A 30-page policy that nobody reads protects nobody. These prompts compress a policy into one page of “what changes for you,” keep a link to the binding full text, and force the model to quote the source instead of inventing it. That last part matters: Stanford RegLab found general models hallucinate on 69-88% of specific legal queries, and even purpose-built legal AI tools still fabricated content 17-34% of the time on hard questions. A summary you can’t trace back to a paragraph is worse than no summary.
TL;DR
- Paste these 12 prompts, swap the bracketed placeholders, and you get one-pagers, diffs, FAQs, and announcement copy in minutes.
- Always upload or paste the actual policy text. A model summarizing from memory will invent provisions.
- Tool fit (June 2026): Claude for accuracy and long docs (1M-token context on Pro $20+), ChatGPT for fast announcement copy, NotebookLM when you need cited summaries across many policy files.
- Every prompt ends with a “quote the source” or “mark policy gap” instruction so you can fact-check in 60 seconds.
Who this is for
HR and People Ops, IT and Security teams, comms leads supporting compliance, and any manager who has to explain a policy change without a legal degree.
Which tool for which job (June 2026)
Policy work is one place where the model choice changes the output quality, because document length and citation discipline matter more than raw “smartness.”
| Tool / plan | Context you can feed | Best for | Notes |
|---|---|---|---|
| Claude Pro ($20/mo) | ~500 pages (1M tokens, Sonnet 4.6 / Opus 4.7) | Long handbooks, careful diffs, plain-English rewrites | Bundles Claude Code + Cowork; Projects keep policy files loaded across chats |
| ChatGPT Plus ($20/mo) | ~320 pages in-app (full 1M only on $200 Pro) | Fast announcement emails, talking points | GPT-5.5 default; US Free tier now shows ads |
| Google AI Pro ($19.99/mo) | 1M tokens (Gemini 3.1 Pro) | Workspace-native drafting, Docs/Gmail tie-in | Formerly “Gemini Advanced” |
| NotebookLM (Free / Plus) | 500K words per source; 50 sources Free, 100 Plus | Cited summaries across many policy docs | Answers cite the exact source passage; runs Gemini 3 |
For anything employees must follow precisely, prefer a tool that shows its source. NotebookLM grounds every answer in an uploaded passage; Claude and ChatGPT will quote when you tell them to, which is why the prompts below ask for it.
Prompt anatomy
Every prompt here carries six parts. Keep them when you write your own:
- Role: who the model plays (HR partner, IT comms lead, analyst).
- Context: the actual policy text, the org, the audience.
- Goal: one deliverable (one-pager, diff, FAQ, email).
- Constraints: word count, must-include fields, reading level.
- Tone: neutral and factual, not preachy.
- Verification hook: “quote the source line” or “mark policy gap” so you can check it.
Placeholders use [brackets]. Swap them, then paste the policy text under the prompt.
12 copy-ready prompt templates
1. Policy to one-page summary
Policy text is below. Audience: [employees / managers / executives].
Summarise to one page: (1) What it is, (2) Why now, (3) What changes
for the reader, (4) Effective date, (5) Who to ask. Plain English,
8th-grade reading level. For each point, quote the source sentence in
parentheses so I can verify. Then paste the policy below this line.Swap: audience.
2. Policy diff for employees
Old policy and new policy are below, labelled OLD and NEW. Show what
changed: (a) Removed, one line each, (b) Added, one line each,
(c) Reworded but same intent. For each, state the impact on daily work.
Skip purely legal restructuring. If a change is ambiguous, flag it
"NEEDS LEGAL REVIEW" rather than guessing.3. FAQ companion
For the policy below, write 8 FAQs employees would actually ask:
boundary scenarios, exceptions, who approves, what happens if I forget.
Plain answers, 35 words or fewer each. If the policy does not answer a
question, write "Policy gap — confirm with [owner]" instead of inventing.4. Manager talking points
Managers must communicate this policy to their team. Provide:
(1) a 60-second verbal explanation, (2) three likely questions plus
answers, (3) the escalation path. Tone: matter-of-fact, not preachy.
Base every answer on the policy text below; do not add new rules.5. New-hire policy briefing
For new hires on day one, distil the five policies they must internalise:
code of conduct, IT acceptable use, security basics, expenses, time-off.
