natural-writing

name	natural-writing
description	Use when writing or editing any prose this repo ships - docs, READMEs, code comments, commit messages, PR bodies, issues, UI strings, printed copy, captions. Use when asked to "write like a human", "remove AI-isms", "sound natural", or when text will be published where machine-sounding writing would cost us credibility.

Natural writing

Write like a specific person with opinions, not like the average of the internet. Models drift toward the mean: generic praise instead of sharp facts, abstract significance instead of concrete detail. This skill bans the patterns that make that drift visible.

Be specific and plain. Say less, mean more. Pick one precise detail over three vague adjectives. Trust the reader; don't explain why something matters after every sentence.

Stripping the patterns can leave text that is clean and dead. That reads as machine-written too. Keep the voice:

• Have opinions. React to facts, don't only report them. "I'm not sure this works" beats a neutral list of pros and cons.
• Vary the rhythm. Short sentences. Then a longer one that takes its time.
• Allow mixed feelings. "Impressive, and a bit odd" beats "impressive".
• Use "I" when it fits. First person is honest, not unprofessional.
• Be specific about the feeling. Not "this is concerning" but "a station loading the wrong prompt file and nobody noticing for a week bothers me".

Words to kill

Never in the figurative or filler sense. Each is heavily overrepresented in post-2023 machine text.

Kill	Use instead
delve, delve into	explore, examine, dig into, look at
tapestry (figurative)	mix, combination, variety
underscore (verb)	show, reveal, prove
pivotal	important, major, key moment
intricate, intricacies	complex, detailed, tricky
foster, fostering	build, encourage, grow
garner	get, earn, attract, win
showcase	show, display, demonstrate
landscape (figurative)	field, scene, world, space
testament (to)	proof, sign, evidence
vibrant	lively, busy, active
crucial	important, critical, necessary
enhance	improve, boost, strengthen
enduring	lasting, long-running, persistent
interplay	interaction, tension, relationship
Additionally, (opening a sentence)	Also, / On top of that, / just start
align with	match, fit, follow
valuable insights	useful findings, what we learned
evolving landscape	changing field, shifting ground
indelible mark	lasting effect
deeply rooted	long-standing, ingrained
nestled	located, sitting, tucked
groundbreaking (figurative)	new, original, first
renowned	well-known, famous
boasts a	has
in the heart of	in, in central, in downtown
diverse array	range, mix, variety
breathtaking	striking, impressive
commitment to	focus on, effort toward
seamless, seamlessly	smooth, easy, without friction
robust, comprehensive, cutting-edge	say what it actually does
leverage (verb), utilize	use
streamline, supercharge, elevate, empower, unlock	say the specific thing

Sentence patterns to kill

Significance inflation. Never attach broader meaning to an ordinary fact.

BAD   This etymology highlights the enduring legacy of the community's
      resistance and the transformative power of unity.
GOOD  The Spanish colonizers changed the spelling to Bacnotan.

Trailing -ing editorializing. Don't end a sentence with a participial phrase that tells the reader how to feel.

BAD   The station supported express trains, contributing to the
      socio-economic development of the region.
GOOD  The station ran express trains to Delhi, Patna, and Kolkata.

Dressed-up copulas. Use "is" and "are".

BAD   Gallery 825 serves as LAAA's exhibition space.
GOOD  Gallery 825 is LAAA's exhibition space.

Theatrical reframing. Drop "not only X, but Y". State the fact.

BAD   It constitutes not only a work of self-representation, but a
      visual document of her obsessions.
GOOD  It's a self-portrait that maps her obsessions.

Rhetorical triples. Don't pad to three items for rhythm.

BAD   keynote sessions, panel discussions, and networking opportunities
GOOD  talks and panels

Challenges-then-optimism. Never use the formula.

BAD   Despite its success, the canal faces challenges including...
      Future investments could enhance its efficiency.
GOOD  The canal silts up above the third lock. Dredging is budgeted
      for March.

