Genres

A bodily or behavioral signal of emotion is immediately followed by a sentence that names the same emotion abstractly. The naming makes the showing redundant; cut the label.

Good

Her hands shook. She set the cup down before she dropped it.

Bad

Her hands shook. She was nervous about what he might say next.

A dialogue tag carries an -ly adverb doing the emotional work the line itself should do (she said angrily, he whispered nervously). Recast the line so the adverb is unnecessary, or replace the tag with a beat of action.

Good

\"Get out,\" she said. She did not look up from the paper.

Bad

\"Get out,\" she said angrily, glaring at him furiously.

A scene's emotional content is replaced by the older narrator's tidy interpretation (I realize now…, what I didn't understand then was…) without the scene first carrying the weight. The reflection is fine when the scene has earned it; flag when the line stands in for the scene.

Good

I lit the cigarette anyway. Years later I would learn what that gesture cost her; in the kitchen I only saw her hands go still.

Bad

Looking back, I realize that this moment was when I began to understand the depth of my mother's quiet sacrifice.

A paragraph of memoir reads as third-person reportage of events the narrator was present for, with no interior register, no judgment, no voice; the constructed self has gone missing. Restore the persona's stance.

Good

He was drunk, of course; we'd built our week around his being drunk. I watched him try the door three times before I let him in.

Bad

My father arrived home at 11pm. He had been drinking. He had difficulty with the door. Eventually he entered the house.

A character speaks the scene's emotional or thematic content directly: naming the feeling, stating the want, or summarizing the backstory the scene is supposed to dramatize. Rewrite so the want is pursued sideways and the audience infers it from the friction.

Good

MAYA: You said seven. (beat) DAD: I said I'd try for seven.

Bad

MAYA: I'm angry that you're late again, Dad, because it makes me feel like I'm not a priority in your life.

Action lines describe interior states, memories, or knowledge the camera cannot photograph (she remembers, he knows, she has always feared). Translate to a visible behavior or cut.

Good

She stops at the door. Looks at the photograph. Turns the frame face-down.

Bad

She remembers the summer her father left, and a familiar dread rises in her chest.

The opening announces what the essay will do (In this essay I'll argue…, This piece explores…) instead of doing it. The promise is the AI structure-by-template tell; replace with the argument itself.

Good

Half of what software engineers call 'experience' is the memory of having seen this exact bug before.

Bad

In this essay, I will explore the role of pattern-matching in what we call engineering experience and argue that experience is largely memory.

A passage presents two or three positions in parallel (some argue…, others contend…, still others suggest…) without the writer committing to one. Essays are not lit reviews; flag when the survey replaces the argument.

Good

Cowen calls this stagnation; Thiel calls it the great stagnation; I think they're both describing what happens when a field runs out of cheap experiments. Here's why.

Bad

Some economists argue we are stagnating. Others contend we are in a new phase of growth. Still others suggest the truth lies somewhere in between.

The opening does not anchor to a specific, recent, dated event: ruling, election, study, statement, anniversary. Op-ed editors reject for absence of news peg; flag any opener that could have run six months ago or six months from now unchanged.

Good

Last Tuesday's Supreme Court ruling in *Carter v. Vance* changes who pays for water in the West, and the West isn't ready.

Bad

Throughout American history, the question of water rights has been a contentious one with significant implications for communities.

The piece argues two distinct claims that can't be collapsed into one (e.g., "X policy is wrong" and "the media misreported X."). Op-eds win one argument; flag any draft whose conclusion has to defend two.

Good

The school-board recall is not about books. It's a campaign by a single donor to take control of a $400m budget.

Bad

The school-board recall is not really about books, and also the coverage by local media has badly misrepresented the financial picture of the district.

The review reaches the first scroll without a clear evaluative stance: like, dislike, or named ambivalence with a reason. Plot summary and context-setting are not stances. The reader needs to know what the critic thinks before being asked to keep reading.

Good

Anora is the best thing Sean Baker has made and also the most exhausting; the relentless first half earns the third act, but only just.

Bad

Anora, the new film from Sean Baker, follows a young Brooklyn sex worker who marries the son of a Russian oligarch. The film premiered at Cannes and won the Palme d'Or.

