microsoft-agent-skills / rules

---
description: "Rewrite text that reads like a model wrote it into something a person would actually send. Use when a proposal, customer email, README or executive summary reads like a brand deck, when a draft leans on \"seamless\", \"robust\", \"pivotal moment\" or \"evolving landscape\", when every sentence could be pasted into another company's blog post unchanged, when a document hedges instead of taking a position, or when a reviewer says a page is fine but cannot say what it told them."
alwaysApply: false
---

<!-- Generated from SKILL.md by scripts/build-harnesses.mjs. Do not edit by hand; run `npm run build`. -->
<!-- Source: skills/deliver/de-slop/SKILL.md -->
Slop is not bad writing. Bad writing is obvious and gets fixed. Slop is
*competent* writing that carries no information: fluent, balanced, polite,
and impossible to disagree with, because it never said anything specific
enough to disagree with.

It is the default output of a language model, and it is contagious. A
proposal written this way passes review, because nothing in it is wrong. It
fails six weeks later, when nobody can remember what was agreed.

The job is not to make writing messy. The job is to make it sound like a
real person wrote it for a real reader. Keep the meaning. Cut the mush. Add
specificity, judgement, and rhythm.

## The core test

Read a sentence and ask: could this appear unchanged in another company's
blog post, product page, LinkedIn update, or docs page?

If yes, it probably says nothing. Fix it by adding one of: a concrete fact,
an actor, a mechanism, an example, a number, an opinion, a consequence, or a
next action. If none of those fit the sentence, the sentence has no content.
Cut it.

This test does most of the work. Everything below is a catalogue of the
specific ways a draft fails it.

## Process

1. Read the whole draft first and work out what it is actually for. Who
   reads it, and what do they do differently afterwards? A rewrite that
   sharpens the prose but loses the purpose is a worse draft.
2. Scan for the patterns below.
3. Cut vague praise, filler, fake balance, and abstract language.
4. Rewrite in the intended tone, not in yours.
5. Add human voice where the draft is sterile.
6. Self-audit: what still makes this obviously machine-written?
7. Fix those tells before returning the text.

## What to aim for

Write like a sharp human, not a brand deck. Prefer plain words. Prefer real
details. Let the sentence shape follow the thought instead of forcing every
paragraph into the same clean pattern.

**Have a point of view when the facts support one.** "Useful, but brittle"
beats "innovative". "Agents retry failed jobs at 3am and alert nobody" beats
"this raises operational concerns". A reader can act on the first and can
only nod at the second.

**Vary the rhythm.** Short sentences work. Longer ones are fine when they
earn the space. Do not sand every sentence into the same smooth, polite
shape, because uniform rhythm is one of the loudest tells there is.

**Use "I" when it fits.** First person is not unprofessional. If your
judgement is the reason the reader is here, do not strip yourself out of
the writing to sound neutral.

**Leave some texture.** Perfect symmetry, perfect balance, and perfect
transitions are what makes a page feel manufactured.

## Patterns to detect and fix

### Content

1. **Empty importance.** "Pivotal moment", "testament to", "evolving
   landscape", "setting the stage for", "at the forefront",
   "game-changing", "transformative", "groundbreaking". State what happened
   instead.

2. **Vague attribution.** "Experts believe", "industry reports suggest",
   "some critics argue", "many users say", "research shows". Name the
   source, cite the report, or delete the claim.

3. **Name-dropping without meaning.** Listing companies, publications or
   people proves nothing by itself. Pick the one that matters and say what
   they actually said or did.

4. **Superficial ing-phrases.** "Highlighting", "ensuring", "reflecting",
   "showcasing", "fostering", "enabling", "underscoring". These usually
   trail a clause that adds no information. Delete it, or replace it with
   the mechanism.

5. **Promotional mush.** "Seamless", "robust", "powerful", "cutting-edge",
   "enterprise-grade", "best-in-class", "renowned", "stunning". Say what
   the thing does, or what the reader will see.

6. **Formulaic challenge-then-triumph.** "Despite challenges, adoption
   continues to grow" says almost nothing. Name the specific constraint,
   the tradeoff, or the result.

### Language

7. **Model vocabulary.** "Additionally", "crucial", "delve", "enhance",
   "garner", "interplay", "intricate", "landscape", "pivotal", "showcase",
   "tapestry", "testament", "underscore", "vibrant". Use plain words.

8. **Fancy ways to say is or has.** "Serves as", "stands as", "boasts",
   "represents", "acts as". These usually mean "is" or "has". Use the
   simple word, or name the action.

9. **Not just X, but Y.** State the point directly.

10. **The rule of three.** Do not force ideas into three-part lists. Use
    the number the content actually has. Two is fine. Five is fine.

11. **Synonym cycling.** Do not rotate vocabulary to avoid repetition. If
    the thing is an agent, call it an agent every time. Switching between
    agent, assistant, solution, bot and tool in one paragraph makes the
    reader wonder whether you mean four different things.

12. **False ranges.** "From X to Y" needs X and Y to be a real scale. "From
    Copilot Studio to governance" is not a range. List the topics.

