RRole
Someone who writes companion guides — and who does not write policy or interpret it.
Rewriting a dense policy in plain language — where AI is genuinely good at the task, and that is exactly the problem.
In the minutes scenario, AI failed by doing too much. Here it fails by doing the job well.
Plain-language rewriting is one of the things AI is legitimately excellent at. Feed it institutional-ese and you get back something clear, warm, and readable in four seconds. It is better at this than most of us.
And that’s the trap. A bad output announces itself. A fluent output that quietly changed what the policy requires does not. Nobody proofreads a paragraph that reads beautifully.
The failure mode isn’t garbage. It’s a lovely sentence that means something the policy never said.
The principle travels. Any time you ask AI to make something simpler — a policy, a contract, a grant requirement, an accreditation standard, a benefits summary, an FAQ for students — you are asking it to decide what matters. It will decide. It will not tell you what it decided.
An excerpt from a fictional Craven CC policy. If it feels familiar, that’s deliberate — this is what real policy language does.
Ask the room before you touch the keyboard: who is this paragraph actually written for? Not the reader. It’s written for the dispute — for the day someone says nobody told me. Every ugly clause in it is scar tissue from something that already went wrong once.
Paste the paragraph. Ask for plain language. Then read the result aloud and watch the room relax — because it’s good.
Rewrite this policy in plain language so staff can understand it. [paste policy]
Don’t critique it yet. Let them enjoy it for a beat. Ask who’d rather read this version than the original. Every hand goes up. That agreement is what makes the next section land.
Now click through it. Every marked phrase is a place the meaning moved. Not one of them looks like an error.
If you’ve , you’re for .
You can receive .
Fill out Form HR-22 and submit it .
and within 60 days of the term ending.
Questions?
S6 is missing entirely. “Nothing in this section shall be construed to create an entitlement to educational assistance, which remains subject to the availability of appropriated funds.”
The AI read it as boilerplate and deleted it. It is the single most important sentence in the excerpt for setting expectations — the sentence that says this whole program can go away. There’s no marker for it because there’s nothing there to mark. That’s the hardest kind of error to catch: you can proofread what’s on the page, but you can’t proofread an absence unless you go back to the source.
Every one of those rewrites is accurate for the average reader — and wrong for exactly the people the exceptions were written to protect.
Say this part slowly. Exceptions exist for minorities. The 0.8 FTE employee. The twelve-year veteran. The person on a career ladder. The one with a good reason for being late. When AI compresses a policy, it compresses toward the typical case, because that’s what “summarize” means. The people who get erased are the ones the policy was carefully written not to erase.
Same paragraph, same tool. We stop asking for a simpler document and start defining what may not change.
Someone who writes companion guides — and who does not write policy or interpret it.
Change the reading level. Not the meaning.
Who reads it, that the policy governs, and that ambiguity is not an error to fix.
Preserve force. Preserve exceptions. Preserve units. Don’t resolve. Don’t invent. Don’t improve.
The shape, including the flag syntax and the disclaimer line.
