01AI has rewired how document work is structured

Traditionally, document work distributed time like this:

After AI, the structure looks like this:

What stands out is that "time spent writing" has shrunk and "time spent thinking and judging" has grown. AI produces a first draft in a minute. The freed time redistributes into "how to structure, how to judge, how to explain".

In the AI era, document work is not the delegation of "writing". It is the extension of "thinking".

02The essence of diagrams — not "drawing what words cannot say"

There is a common misunderstanding about diagrams: "diagram what is hard to express in words" — this is wrong. A proper diagram is "an aid to put what can be said in words into the head faster".

Order matters.

  1. First, reach the state where you can fully explain it in words
  2. Then think about which part of that explanation shortens with a figure
  3. Diagram only that part

The reverse order — "I cannot explain well in words, so I will draw it" — lets the diagram cover the gap in verbalization. The reader will feel "I sort of get it", but if the author themselves cannot explain, the diagram is actually explaining nothing.

Test for a diagram: can the author explain, in 30 seconds, "why this structure, why this arrow, why this order"? If not, the diagram came too early. Finish verbalizing first, then diagram.

03Four common diagram patterns and how to prompt AI

Pharma documents return repeatedly to four patterns. Each has its own AI instruction template.

PatternUseAI instruction example
HierarchyRegulatory pyramids, org charts, principle layers"Arrange the five elements below in a top-to-bottom hierarchy. Output in Mermaid syntax. One-line role for each layer."
FlowDecision processes, work flows, patient journeys"Render the four steps as a left-to-right flowchart. Mark any branches explicitly."
ContrastNG/OK, before/after, competitive"Compose the following two columns as a contrast table. Correspondence between rows must be clear."
RelationStakeholder maps, cause-effect, concept maps"Render the relationships among the six elements as a Mermaid graph. Label each edge with the type of relation."

When prompting AI for diagrams, the trick is to "constrain output to Mermaid or SVG code". Mermaid output drops into Mermaid Live Editor for rendering. SVG embeds directly into HTML. The modern usage is not "make an image" — it is "draw out a diagram in a representable text syntax".

04Three principles for becoming someone who can explain

In the era when AI writes the first draft, the differentiator collapses onto "can you explain it?" Three principles.

Principle 1 — Hold "why" at three depths

For each section, prepare three levels of "why":

With three depths, any question can be answered by shifting one layer. "Someone who can explain" is someone who can move freely between Why levels.

Principle 2 — Separate "AI wrote it" from "I judged it"

People who present AI output as if they wrote it never grow explanation muscle. The discipline that grows judgment is "this part AI drafted; this is what I decided about it" — kept distinct internally.

Concretely, on your own notes, tag each section (AI / Me). The final presentation can blend them, but internally always separate.

Principle 3 — Train "cutting" with AI's help

Most documents are not read because they are too long. AI makes length easy. So in the AI era, the differentiator becomes "the ability to cut". Ask AI "cut this to 30% of its essence" — over and over. Observing what remains and what dies clarifies, inside you, "what I truly want to say".

"Cutting" is not AI's hardest task. The hardest part is the human side, after cutting, deciding "this is enough". The explanation muscle of the AI era lives in the weight of that post-cut judgment.
Claude 4.7 + 4.8

What Claude 4.8 makes possible

Anthropic's latest model, Claude 4.8 (released May 28, 2026), enables document workflows that were previously unreachable.

  • Long-context retention: load the package insert PDF, the trial protocol, and prior materials simultaneously, and draft with consistency across all of them
  • Automated 4-layer regulatory check: understands the Pharma Act §§66/68 → Appropriateness Standards → HanteiG hierarchy and can run the 4-layer pass internally
  • Dynamic Workflows: from one instruction, up to 1,000 sub-agents draft in parallel. Ten MR materials across ten products can be drafted concurrently in an hour
  • Diagram structure proposal: from text drafts, picks the optimal diagram pattern (hierarchy / flow / contrast / relation) and outputs Mermaid syntax
  • Effort Control: /effort xhigh maximizes reasoning depth. For materials requiring complex interpretive coherence, it approaches the quality of human pre-review
  • Honesty / adversarial simulation: 4.8 scores higher on honesty than the previous generation. Easier to have AI itself generate sharp counterarguments to its own output

Even with Claude 4.8, the final judgment is human. "What to keep, what to drop" — there sits the material reviewer's expertise. AI multiplies the options for judgment, but the judgment itself stays with the human.

05Walkthrough — building one diagram with AI from scratch

Concrete example: building the five-tier pyramid from Material Review Vol. 2 with AI from zero.

Step 1 — Verbalize in your own head

First, bullet out what you want to convey, using only your head.

# Verbalization notes
- Five regulations relate to pharma advertising
- All must be honored
- But the risk of violation differs
- Top to bottom: Pharma Act, Appropriateness Std, HanteiG, JPMA Code, Internal SOP
- Higher = larger consequences if breached
- Lower = finer operational detail
- So a layered structure matches field experience

Step 2 — Have AI propose options

Prompt:
Based on the verbalization above, propose three diagram options:
(1) Pyramid (stacked hierarchy)
(2) Concentric rings (inner → outer)
(3) Flowchart (top → bottom)
Three lines of pros/cons for each.

Step 3 — You choose, and write the why

From AI's proposals, pick the pyramid. Write why the pyramid in one line: "the visual intuition of 'larger risk = physically higher' matches operations".

Step 4 — Have AI write the Mermaid (or HTML) code

Prompt:
Implement the pyramid in HTML + CSS:
- Five layers L1–L5, top to bottom
- Each layer carries the regulation name + a one-line role
- Higher layers are darker (suggesting larger risk)
- Responsive (does not break on mobile)

Step 5 — You write the explanation prose

Once the figure is done, the prose around it you write yourself. Because that prose carries "why this figure". Letting AI write it produces generic copy and erases the document-specific intent.

The result: only steps 1 and 5 are human work, but those are the core of explanation muscle. Steps 2–4 AI handles fast. Total document time shrinks, but human judgment time does not. The judgment time instead concentrates on raising its quality.

06Four common traps

07Three habits for tomorrow

  1. Spend the first 5 minutes of document work on "designing what to make AI write" — do not let AI write immediately; set structure and axis first
  2. Before diagramming, reach the state where you can fully explain in words — diagrams are aids, not substitutes for verbalization. Do not reverse the order
  3. Before submission, write the 3-depth "Why" on one page — Q&A preparation finishes at the same time. Pre-presentation explanation time shrinks
In closing

The era when AI writes drafts is not the era where "people who can write" lose value. It is the era where "people who can articulate why it is written that way" stand out more than ever.

When AI began drafting, three kinds of people emerged. First — those who hand off the AI output as-is. Second — those who edit the AI output and use it. Third — those who, after letting AI write, can articulate "why it became this". The first two may, before long, be replaced by AI itself. Only the third is "someone who can explain" in the AI era.

Vol. 5 (in preparation) takes up data interpretation in dialogue with AI — the intersection of analysis and AI use.