produce a research brief (300–500 words) answering: {{research_question}}. Structure the brief: Key Findings, Supporting Evidence (with short citations or URLs), Implications for our work, and Recommended Next Steps. Mark any uncertain claims clearly as 'uncertain' and include suggestions for verification. Use neutral tone suitable for an internal decision memo."

Input: {{background_context}} + {{source_links}}

Draft email (polished, role-aware)

Prompt:

"Write a {{tone}} email to {{recipient_role}} about {{topic}}. Include a clear subject line, a brief context sentence, the main request or information, and a proposed closing with next steps. Keep it to 150–250 words. If the email contains sensitive or personal data, tag as SENSITIVE and include a note identifying the sensitive elements."

Placeholders: {{tone}}, {{recipient_role}}, {{topic}}, {{key_points}}

Handoff rules (who does what and when)

Use these simple, enforceable rules to keep responsibility clear:

• Every output must carry a sensitivity tag. If not set automatically, the requester must assign it when launching the task.
• PUBLIC & INTERNAL outputs may be used without mandatory review, but the requester remains accountable.
• SENSITIVE outputs require review by a named reviewer before distribution outside the originating team.
• RESTRICTED outputs require approval by a designated approver (legal/compliance/safety) and cannot be published or used for decisions until approved.
• When a task leads to a decision recommendation, the recommendation must be accompanied by evidence and a named human approver before execution.

Suggested reviewer roles and review scope

• Content owner (subject-matter expert): accuracy and completeness.
• Data/privacy owner: PII, data-sharing risks, and anonymization needs.
• Legal/compliance: regulatory or contract risks (RESTRICTED items).
• Manager or approver: alignment with strategy and authority for execution.

Example agent flow (meeting to action)

Flow: Meeting capture agent → Summarization agent → Decision extraction agent → Human reviewer → Task creation in tracking tool.

Notes for implementers:

• Each step writes metadata including model, prompt template ID, sensitivity tag, and timestamp.
• If any agent tags an output SENSITIVE or RESTRICTED, the flow pauses and notifies the designated reviewer(s).
• Keep the human in the loop for action assignment and final decision approval.

Attribution and messaging

Use simple, transparent attribution for outputs that will be shared:

"Drafted by AI Assistant (model: {{model_name}}). Reviewed by: {{reviewer_name}} (if applicable)."

Avoid implying the AI is an independent expert. Always pair with a human reviewer when the output influences decisions or external communications.

Customization & operationalizing this playbook

How to adapt:

• Replace placeholders ({{...}}) with your team’s terms and integrate with ticketing or task tools so action items create tasks automatically.
• Define reviewers for SENSITIVE and RESTRICTED categories in an accessible team roster.
• Assign prompt-template IDs and store them so you can track performance and iteratively improve prompts.

Suggested metrics to watch

• Percentage of AI outputs tagged SENSITIVE or RESTRICTED.
• Time from AI draft to human approval for reviewed outputs.
• Reviewer correction rate (how often reviewers change AI recommendations).
• User satisfaction rating for AI-drafted items (simple 1–5 after review).

Starter checklist before publishing or acting on AI output

1. Verify sensitivity tag is present and appropriate.
2. Confirm necessary reviewers were notified and have approved if required.
3. Check evidence and citations for research or decision outputs.
4. Add attribution line showing AI model and reviewer (if reviewed).
5. Record metadata (model, prompt ID, requester, timestamp) for auditability.

Next steps & recommended improvements

Start by piloting these templates with a single team or use case (e.g., meeting summaries). Collect reviewer feedback and the suggested metrics for 4–6 weeks, then iterate on prompts, reviewer assignments, and tags. Consider packaging the most-used prompts and reviews into a team-owned collection so they can be copied, versioned, and tailored across the organization.

Want help tailoring these templates to your team’s roles, systems, or compliance requirements? Consider creating a site-specific copy of this playbook and mapping reviewer roles, approval thresholds, and task integrations to your operating practices.