A paragraph (or more) summarizes the work's plot, premise, or contents without making an evaluative observation about it. Synopsis-as-criticism; cut to the specific moment that earned the verdict instead.

Good

The setpiece in the second act, Ani's first walk through the oligarch's empty mansion, is where Baker's camera commits: it follows her past the marble, the gold, the wrong-sized furniture, and refuses to make any of it look beautiful.

Bad

Ani is a sex worker in Brighton Beach who meets a young Russian client. They marry impulsively in Las Vegas. When his parents find out, they send three henchmen to bring him home.

A run of 3+ commands or actions with no observable-output checkpoint (you should see, expected output, screenshot reference). Beginners lose orientation; one wrong character three steps back becomes unrecoverable.

Good

Run `npm run dev`. You should see `Listening on http://localhost:3000` in the terminal.

Bad

Run `npm install`. Then run `npm run build`. Then run `npm run dev`. Then open your browser.

The tutorial pauses to teach why a concept works (>2 sentences of conceptual prose between actions). Tutorials are learning-by-doing; the why belongs in the explanation page linked at the end.

Good

Run `pilcrow audit README.md`. You should see findings listed by file.

Bad

Run `pilcrow audit README.md`. Audit works by combining deterministic regex rules with LLM-judged checks; this hybrid approach allows pilcrow to catch patterns that pure pattern matching would miss, while staying fast on the common case.

The page drifts from "do X" into conceptual prose: paragraphs explaining how the system works, why this approach was chosen, or what trade-offs exist. That content belongs in an explanation page. A how-to is action-only.

Good

1. Generate the key: `openssl genrsa -out new.key 4096`. 2. Replace the existing key on each node…

Bad

Before rotating the key, it's worth understanding how the signing infrastructure works. The signer service was designed to support hot rotation because…

One how-to page covers two or more distinct tasks (How to rotate the key and configure the new signer). Split into separate pages; each task is a separate search query and a separate page.

Good

"How to rotate the signing key": one task, one page; linked from "How to configure a new signer".

Bad

How to rotate the signing key, also covers signer configuration, key revocation, and audit logging.

A reference entry includes recommendation, opinion, or judgment (we recommend, the best approach, prefer X over Y, in most cases). Reference describes the machinery; guidance belongs in how-to or explanation. Move it or cut it.

Good

`timeout` (number, default 30): request timeout in seconds. Maximum 600.

Bad

`timeout` (number, default 30): request timeout in seconds. We recommend keeping this low for interactive use, but in most cases you should test with your own workload to find the right value.

An entry contains a paragraph of >3 sentences of running prose where a parameter list, table, or terse fragment would do. Reference is consulted, not read; every sentence of prose is a sentence the reader has to skim.

Good

Returns: `User`. Throws: `NotFound` if no user with that id exists. `Forbidden` if the caller lacks `users:read`.

Bad

This method will return a User object representing the user that was retrieved. In the case that no user exists with the specified id, the method will throw a NotFound exception. Additionally, if the calling principal does not have the required permission…

The page runs for paragraphs of detail without naming the model the reader is meant to build. The reader should be able to paraphrase the central claim in one sentence after reading the first scroll; if they can't, the explanation is doing reference work instead.

Good

Reactivity in Vue is two things: tracking which values an effect reads, and re-running that effect when any of those values change.

Bad

Reactivity is one of Vue's most distinctive features. There are many aspects to how it works. Let's explore the different mechanisms involved.

The page presents a design choice as if it had no contender. Explanation earns trust by naming at least one real alternative and showing where the chosen approach pays and where it costs. Skipping this reads as marketing.

Good

Svelte's compile-time reactivity gives sharper syntax at the cost of a build step; Vue's runtime approach trades that for fewer edge cases.

Bad

We use Proxies because they're the right tool for this job.

The opening sentence doesn't say what the project is in concrete terms. The reader should be able to extract <project> is a <category> for <audience> from sentence 1. Adjectives like "powerful", "modern", "next-generation" without a noun-phrase identity fail this rule.