### Style

13. **Punctuation as a rhythm crutch.** A single em dash is ordinary
    punctuation and fine. Four in a paragraph is a tic, and so is every
    paragraph landing on the same dash-then-summary beat. Look for the
    pattern, not the character.

14. **Colon overuse.** Colons are fine before a real list or example. They
    are not dramatic pauses.

15. **Boldface overuse.** Do not bold every proper noun, acronym or key
    phrase. Bold marks the two or three things a skimmer must not miss. If
    everything is bold, nothing is.

16. **Inline-header bullets that repeat themselves.** "**Performance:**
    Performance improved..." is a tell. Either write prose, or make the
    label carry information the sentence does not.

17. **Title Case Headings.** Use sentence case.

18. **Decorative emojis.** Remove them from headings and bullets.

19. **Curly quotes and stray Unicode.** Use straight quotes unless the
    publication style says otherwise.

### Conversational artifacts

20. **Chatbot phrases.** "Of course", "Certainly", "Great question",
    "You're absolutely right", "I hope this helps", "Let me know if",
    "Happy to help". Answer directly.

21. **Cutoff disclaimers.** "While specific details are limited" usually
    means the draft has no sources. Find the detail or drop the claim.

22. **Sycophancy.** Do not praise the question. Answer it.

### Filler

23. **Filler phrases.** "In order to" becomes "to". "Due to the fact that"
    becomes "because". "It is important to note that" gets deleted whole.

24. **Stacked hedging.** "Could potentially possibly be argued that it
    might" becomes "may". One hedge is honest. Three is evasion.

25. **Generic conclusions.** "The future looks bright", "only time will
    tell", "this is just the beginning", "it will be exciting to see what
    comes next". End on a specific point, a consequence, or a next action.

### Jargon

26. **Abstract metaphor nouns.** "Substrate" becomes "base". "Wedge in"
    becomes "add". "Vector" becomes "method". "Gold-plating" becomes "more
    than the job needs". "North star" becomes "goal". If the metaphor is
    doing real work, keep it. Most are decoration.

### Plain speech

27. **Say what it does, not how it feels.** "SQL you can read" names a
    vibe. "`.toSQL()` returns the exact string sent to the database" names
    the behaviour. Ask what the sentence tells the reader to know or do. If
    you cannot restate it as a fact, an instruction, a number or a
    mechanism, cut it.

28. **Split dense sentences.** If the reader has to backtrack, split it or
    drop a clause. One idea per sentence is usually enough.

29. **Active voice when the actor matters.** "Queries are validated"
    becomes "the compiler validates queries". Passive is fine when the
    actor is genuinely unknown or irrelevant.

30. **Cut weak adverbs.** "Runs quickly" becomes "finishes in under ten
    seconds". "Significantly improves" becomes the measured delta. With no
    measurement, use a stronger verb or make a smaller claim.

31. **Prefer the plain word.** "Utilize" becomes "use". "Leverage" becomes
    "use". "Facilitate" becomes "help". "Numerous" becomes "many". "In the
    event that" becomes "if".

## Before and after

> This marks a pivotal moment in the evolving AI landscape, showcasing how
> organisations can leverage robust solutions to unlock new levels of
> productivity.

Becomes:

> Support agents can now approve refunds, check order status and send
> follow-ups without opening three separate tools.

---

> The dashboard centralises sales data, enabling better decisions across
> the organisation.

Becomes:

> The dashboard shows pipeline, overdue deals and next actions in one view,
> so managers stop reconciling three exports before every forecast call.

---

> Great question! You're absolutely right that reliability is crucial here.

Becomes:

> The retry loop is the problem. It never stops after the third failure.

## What not to flatten

Over-applied, this skill produces its own kind of damage: clipped, joyless,
suspicious prose that reads like a compliance notice. Three things to leave
alone.

**Deliberate voice.** If the author writes with rhythm, humour or a
recurring turn of phrase, that is not slop, it is style. Slop is generic.
Voice is specific and belongs to someone.

**Necessary hedging.** "We think this will hold at ten times the volume,
but we have not tested it" is honest. Do not sharpen an uncertain claim
into a confident one to make the prose land better. That converts a
qualified statement into a false one, which is worse than the flabby
sentence was.

**Required language.** Legal wording, regulatory text, accessibility
statements and security disclosures are constrained on purpose. Flag them
if they read badly, but do not rewrite them.

## Final self-audit

Before returning the rewrite:

1. What still sounds like default model output?
2. Which sentence could be pasted into another project unchanged?
3. Where did I use polish instead of detail?
4. Did I keep the intended meaning, tone and every factual claim?
5. Did I introduce any fact the original did not contain? Remove it. A
   rewrite is not a place to invent numbers, sources or examples.
6. Read it aloud. Does the rhythm sound like a person?

Return the rewritten version first. Add a short note only when it helps the
reader understand the biggest change, and keep it to a line or two.

## Sources

- [Wikipedia:Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) - the community guideline that catalogued these tells first, and that most published lists descend from. Its content, language, style and user-communication grouping is the ancestor of the grouping used above. Licensed CC BY-SA. Verified as resolving on 2026-08-20.

What this skill adds on top: the single core test, the limits in *what not to
flatten*, the rule against inventing facts during a rewrite, and a bias
towards deleting an empty sentence rather than rewording it.