A senior designer spends a startling fraction of every week writing words, not designing: the design-review writeup, the design-system PR description, the change-log, the crit feedback, the 1:1 prep, the status update. By 2026 Claude and Notion AI will draft all of these in seconds, which is a genuine gift and a genuine trap, because the AI draft is always too long, too generic, and occasionally wrong in a way that matters - a redacted detail invented, a heuristic misapplied, a status that overstates. This lesson teaches the editorial discipline that turns AI drafts into shippable design communication: draft fast, edit down by forty percent, and redact what the AI got wrong before send. The artifact is a pack: a published review post, a PR template, a design-crit prompt, and a manager's weekly template, each one an AI draft you made true.

The Draft Is Not the Deliverable

Start with the mistake that makes AI writing dangerous in a design context: treating the draft as the deliverable. A designer asks Claude to write up a design review, gets back a fluent, well-structured, three-page document in nine seconds, and sends it. It reads professionally. It is also forty percent too long, hedges every real opinion into mush, and includes a confident sentence about the team's accessibility status that is simply not true because the model inferred it. The review went out fast and made the designer look less sharp than a two-paragraph note they wrote themselves would have.

The AI draft is a first draft, in the most literal sense, and the value you add is editorial. The model is genuinely good at the part that is tedious - producing structured, complete, grammatical prose from a pile of notes - and genuinely bad at the part that matters, which is judgment about what to cut, what to sharpen, and what is actually true. So the discipline is not "write with AI." It is "draft with AI, then edit like an editor," and the single most reliable move is the forty-percent cut, because an AI draft is almost always padded, and padding is where the genericness and the false confidence hide.

This is the L2 pattern in the communication domain. Claude and Notion AI supply the speed: a complete draft of any design-team document in seconds. You supply the verification and the cut, because a design communication is a thing people act on, and an AI draft that ships unedited acts on your behalf with the model's padding and the model's occasional fabrication attached to your name.

The Forty-Percent Cut: The Edit That Does the Most Work

The forty-percent cut is the core editorial move and it is worth being specific about what you are cutting, because it is not arbitrary trimming. AI drafts pad in predictable places, and each kind of padding hides a different problem. You cut the throat-clearing introductions that say nothing ("In today's fast-paced design environment, it is important to consider..."). You cut the hedging that turns an opinion into mush ("it could be argued that the spacing might potentially benefit from some refinement" becomes "the spacing is off"). You cut the redundant summaries that repeat the body. You cut the generic filler that could apply to any design at any company, because filler is the tell that the model is averaging instead of saying something specific about this work.

The cut is not just about length; it is about truth and force. A two-paragraph design review that says exactly what is wrong and what to do is more useful, more respected, and more honest than a three-page document that buries one real point in two pages of AI scaffolding. When you cut forty percent, what survives is the part with actual judgment in it, which is the part only you could have written. The model gave you the structure and the completeness; the cut is how you reintroduce the sharpness the model averaged away. A designer who ships the uncut draft has let the model flatten their voice; a designer who cuts forty percent has used the model for speed and kept their voice.

The AI gives you a complete draft and flattens your judgment into it. The forty-percent cut is how you take the speed and reclaim the sharpness. What survives the cut is the part only you could have written.

The Design-Crit Pattern: AI as Structured First-Pass Critic

The most interesting use of AI in this lesson is the design crit, where the model acts as a structured first-pass critic on another designer's work before a human gives the real feedback. The value is that AI is genuinely good at systematic, exhaustive checking against a rubric, which is exactly what a tired human crit often skips. You give Claude a clear rubric - Nielsen's usability heuristics, the relevant WCAG 2.2 criteria, and your team's design-system rules - and a screen, and it produces a thorough first-pass critique that flags every heuristic violation, every contrast miss, every off-system component, systematically, without the fatigue or politeness that makes human first-passes incomplete.

But the crucial discipline is that this is a first pass the human designer then edits before sharing, never feedback the AI sends directly. The model is good at catching the systematic, rule-based issues and bad at the judgment a real crit needs: which issues actually matter for this user and this context, which "violations" are intentional and correct, how to frame the feedback so it helps rather than deflates the person receiving it. An AI crit sent raw to a junior designer is a wall of un-prioritized, tone-deaf nitpicks that damages more than it helps. The human takes the AI's exhaustive first pass, discards the issues that do not matter, prioritizes the ones that do, adds the contextual judgment the model lacks, and reframes it in a human voice. The pattern is AI-for-coverage, human-for-judgment-and-care, and the design-crit prompt that encodes the rubric is one of the artifacts.