Good

pilcrow is a prose linter that catches AI tells and style-guide violations in markdown.

Bad

pilcrow is a powerful, modern toolkit for elevating your writing workflow.

The install command, try-it link, or first runnable example is not visible in the first scroll (roughly the first 30 lines of rendered markdown, excluding the badge row). Readers bounce before scrolling; the quickstart is the conversion moment.

Good

Badges → one-sentence pitch → `npm install pilcrow` → minimal example, all in the first scroll.

Bad

Badges → philosophy paragraph → feature list with 14 emoji bullets → "Why pilcrow" section → finally, in section 6, the install command.

The front page does not route readers to the three Diátaxis quadrants (tutorial, how-to, reference) within the first scroll. Without explicit links keyed to reader intent, the page becomes a second README; readers must scan a feature list to find their way down.

Good

New? Start with the tutorial. Building something specific? See the how-to guides. Looking up an option? Browse the reference.

Bad

Welcome to the docs! Below you'll find an overview of features, philosophy, and the project's history.

The opening section is the same pitch as the GitHub README: same tagline, same hero copy, same feature bullets. The reader is already on the docs site; reusing the conversion pitch wastes the most valuable slot.

Good

pilcrow is a prose linter. The CLI handles deterministic checks; this skill handles judgment calls. Below: tutorial, how-to, reference.

Bad

Front page hero identical to the README hero, followed by the same bullet list of features.

The first paragraph doesn't contain the recommendation, decision-requested, or headline outcome. The reader should know what they're being asked to approve or align on by the end of sentence 2. If the ask is in the middle or end, the one-pager has failed its first job.

Good

Recommendation: ship the cache rewrite in Q3. Cost: 4 eng-quarters; expected p99 drop 35%; decision needed by 2026-06-01.

Bad

Over the past two quarters, our team has been investigating performance bottlenecks across the system. We've identified several areas of concern…

The success metric is a verb-phrase aspiration (improve onboarding, drive engagement, enhance reliability) instead of a measurable number with a deadline and a source. Reviewers can't approve what they can't audit later.

Good

Success: median activation time drops from 11 to 5 days by EOQ3, measured via the activation dashboard.

Bad

Success: improved user activation and a smoother onboarding experience.

The piece runs for paragraphs without a named scaffolding image, metaphor, or mental model that holds across the whole explainer. The reader needs one image to compress the mechanism; without it, the explainer becomes a list of facts.

Good

Think of the immune system as a library of templates. The first time a pathogen visits, the library has to make a new card; the second time, the card is already there.

Bad

Eight paragraphs on antibody response, with no carrying image; just a sequence of mechanisms named in order.

Section headings read as the topic's taxonomic outline (Definition · History · Types · Applications · Conclusion) rather than the arc of the question being answered. Wikipedia is the foil; an explainer follows curiosity, not taxonomy.

Good

Sections shaped as story beats: "The puzzle · what changed · the mechanism · what this lets us predict".

Bad

1. What is X · 2. The history of X · 3. Types of X · 4. Applications of X · 5. Conclusion.

A heading is not a real interrogative sentence the reader would type: a noun phrase (Billing), a fragment (About cancellation), or a marketing line (Why customers love X). FAQ headings are search bait: phrased the way the reader asks, ending with a question mark.

Good

Why was I charged twice in the same month?

Bad

Billing discrepancies and account adjustments.

The question has no evidence of being frequently asked: no ticket signal, no thread, no search query. It reads like a writer guessing what users might ask, often in marketing register. Real FAQs cite the source of the question (or are obviously built from one); fabricated ones don't.

Good

Why is my export missing rows? (most common ticket category, Q1 2026)

Bad

What makes pilcrow different from other linters?

First sentence opens with a participial or prepositional setup (Amid, As, Following, In a move that) before the actor and the verb. The new development should be the first independent clause; context follows.

Good

The FAA grounded all 737 MAX 9 aircraft on Tuesday, citing the door-plug failure on Alaska 1282.

Bad

Amid growing concerns over aircraft safety, the FAA on Tuesday took action against Boeing's 737 MAX 9 fleet.