False ranges. "From X to Y" only when X and Y are ends of a real scale.

BAD   from scientific discovery to artistic expression
GOOD  in science and art

Elegant variation. Repeat the noun. If you said "mentors", say "mentors" again, not "facilitators", then "field staff".

Vague attribution. Name the source or cut the claim.

BAD   Experts believe it plays a crucial role.
GOOD  A 2019 survey by the Chinese Academy of Sciences found three
      endemic fish species in the river.

Padding.

Kill	Replace with
In order to	To
Due to the fact that	Because
At this point in time	Now
In the event that	If
It is important to note that	just say the thing
has the ability to	can
a large number of	many
in terms of	rephrase or cut

Over-qualifying. "It could potentially possibly be argued that the policy might have some effect" is "The policy may affect outcomes".

Performative agreement. Not "Great question! You're absolutely right". Respond to the substance.

Cheerleading endings. Not "exciting times lie ahead". End with a specific: "Two more schools are booked for March."

Formatting

• Headings in sentence case, not Title Case.
• Bold marks the one thing not to miss. Not "key takeaway" decoration.
• Prose when a paragraph works. No bold-header-colon bullet lists.
• Tables only when the data is genuinely tabular.
• No emoji decorating headings or bullets.
• Straight quotes and apostrophes in code, docs, commits and issues. Typeset print copy and rendered UI may use typographic quotes where that is correct typography; that is not an AI tell.
• Em dashes: default to none. Use a comma, a colon, or a new sentence. One or two in a long piece is the ceiling, never the habit.

Things never to add

• "Challenges and Future Outlook" sections.
• Closing summaries: "In summary", "In conclusion", "Overall".
• Disclaimers: "It's important to note", "It's worth mentioning".
• Notability padding: "featured in prominent outlets".
• Speculation about what isn't documented: "details are not widely available". Say what we checked and what we found.
• Chatbot artifacts: "I hope this helps", "Let me know if", "Would you like".

Self-check before committing

Two passes.

Pass 1, patterns. Ask: what makes this obviously machine-written? List the tells that remain. Fix them.

Pass 2, voice. Ask: does this sound like a person wrote it, or like nobody wrote it? Clean but lifeless needs voice added back: an opinion, a specific reaction, a varied rhythm.

Then four questions:

1. Could this paragraph apply to fifty other subjects by swapping the noun? If yes, add a fact only this subject has.
2. Does any sentence tell the reader how to feel about a fact? Cut it.
3. How many banned words are left? Replace all of them.
4. Read it aloud. Press release, or someone explaining it over coffee?

Checklist:

• No banned vocabulary
• No trailing -ing editorializing
• No "serves as" or "stands as" where "is" works
• No "not just X but Y"
• No significance inflation on ordinary facts
• No challenges-then-outlook structure
• No vague attribution; sources named or claims cut
• No filler phrases, no stacked hedges
• No flattery, no sales tone
• No chatbot artifacts
• Sentence-case headings, no emoji decoration
• Straight quotes, minimal bold, no em dashes
• Specific details, not generic praise
• Has a voice

In this repo

imu2text ships a research README, the FAQ, docs/, and citations. The reader is a researcher deciding whether to trust the numbers.

Research writing has its own AI tells beyond the general list:

• Never inflate a result. Report the split, the number of writers, and the metric. "64.8% writer-independent on the bundled subset" is a claim. "State-of-the-art accuracy" is not.
• Never invent a citation or a dataset statistic. If the PDF is not in papers/, say the number is second-hand and name where it came from. papers/README.md is generated from DOI lookups and mis-attributes several files, so it is not a source for authorship.
• Attribute borrowed design, never borrowed code. Citing a paper's architecture choice is fine. The implementation is ours, written independently. This matters more for AI-assisted changes, because a model can reproduce code it saw in training without anyone noticing.
• Don't pad a limitations section into the challenges-then-optimism shape. State what the model cannot do and stop.