The PR Description and Change-Log Pattern

Design-system PR descriptions and change-log entries are a perfect AI-draft target because they are structured, repetitive, and tedious, which is exactly the kind of writing the model produces well and you produce slowly. A good design-system PR description explains what changed, why, what consumers need to do, and whether anything is breaking. Claude can draft all of that from your notes and the diff in seconds, in a consistent structure, which is a real time saver on a chore that designers chronically under-document.

The verification here is specific and non-negotiable: you check that the PR description accurately describes what actually changed, because a PR description is a promise to every consumer of your design system, and an AI draft can confidently describe a change that is subtly different from the real one. If the model says "the button padding increased from 12 to 16" and it actually went from 12 to 14, that error propagates to everyone who reads the PR instead of the code. So you draft fast, then verify every factual claim against the actual diff, and you cut the generic preamble the model adds. The PR template you build encodes the right structure - summary, motivation, breaking changes, migration notes - so the model fills a known-good skeleton and your verification has a checklist. The template is the second artifact, and it makes every future PR both faster to draft and easier to verify.

The Manager's Pattern: Notion AI With the Errors Redacted

For the design manager, the same discipline applies to a different set of documents: 1:1 prep, OKR check-ins, and design-team status updates, which Notion AI will draft from your notes in seconds. The time savings are real - a manager who used to spend an hour assembling a status update can draft it in five minutes - but the stakes are different and the redaction discipline is the heart of the pattern, because these documents go to reports, to peers, and to leadership, and an AI error in them is not just sloppy, it is a misrepresentation with your management judgment attached.

The redaction move is explicit and it is the manager's version of the verification tax: before any AI-drafted management document is sent, you read it specifically hunting for the parts the AI got wrong or invented, and you redact them. Notion AI drafting a status update will confidently summarize a project as "on track" when your actual read is "at risk," because it pattern-matched to the optimistic default. It will invent a detail about a report's progress it has no basis for. It will smooth a nuanced OKR status into a green that should be yellow. None of these is malicious; all of them are the model averaging toward the plausible and pleasant. The manager's job is to catch every one before send, because a status update is a commitment and a 1:1 note is a record, and the model's pleasant inaccuracies become your inaccuracies the moment you hit send. The manager's weekly template - with the redaction step built into it as a required pass - is the fourth artifact.

Why the AI Errors Cluster Where They Do

It is worth naming why AI writing fails in these specific ways, because knowing the pattern makes the verification targeted instead of paranoid. The model is optimizing for fluent, plausible, agreeable text, so its errors cluster in three places. First, it pads, because more text reads as more thorough, so the forty-percent cut targets the padding. Second, it smooths toward the pleasant and the average, so a yellow becomes a green and a sharp opinion becomes a hedge, which is why you hunt for overstated optimism and re-sharpen flattened judgment. Third, it fabricates plausible specifics, inventing a number or a status it has no basis for, because a confident specific reads better than an admission of uncertainty, which is why you verify every factual claim against the source.

These three failure modes - padding, smoothing, and fabrication - map directly onto the three edits this lesson teaches: cut the padding, re-sharpen the smoothing, verify against fabrication. You are not reading the whole draft with equal suspicion; you are reading it knowing exactly where the model's optimization for pleasant fluency will have introduced length, false agreeableness, and invented specifics. That targeting is what keeps the verification fast enough to preserve the time savings. An AI draft you edit blindly takes as long as writing it yourself; an AI draft you edit knowing the three failure modes is genuinely faster and genuinely true.

The Three-Pass Edit That Keeps Verification Fast

The single most useful operational habit in this lesson is to stop reading an AI draft as one undifferentiated document to be vaguely improved, and to read it instead as three known failure modes to be hunted in three targeted passes. Reading the whole draft with equal suspicion is what makes AI editing feel as slow as writing from scratch; reading it knowing exactly where the model's errors cluster is what keeps the time savings real. The three passes map one-to-one onto the three failure modes, and running them in order is the whole technique.