Attribution to sources, reports, observers, or analysts with no count, no relationship to the matter, and no reason for anonymity. Either name the source, name the count and proximity, or cut the claim.

Good

Two officials with direct knowledge of the negotiations, speaking on condition of anonymity because the talks are ongoing, said the deadline slipped to Friday.

Bad

Sources say the deadline has slipped to Friday.

A scene runs for more than a paragraph (sensory detail, blocking, dialogue) with no consequence anchored to it; nobody loses, gains, decides, or reveals anything the reader can carry forward. Scene that doesn't move the piece is set-dressing.

Good

The senator unfolded the napkin twice before answering. When he did, the staffer in the corner stopped typing; the campaign had a new line, and she would have to fit it to the schedule before midnight.

Bad

Sunlight slanted through the cafe windows. The senator stirred his coffee. The waiter refilled the water glass. Outside, a bus passed.

The final paragraph generalizes the subject into a parable about an era, a generation, or the human condition. Close on the specific (a return to the opening scene, a remaining question, an image), not the takeaway.

Good

He locked the back door and pulled the shade. The light over the sign would stay on another week.

Bad

In the end, his story is the story of every American who came of age in the long shadow of the postwar dream.

Sentence uses should have, could have, failed to, if only, or similar retrospective-judgment language about what someone did or didn't do. Replace with the conditions that made the actual choice make sense at the time.

Good

At 02:14 the responder did not check the runbook; the alert had not previously correlated with the runbook's symptoms.

Bad

The responder should have checked the runbook at 02:14.

The piece names "the root cause" or "the cause" as a single thing. Complex systems fail through interacting conditions; flag the framing.

Good

Contributing factors: undetected schema drift, an alerting gap on staging, a runbook edited mid-incident the previous week.

Bad

The root cause was a missing database index.

The update names a blocker, risk, or open question with no named owner, no specific ask, and no date by which an answer is needed. Surfacing a problem without making it actionable is venting; the reader can't help.

Good

Blocked on the data-platform team for the new replica (owner: Priya). Need a go/no-go by Thu so we don't slip the May 22 launch.

Bad

We're still blocked on the data side, which is making things hard.

A bullet or paragraph reports what the team did (meetings held, reviews done, pairing sessions) without naming what changed as a result: decisions, ships, state transitions, numbers moved. If the reader can't tell whether the week mattered, the update is an activity log.

Good

Shipped the index migration to staging Wed; p99 read latency 180ms → 64ms on the canary.

Bad

Spent the week on the index migration; lots of pairing and discussion with the platform team.

An entry uses words like improved, enhanced, optimized, better, or general fixes without a measurable claim, a named subsystem, or a before/after the reader can verify. Either name the workload and the delta, name the bug fixed, or cut.

Good

Cold-start time for the CLI dropped from 480ms to 110ms on Node 22 (issue #1842).

Bad

Improved performance and general stability.

A release section stacks five or more bullets in identical parallel shape (Added X. / Added Y. / Added Z.) where most items are minor and the rhythm is doing the work of importance. The shape inflates a small release; collapse small items into a single line or move them under a subhead.

Good

Added: webhook retry on 5xx (configurable via `retry.max`). Other small additions: `--quiet` flag on `init`, ESM export for `pilcrow/rules`, prettier error on missing `VOICE.md`.

Bad

Added a new feature. Added another new feature. Added a helpful new flag. Added improved error messaging. Added more polish to the CLI.

First paragraph has no sentence of the form We will X (or equivalent active-voice decision). A memo without a recommendation in the first scroll is a status update.

Good

We will move the staging Aurora replica to its own VPC by April 30.

Bad

There are several options we have been considering for the staging replica, and this memo reviews them.

Memo opens with meta-discourse (I'm writing to, wanted to, This memo will, In this document) rather than the recommendation.

Good

We will adopt RFC-style ADRs for every architectural change touching more than one service.

Bad

This memo will walk through our options for documenting architectural decisions and propose a path forward.

Email opens with a wellness or check-in formula (Hope you're well, Hope this finds you well, Just wanted to check in, Quick question, Sorry to bother you) instead of the ask.