Each: a 2-sentence summary plus a link placeholder [link]. Flag any
policy where the summary loses a rule the new hire must follow.6. Annual policy refresh
This policy needs its annual review. Audit it and output a fix list:
(1) sections referencing old systems or roles, (2) provisions no longer
enforced, (3) new scenarios not covered, (4) anything needing legal
review. Cite the line number or heading for each finding.7. Cross-jurisdictional summary
This policy applies across [regions]. Summarise: (1) universal
provisions, (2) region-specific differences in a table, (3) when to
consult local counsel. Do not bury regional differences in footnotes.
If a region is named but not actually covered, say so.Swap: regions.
8. Plain-English rewrite
Rewrite the policy paragraph below at an 8th-grade reading level WITHOUT
weakening intent. Show original and rewrite side by side. Flag any line
where the meaning may have shifted and mark it "LEGAL MUST REVIEW".9. Compliance Q&A simulator
Stress-test my policy summary below. Ask five tricky questions an
employee might face (e.g., "Can I use my personal phone for work
email?"). Answer each strictly from the policy. If the policy is
unclear, mark the answer "POLICY GAP" instead of guessing.10. Policy notification email
Draft the all-hands email announcing this new policy: (1) what, when,
and why in three sentences, (2) what changes for the reader, (3) where
to read the full text, (4) Q&A office-hours time. 200 words or fewer.
No legalese in the body. Keep the full text as a link, not pasted in.11. Acknowledgement copy
Draft the e-signature acknowledgement language: 50 words or fewer.
It must confirm the reader has read, understood, and will adhere to the
policy. Do not bury any new obligation inside the attestation wording.12. Exception-request template
Some employees will need exceptions. Draft the exception-request
template: (1) what to provide, (2) approval path, (3) timeline
expectation, (4) confidentiality. Avoid vague "discretion of
management" wording; name the approver role.Two-pass workflow that keeps you out of trouble
- Ground it. Upload the policy file (Claude Projects, a ChatGPT chat, or a NotebookLM source). Never let the model summarize a named policy from memory — that is where invented clauses come from.
- Draft. Run the prompt for your deliverable.
- Verify. Use prompt 9 to stress-test the draft, then spot-check the quoted source lines. Anything marked “POLICY GAP” or “NEEDS LEGAL REVIEW” goes to the policy owner.
- Ship. Send the one-pager; keep the binding full text one click away.
This is the same “source-controlled” pattern enterprises adopted in 2026 after a string of fabricated-citation incidents — including South Africa’s draft National AI Policy, withdrawn in April 2026 because its reference list contained AI-hallucinated sources.
Common mistakes
- Summarizing from memory instead of uploading the actual text.
- No verification hook, so you can’t tell invented provisions from real ones.
- Vague audience, so the tone over- or undershoots seniority.
- No word limit, so nobody finishes reading.
- Treating the one-pager as the binding document instead of a guide to it.
- Over-simplifying provisions employees must follow precisely.
How to push results further
- State the audience level explicitly (frontline vs. executive).
- Cap length: one page tactical, three bullets for executives.
- Lead with the decision or action needed.
- Keep the binding full text as a link, never pasted into the summary.
- Read the draft aloud before sending.
- Let the model draft; a human owns the final sign-off.
FAQ
- How long should a policy one-pager be?: One page for the tactical version; three bullets plus a link for executives. If it runs past a page, you are copying the policy, not summarizing it.
- Which model handles long policies best?: As of June 2026, Claude (1M-token context on Pro $20/mo) reads roughly 500 pages in one pass, versus about 320 in-app pages on ChatGPT Plus. For multi-document policy sets, NotebookLM cites the exact source passage.
- Can AI replace the HR partner or compliance lead?: It drafts and templates well; it does not own judgment or legal accountability. A human signs off.
- How do I stop it from inventing provisions?: Always upload the source text and use a prompt that asks the model to quote the source line or mark a “policy gap.” Then spot-check the quotes.
- Should the summary mention risks or penalties?: Yes. Omitting them to seem friendly erodes trust the first time someone is caught off guard.
- Is the AI summary legally binding?: No. The signed, full policy is the binding document. The one-pager is a reading aid that should always link to it.
Related
- SOP drafting prompts
- Executive summary prompts
- Stakeholder update email prompts
- Handover document prompts
- Summarize a Long Policy Document With AI
- Productivity & Office Prompts hub
Tags: #Prompt #Productivity #Policy