The first pass is the padding pass. You cut throat-clearing introductions, hedging that turns an opinion into mush, redundant summaries that repeat the body, and generic filler that could apply to any design at any company, aiming for roughly a forty-percent reduction and stopping when only judgment-bearing content remains. This pass is fast because padding is visible on sight once you are looking for it, and it does double duty: cutting the padding also exposes the smoothing and the fabrication that were hiding inside it. The second pass is the smoothing pass. You hunt every assessment, status, and recommendation that reads pleasantly agreeable - the yellow rendered as green, the sharp opinion softened to a maybe - and you re-sharpen each one to your actual read. The third pass is the fabrication pass. You enumerate every concrete factual claim - every number, status, behavior, and attribution - and verify each against its source: the diff for a PR, your notes for a status update, the real video for a research quote. Anything the source does not support gets corrected or cut.

Three directed passes are dramatically faster than one suspicious read, because each pass looks for one specific thing and ignores everything else, which is how human attention actually works well. The padding pass does not stop to verify facts; the fabrication pass does not stop to trim prose. By the end, the document is concise (padding gone), honest (smoothing reversed), and true (fabrication caught), which are exactly the three properties an AI draft lacks and a shippable communication needs. The same three passes apply whether the document is a design review, a PR description, a crit, or a manager's status update; only the emphasis shifts - a PR weights the fabrication pass heavily against the diff, a manager's update weights the smoothing pass heavily against the real status. Learn the three passes as a single reflex and you get the model's speed on the drafting and your own judgment on the editing, which is the entire deal this lesson is built to make, and the deal that keeps your name attached to communication that is short, sharp, and true rather than long, bland, and occasionally false.

Putting It to Work This Week

Build the pack this week from real documents. Take your next design-review writeup, draft it with Claude, and cut forty percent, targeting throat-clearing, hedging, and generic filler until only the judgment survives; publish that as the review post. Build a PR template with the summary-motivation-breaking-changes-migration structure, draft your next PR into it, and verify every factual claim against the diff. Write a design-crit prompt that encodes Nielsen heuristics plus WCAG 2.2 plus your design-system rules, run it as a first pass on a real screen, and edit it down to prioritized, kindly-framed feedback before sharing. And if you manage, build the manager's weekly template with the redaction pass built in, and use it for your next status update, hunting specifically for the optimistic overstatements and invented details before send.

You will know the discipline has landed when your AI-drafted communications are shorter, sharper, and more trusted than what you used to write by hand, because you got the model's speed and kept your judgment. The model writes the complete draft in seconds, which is the gift. The forty-percent cut, the rubric-based crit edited for care, the PR verified against the diff, and the redacted manager's update are the four moves that make the draft true and make it yours. In a year when everyone's documents are getting longer and blander because they ship the AI draft uncut, the designer who edits is the one whose words still carry weight.

Key Takeaways

• The AI draft is a first draft, not the deliverable: the model is good at producing structured, complete prose and bad at judgment about what to cut, sharpen, and verify. Draft with AI, then edit like an editor, with the forty-percent cut as the core move.
• The forty-percent cut targets predictable padding - throat-clearing intros, hedging that mushes opinions, redundant summaries, generic filler - and what survives is the judgment only you could have written. It is about truth and force, not just length.
• The design-crit pattern is AI-for-coverage, human-for-judgment-and-care: give Claude a Nielsen-plus-WCAG-2.2-plus-design-system rubric for an exhaustive first-pass critique, then edit it down to prioritized, contextual, kindly-framed feedback before sharing. Never send the raw AI crit.
• For PR descriptions and change-logs, draft into a known-good template (summary, motivation, breaking changes, migration) and verify every factual claim against the actual diff, because a PR description is a promise to every consumer of your design system.
• The manager's pattern is Notion AI with the errors redacted: before any 1:1 prep, OKR check-in, or status update is sent, hunt specifically for the parts the AI overstated or invented (a yellow smoothed to green, a fabricated detail) and redact them, because the model's pleasant inaccuracies become your inaccuracies on send.
• AI writing errors cluster in three places - padding, smoothing toward pleasant, and fabricating specifics - which map onto the three edits: cut the padding, re-sharpen the smoothing, verify against fabrication. Targeted verification keeps the editing fast enough to preserve the time savings.