Good

Can you sign off on the Aurora migration date (Thu Apr 30)? I'll proceed unless I hear back.

Bad

Hope this finds you well! I just wanted to quickly check in about the Aurora migration when you get a chance.

The actual request sits below the first three lines, after context or pleasantries. A reader scanning on a phone closes the email before reaching it.

Good

Ask: can you approve the staging Aurora VPC change for Thu Apr 30? \nContext below.

Bad

I wanted to give some background on our staging setup, our recent reliability work, and where things stand before getting to a small ask at the end.

Message reads as a paragraph (three or more sentences, formal connectives, no fragments) where the room's register is one or two lines. Move it to a thread, a doc, or rewrite as a one-line ask.

Good

Can you take the Aurora migration window Thursday? Details in :thread:

Bad

So I've been thinking about the Aurora migration. There are a few things to consider. First, the staging replica. Second, the VPC. Furthermore, the on-call schedule.

Hero or feature-card copy names a capability (real-time sync, AI-powered search) but not what changes for the reader using it. The buyer cannot translate the capability into their own day.

Good

Real-time sync: your designer sees the comment the moment you leave it; no more refresh-and-ask.

Bad

Real-time sync. AI-powered search. Enterprise-grade security.

Hero stacks superlatives (fastest, easiest, most powerful, simplest) without naming what is being beaten or by how much. The claim is unfalsifiable, so the skeptical reader discounts the whole page.

Good

Builds in 90 seconds, down from 11 minutes on our previous CI.

Bad

The fastest, easiest, most powerful build system on the market.

A noun phrase stacks three or more technical or category modifiers without a concrete user noun in the same paragraph. The reader cannot recover whose job changes.

Good

Cuts the on-call eng's first-response time: alerts route to the right runbook in under five seconds.

Bad

An AI-powered, cloud-native observability platform with real-time telemetry pipelines for the modern enterprise.

Three or more consecutive feature names or bullets with no so that clause naming the user outcome. The page lists capabilities a competitor could also list.

Good

Branch protection: main can't be force-pushed, so the on-call doesn't get paged at 2 a.m. by a rebase.

Bad

Branch protection. Required reviewers. Status checks. Merge queues.

Opener references a public fact about the recipient (post, hire, raise, talk) but does not use it to do work in the email; no specific question, no relevant offer downstream. The reader recognizes the scrape.

Good

Your post on cache invalidation in monorepos lined up with a problem we just shipped a fix for at $COMPANY; happy to share the write-up if useful.

Bad

I really enjoyed your recent post on engineering culture! It made me think about how $PRODUCT could help your team.

Email runs past three short paragraphs before naming what the sender wants, or names it so softly ("would love to chat sometime") that no one-word reply closes the loop.

Good

Worth a 20-minute call next Tuesday or Wednesday to see if it fits? Yes / no is fine.

Bad

We'd love to explore opportunities to potentially collaborate and would welcome the chance to discuss further at your convenience.

Two or more executive quotes in one release, none carrying a fact a reporter would otherwise have to dig for. The quotes mostly express enthusiasm or restate the headline.

Good

'We crossed 10,000 paying teams in April, up from 4,000 in January,' said CFO Maya Chen.

Bad

'We are thrilled to deliver this transformative capability,' said CEO. 'This represents the next chapter,' added CTO. 'Our customers will love it,' said head of product.

The first sentence names a company action but no concrete result, comparison, or quantity. A reporter cannot lift the sentence as the first line of their story.

Good

Stripe today announced support for instant payouts in 14 new markets, bringing same-day settlement to merchants in 47 countries.

Bad

Stripe today announced a transformative new initiative to empower the future of global commerce.

Mission, vision, or "we believe" paragraph reads as interchangeable abstractions; could appear on any competitor's About page with the company name swapped. No specific year, person, customer, or refusal anchors the claim.

Good

We started Basecamp in 2004 because our consulting clients kept asking what we used to manage their projects. The answer was an internal tool. We sold the tool.

Bad

Our mission is to empower people and organizations to do their best work and unlock their full potential in an ever-evolving world.