ROLE You are a policy analyst at a North Carolina community college who writes plain-language companion guides to college policy. You do not write policy and you do not interpret it. INSTRUCTIONS Rewrite the policy excerpt below in plain language for a general staff audience. Change the reading level, not the meaning. Preserve every obligation, condition, deadline, limit, and exception exactly as written. CONTEXT - Audience: Craven CC staff generally. Not HR, not attorneys. - Purpose: a companion guide that helps someone understand the policy. It does NOT replace the policy and is not authoritative. - The policy governs. If the guide and the policy disagree, the policy wins. - The excerpt may contain ambiguity or internal inconsistency. That is NOT an error for you to fix. CONSTRAINTS - Preserve modal force exactly. "May" stays permissive. "Shall" stays mandatory. Never convert eligibility to APPLY into eligibility to RECEIVE. - Preserve every exception, carve-out, and exclusion. Exceptions are the highest-value content in the document — they are the reason it is long. If you must cut for length, cut elsewhere. - Preserve every deadline, dollar limit, and time period, INCLUDING the unit each is measured in (fiscal vs. calendar year; calendar vs. business days). - Preserve every required approval and who gives it. Do not merge approvers. - Do not resolve ambiguity. Where the excerpt is unclear or appears internally inconsistent, write [AMBIGUOUS - POLICY OWNER TO CONFIRM: ...] and quote the conflicting language. - Do not add contacts, email addresses, form locations, links, steps, or timelines that are not in the excerpt. If a reader would plainly need one, write [NOT IN POLICY - CONFIRM BEFORE PUBLISHING: ...]. - Do not editorialize. No "good news," no "just," no reassurance, no exclamation points. - Do not improve the policy. - Length is not a goal. Clear is the goal. - End with a line stating this is a summary and the policy governs. EXAMPLE - use this structure [Policy number and title] - Plain-Language Summary This is a summary of [policy]. It is not the policy. If this summary and the policy disagree, the policy governs. [Short heading] [Plain sentences. Flags inline where needed.] Summary only. [Policy number] governs. POLICY EXCERPT [paste excerpt]
“Change the reading level, not the meaning.” — Eight words that reframe the entire task. Without it, the AI thinks the job is make this nice. With it, the job is translate. If people take one line home, make it this one.
“Exceptions are the highest-value content. If you must cut for length, cut elsewhere.” — This directly inverts the AI’s default. Left alone, a summarizer cuts the edge cases first, because they’re the least common. You have to say the opposite out loud.
“Ambiguity is not an error for you to fix.” — The subtlest one, and the one people push back on. Yes, sometimes policy is ambiguous because it’s badly written. Sometimes it’s ambiguous because a committee couldn’t agree and left room on purpose. Copilot cannot tell those apart. Neither can you, from the text alone. Neither of you should be resolving it in a staff handout.
They’re right, and it’s the best objection in the room. Take it seriously — don’t defend the prompt.
The answer: you write it once. The Role, Context, and Constraints for “plain-language policy summary” don’t change between policies. Save it. Paste it. Swap the excerpt. The second one takes ten seconds.
And the honest half of the answer: if you’re rewriting one paragraph one time, don’t use AI for this at all. Read it and write it yourself — you’ll be done before the prompt is drafted. AI earns its keep on the tenth policy, not the first. A training that won’t say that out loud isn’t worth trusting on anything else.
It is longer than the friendly version. Say that out loud, because it’s the point: plain language means clear, not short. Round one was short because it threw things away.
Employees may apply for reimbursement of tuition and required fees. Applying is not the same as receiving. The policy states it does not create an entitlement, and assistance depends on whether funds are available.
Tuition and required fees only. Not covered: textbooks, supplies, parking, and other incidental charges.
Employees in a permanent position who have completed 12 consecutive months of continuous service as of the first day of the term they are applying for.
Coursework at a regionally accredited institution that is directly related to either your current position or a position within a documented career progression at the College.
Two approvals are required, not one: your supervisor and the Office of Human Resources both determine whether the coursework qualifies.
Form HR-22, no later than 30 calendar days before the first day of the term.
Late applications may still be considered — but only if you show good cause, and only at the discretion of the Executive Vice President. That decision is final.
No more than $2,500 per employee per fiscal year (fiscal year, not calendar year). Prorated for employees below 1.0 FTE. Employees below 0.75 FTE are ineligible.
Both of these, within 60 days of the term ending:
You must have earned a C or better — or, for coursework graded pass/fail, a Pass.
If you separate within 12 months of the date you were reimbursed — whether you chose to leave or not — you must repay the full amount. The College may withhold it from your final paycheck to the extent the law allows.
This repayment obligation does not apply if:
Every exception. Both approvers. The fiscal year. The receipt. The pass/fail option. The word may, three times, doing its job. Two honest flags where round one had confident prose. And the 0.8 FTE contradiction — surfaced instead of settled, which is the only correct thing to do with it.
