Do not ask AI to write a handbook. Build a rules system that can keep growing.

01 / BOUNDARYDecide what must stay out of the handbook.

Without layers, AI turns every source into one encyclopedia. The result is often thick and complete-looking, but no one wants to finish it. We separated the material into three levels:

The public edition contains only transferable judgment. Anything involving clients, cases, internal paths, or account permissions must be removed, even if including it would look “more complete.” Completeness is not the value; the boundary is.

02 / EVIDENCECollect evidence before collecting opinions.

Inputs included old work manuals, management meeting notes, group-chat feedback, former interns' reflections, role profiles, and repeated field notes asking, “Why did this go wrong again?” AI's first job was not polishing. It was labelling:

• Does this problem recur?
• Does it threaten efficiency, quality, confidentiality, or trust?
• Can a reader observe one clearly correct action?
• Is the requirement only for interns, or does it also hold managers accountable?

A rule with no real problem behind it may simply be a manager's frustration. Encouragement that cannot lead to a next action may be no more than a slogan.

03 / STRUCTUREOrganize around a growth path, not the source folders.

Readers are not visiting an archive. A newcomer needs to learn quickly: Why am I here? How do I work with others? What do I learn first? What makes a delivery reliable? What can I ultimately take away?

So the contents became five parts: Purpose, Growing Together, Getting Started, Casework & Delivery, and Direction. Who said what in which meeting, or which old manual contained it, moved backstage. The front stage contains only the reader's next step.

04 / EDITORLet AI be an editor before asking it to be an author.

A one-shot request to generate a whole handbook usually returns tidy, correct, anonymous average prose. A better method uses AI in four distinct passes:

1. Merge: find repeated requirements and group different descriptions of the same problem.
2. Calibrate: flag contradictions, missing boundaries, and language that may damage trust.
3. Challenge weakness: find empty headings, template phrases, slogans, and paragraphs with no actionable meaning.
4. Replace: within the page budget, substitute shorter, more concrete sentences that make consequences visible.

Every addition should explain what it replaced. Adding without removing is the easiest mistake to make when a knowledge base meets generative AI.

05 / VOICETranslate “sharp and capable” into constraints you can test.

A voice cannot be specified by saying “write like this book.” We turned the feeling into checks: keep body paragraphs short; reserve callouts for actual judgment; make headings name a consequence; do not repeat template sentences; cap quotations and tables; run a deletion audit before every revision.

The first says what is here. The second says why to look now. Voice is not imitation of attractive sentences; it is the repeated act of choosing information on the reader's behalf.

06 / DELIVERYTreat Word as a product, not an attachment.

Final text was only half the work. To keep future edits from becoming manual copy-and-paste, we generated Word from a single content source and standardized fonts, hierarchy, tables, page numbers, table-of-contents links, bookmarks, and the public-account QR code. We then exported the PDF through Microsoft Word itself.

Each revision was compared page by page with the prior version. The program first located every changed page, then reviewed each at 100% scale; unchanged pages were pixel-compared to confirm that nothing drifted unexpectedly, while the complete structure still passed automated checks.

A file opening successfully does not mean delivery is complete. The work is complete only when the actual Word document, PDF, and every page can withstand real use.

07 / LOOPMake every cohort of interns an input to the next edition.

A useful handbook should not stop on publication day. New problems enter a simple loop: record recurring issues; propose a candidate rule; have a manager confirm the boundary; move stable actions into a checklist or SOP; reserve the survival guide for durable judgment.

The same mechanism constrains managers. “Be more serious” and “take more initiative” are not enough. Managers must provide observable standards, timely feedback, and gradually harder assignments. Interns are people to develop, but also a mirror: poor mentoring is not always proof that the intern is incapable.

08 / REUSABLE METHODA seven-step workflow another team can reuse.

1. Inventory sources: collect only real records, with source, date, and public/private status.
2. Draw boundaries: separate the survival guide, onboarding checklist, and internal SOPs; redact before processing.
3. Build a problem map: classify by recurrence, consequence, affected person, and correct action.
4. Design the reader journey: organize around “what next,” not “where this source came from.”
5. Use AI in passes: merge, calibrate, challenge weakness, and replace—one problem per pass.
6. Engineer delivery: generate Word and PDF from one source with consistent layout and navigation.
7. Validate and feed back: compare pages, run automated audits, manage versions, and bring new problems into the next edition.

This base instruction can begin each editing pass:

Goal: edit the materials into a handbook people can use to get started quickly, without increasing total length.

First:
1. List repeated, contradictory, vague, and potentially confidential material.
2. Separate durable principles, onboarding checklists, and internal SOPs.
3. For every added paragraph, explain what it replaces or compresses.

Writing requirements:
- Headings should name a judgment or consequence.
- One paragraph, one idea.
- Abstract expectations must become observable actions.
- Put quotations and sources in one consistent section at the end.
- Keep uncertain facts as verification items; never invent the missing detail.

Output: issue list, proposed cuts and changes, revised draft, and risks that still require human judgment.