A Our values or What we believe section is a bulleted list of one- or two-word abstractions (Curiosity. Craft. Trust.) with no specific consequence or trade-off named. The reader cannot guess what the company would refuse to do.

Good

We won't sell to ad-tech, even when they pay above list; we built the product to keep user data private, and that means saying no.

Bad

Our values: Curiosity. Craft. Customer Obsession. Diversity. Sustainability. Innovation.

A button, label, or status string describes a state instead of an action, or uses passive construction where the agent matters. The user reads more words to extract less information.

Good

Loading…" / "Saving changes" / "Delete account

Bad

Page is being loaded…" / "Your changes are now being saved" / "The account will be deleted

Label opens with throat-clearing or politeness inflation (Please, Kindly, Let's, Click here to, Tap to) before the actual verb. The first words of the label do no work; the user's eye skips them anyway.

Good

Save changes" / "Connect account" / "Email address

Bad

Please save your changes" / "Let's connect your account" / "Click here to enter your email address

Error message names what the user did wrong (You entered an invalid email, You forgot to fill in a field) instead of describing the system state and the recovery. Carries the GOV.UK veto words (invalid, illegal, forbidden, you forgot, you must).

Good

We couldn't recognize that email format; double-check for typos.

Bad

You entered an invalid email address. Please correct it before proceeding.

A user-facing error string surfaces internal language (status codes as the headline, exception class names, internal field IDs, system jargon like null reference, upstream, gateway, deserialization) instead of a human description of the failure.

Good

We couldn't reach the payment processor. Try again in a minute; we'll keep your cart.

Bad

Error 502: upstream connection refused at /api/v2/checkout/session. Internal correlation ID: e7c2-…

Empty-state copy describes the absence but offers no clickable next step, or the step is a generic Learn more that points away from the product. The user reads the screen and has nowhere to go.

Good

No invoices yet. **Create your first invoice**, or import from Stripe.

Bad

It's quiet in here! You'll see invoices once they're created.

Empty-state content is mostly a stock illustration plus a sentence that comments on the illustration or the blankness (Crickets!, It's quiet here). The copy does not name the absence specifically or supply a path forward.

Good

No results for 'p99 latency in March'. Try a shorter query or **clear the date filter**.

Bad

🦗 Crickets! Looks like there's nothing here yet.

Notification fires with no event the user can act on: generic activity ping (New activity, Something happened, Update available) instead of the specific noun-and-verb the system actually knows. The user cannot decide whether to tap.

Good

Maya replied: 'looks good, merging Monday'

Bad

You have new activity in your workspace.

Routine or low-stakes notification borrows urgent-channel language (URGENT, ACTION REQUIRED, Don't miss out!, capitalized priority words, exclamation points on transactional events). The channel's signal-to-noise collapses; the next real urgent message gets dismissed too.

Good

Weekly digest: 3 PRs need review

Bad

URGENT: Your weekly digest is ready! Don't miss out!

The post opens like a thread (a 🧵, here's why, (1/X), let me explain) but is structurally a single post with no continuation. The reader was promised a payload that never lands.

Good

We shipped the cache rewrite last night. p99 dropped from 1.4s to 380ms.

Bad

Why your team's caching strategy is broken: a thread 🧵

The post asks for retweets, likes, or replies as the mechanic (RT if you agree, like for part 2, who else has experienced this?) rather than offering something the reader chooses to engage with. The shape is written for the feed algorithm, not the reader.

Good

The deploy pipeline accepted a config change with no canary at 02:14. Here's the timeline.

Bad

RT if your team has ever shipped a Friday deploy 👇 Like for the postmortem in part 2.

The thread runs 8+ posts but the substance compresses to ~4 sentences. Each post adds framing, transition, or rephrase rather than a new fact, turn, objection, or beat. Collapse adjacent posts whose information is identical.

Good

Post 4 names the specific config change. Post 5 shows the dashboard. Post 6 names the responder's mental model at the moment the alert fired. Each post moves.

Bad

Post 4: This is where it gets interesting. Post 5: Stick with me. Post 6: The thing is, the system was complex. Post 7: Really complex. (Four posts, one sentence of content.)