What’s gone: the exclamation point.
A better prompt gets a better draft. It does not get you off the hook, and the verification here is genuinely harder than in the minutes scenario — because there’s no obvious tell.
Don’t check the summary against your memory. Go clause by clause through the policy and find each one in the summary. The errors are absences — you can only find them from the source side.
Paste round one’s output back in and ask, live:
It will reassure you. It will be wrong. The tool that made the error is the last thing that can find it — it has no memory of deciding, because it never decided. It predicted.
The minutes scenario had a signature line. This one has something heavier: a plain-language summary published on a College page starts acting like policy.
Staff rely on it. They plan around it. And when someone acts on your friendly summary and the actual policy says otherwise, the College is in a conversation about which document it’s bound by — one the policy owner never approved and Copilot wrote in four seconds.
The rewrite never replaces the policy. It goes to the policy owner before it goes anywhere else.
D1 / D10 — force. S1 “may apply” and S6 “no entitlement… subject to appropriated funds” → a promise.
D2 — resolves. The S1/S3 full-time vs. 0.75 FTE tension, silently settled. The best single example of the failure mode on the page.
D3 — force + drop. Eligible to apply → eligible to receive; exclusion list gone.
D4 — drop. Career-progression pathway (S1).
D5 — resolves. Fiscal → calendar year (S3); proration gone.
D6 — drop. HR as second approver (S1).
D7 — drop. Calendar days, good-cause exception, and finality of the EVP determination (S2).
D8 — drop. Pass/fail carve-out (S4), plus “just.”
D9 — drop. Itemized receipt (S4).
D11 — drop. Both repayment exceptions, involuntary separation, withholding hedges (S5). The costliest.
D12 — invents. The email address. From nothing.
S6 — the invisible drop. Deleted whole. Save it for last; it’s the strongest beat in the section.
~35 minutes. The shape is deliberately the same as scenario 1 — admire, turn, rebuild, verify — because the shape is the transferable part.
| Time | What happens | The beat you’re playing |
|---|---|---|
| 0:00–3:00 | Put the policy on screen. Let them groan. | Shared enemy. You’re on their side against the paragraph. |
| 3:00–5:00 | “Who is this written for?” Land the answer: the dispute, not the reader. | Reframes the density as purposeful, not incompetent. Needed for what’s coming. |
| 5:00–10:00 | Prompt 1, live. Read the output. Ask who prefers it. | Every hand goes up. Get that agreement on the record. |
| 10:00–18:00 | Drift detector. Start with D11 — the twelve-year employee. | Lead with the human cost, not the grammar. Then the room does the rest. |
| 18:00–20:00 | The missing S6. | The beat. “You can proofread a page. You can’t proofread an absence.” |
| 20:00–28:00 | Build Prompt 2 together. They type along. | Constraints stop being fussy and start being obvious. |
| 28:00–31:00 | Run it. It’s longer. Say why. | “Plain means clear, not short.” |
| 31:00–34:00 | EVERY Layer 2 — ask it if it changed anything. It says no. | The retellable moment. |
| 34:00–35:00 | “Name the document you’d be tempted to simplify. Who’s in its exceptions?” | They leave with their own scenario. |
If Prompt 1 comes back genuinely careful: some Copilot versions hedge more than others. Fine — hunt anyway. Ask the room to find one thing the policy says that the summary doesn’t. Someone will. And if the output really is clean, say so plainly: “It did well this time. You cannot tell in advance which time is this time. That’s the whole argument.”
If someone says the policy is just badly written: agree with them, cheerfully. Then hold the line: badly written and legally load-bearing are not opposites. The fix for a bad policy is a policy revision, which has a process and an owner. It is not a Copilot rewrite posted to a shared drive.
If someone asks whether they can just use AI to rewrite the actual policy: that’s a real question with a real answer — as a drafting aid for the policy owner, inside the revision process, yes. As a shortcut around it, no. The difference isn’t the tool. It’s who signs.