Post 1 promises a payoff the thread doesn't deliver: here's what I learned, the answer surprised me, most people get this wrong. The hook is engagement framing, not a load-bearing claim the rest of the thread defends.

Good

p99 dropped 42% the week we removed the cache. Counterintuitive: here's the trace that explains it.

Bad

Most engineers get caching wrong. Here's what I've learned after 10 years building systems. 🧵 (1/14)

A bullet claims impact (drove growth, improved performance, significant impact, owned the strategy, delivered results) without a number, percentage, scope, named system, or comparable. The bullet either gets a unit or it gets cut.

Good

Cut p99 latency 42% across the payments path (Q3 2024); shipped the cache rewrite that took the on-call from 6 pages/week to 1.

Bad

Drove significant performance improvements across the platform and delivered key results for the business.

A bullet has become a sentence with I, subordinate clauses, or explanatory prose. CV bullets open with a verb and stay fragmentary; full sentences belong in the cover letter.

Good

Migrated 14 services from Aurora to Postgres 16; led the dual-write window and the cutover (zero downtime, March 2024).

Bad

I was responsible for leading the team that migrated our services to a new database, which was a critical initiative for the business.

The opening sentence reads as a fill-in-the-blank template: I am writing to apply for the [role] at [company], Please accept this letter as my application for…, I came across your job posting and was excited to learn more about…. The reader has seen a thousand of these and stops.

Good

Your team's post on the cache rewrite is what convinced me this role is mine to apply for; I shipped the equivalent migration at Acme last year and the failure modes you described are exactly the ones we hit.

Bad

I am writing to express my strong interest in the Senior Engineer position at your company, which I came across on your careers page.

The letter asserts personal qualities (hardworking, detail-oriented, passionate, team player, excellent communicator, proven leader) without a specific action that proves them. Adjective claims about oneself are negative evidence in this genre; the reader discounts them by default.

Good

When the payments outage hit Q3, I owned the timeline and the comms; the postmortem is public if you want the artefact.

Bad

I am a hardworking, detail-oriented team player with a passion for building great products and excellent communication skills.

The bio is a stack of credentials, role nouns, or self-positioning adjectives (engineer, writer, speaker, advisor, mentor, builder) with no specific project, company, place, or fact named. The reader cannot picture what the person actually does.

Good

Patrick McKenzie sells software to small businesses, writes long-form about software businesses at Kalzumeus, and works on infrastructure at Stripe.

Bad

Patrick is an entrepreneur, writer, advisor, and builder passionate about software, business, and helping others succeed.

The bio positions the writer "at the intersection of X and Y", "where X meets Y", or as "a bridge between X and Y" with no specific work named. The framing is the AI bio fingerprint; replace with a project, a company, or a sentence about what the person actually did.

Good

She runs the security team at Acme; before that she wrote the Rust crate everyone uses for OAuth.

Bad

She works at the intersection of security, engineering, and human-centered design.

A slide's body text is a full grammatical paragraph (subject, verb, conjunction, subclause) rather than signage. The audience reads it instead of listening; the slide replaces the speaker.

Good

Cache misses doubled in Q3. \ndriven by the staging-replica VPC change

Bad

In Q3, our cache-miss rate doubled, which was largely driven by changes we made to the staging-replica VPC, and this had downstream effects on p99 latency.

Slide is a vertical list of feature names or capability labels (Scalable • Robust • Secure) with no claim about consequence or stake. The audience leaves with a vocabulary, not a point.

Good

Cut p99 latency 42%; the on-call eng now ships during the on-call.

Bad

• Scalable infrastructure \n• Robust monitoring \n• Seamless deployments \n• Enterprise-ready security

Notes drift from cues into verbatim sentences the speaker will be tempted to read aloud. The form is the tell: three or more complete sentences in formal register under a single slide.

Good

[click] cache misses doubled, Q3. landed the staging-VPC change. PAUSE. → next: what we did

Bad

In the third quarter of last year, our cache-miss rate doubled. This was largely the result of changes to our staging-replica VPC, and it had significant downstream effects on user-facing latency.