ai-eos

Files (1)

• SKILL.md

SKILL.md (first lines)

# AI Elements of Style

Hunt the stylistic tells that mark a piece of technical prose as AI-generated, and report them. Edit nothing.

The counterpart to `eos`: where `eos` holds prose to Strunk & White, this holds prose to the tells that a machine wrote it. Same target, different standard.

Invoked directly as `/phil:ai-eos`, or as one of the three passes in `/phil:red-team-prose`, which adds an `eos` clarity pass and a document-fitness pass over the same text.

Take an adversarial stance: assume the text is trying to pass as human-written and look for what gives it away. But hold the opposite discipline just as hard — **a tell is a specific span you can quote, name, and fix.** Vague suspicion is not a finding.

This skill owns **LLM tells only**. Clarity, concision, and voice belong to `eos`; document fitness for audience and purpose belongs to `${CLAUDE_PLUGIN_ROOT}/rules/technical-communication.md`; comment quality belongs to `clean-comments`. Do not review against those here.

**Density beats any single token.** The sources cited below all treat frequency, not single tokens, as the signal. One "pivotal" proves nothing; four mechanisms clustered in one paragraph is damning. Lead with the density verdict, then the spans.

## Sources

The catalog below summarizes [Wikipedia's *Signs of AI writing*](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) (WikiProject AI Cleanup, CC BY-SA), the excess-vocabulary measurements in Kobak et al., *Delving into LLM-assisted writing in biomedical publications through excess vocabulary* (Science Advances, 2025) and Liang et al., *Mapping the Increasing Use of LLMs in Scientific Papers* (COLM, 2024), and George Orwell's "Politics and the English Language" (1946) on the dying metaphor and the ready-made phrase.

---

## Parse the Argument

Determine what `$ARGUMENTS` refers to:

| Pattern | Type | Example |
|---------|------|---------|
| `--changes` | Uncommitted changes | `--changes` |
| Digits separated by `-` or `:`, with an optional `lines:` prefix | Line range | `42-67`, `42:67`, `lines:42-67` |
| Has a file extension | File path | `README.md`, `docs/setup.md` |
| Ends with `/` or a directory with no extension | Directory path | `docs/`, `skills/` |
| No argument | Prose in the current context | — |

If the argument contains both a file and a line range (`docs/setup.md:12-30`), split them.

---

## Step 1: Gather the Prose

Review prose only. Skip code blocks, frontmatter, config, link URLs, and generated output. Review table cells only when they contain sentences.

### `--changes`
Run `git diff --name-only HEAD` to list files with uncommitted changes, filter to prose files (`**/*.{md,txt,rst}`), then run `git diff HEAD -- {file}` per file to get the hunks. Review only the added and modified lines — but read the whole file for context, since density is a property of the passage, not the diff. The density denominator is the reviewed lines, not the file. Pass `git diff` neither `--output` nor `--ext-diff`: the first writes any path, the second runs a repo-configured diff driver, and this command declares `mutates: false`.

### Line Range
Read the file with surrounding context. Review the selected block only.

### File Path
Read the entire file.

### Directory Path
Glob recursively for `**/*.{md,txt,rst}`. Review each.

---

## Step 2: Hunt the Tells

Two tiers. **Tells** are over-the-top enough to flag on a single occurrence. **Whiffs** are weak signals — flag them only when they cluster with tells or with each other.

For every finding, name the **mechanism** (why an LLM produces it), not just the phrase. The mechanism is what makes the finding actionable and what keeps the catalog useful as models change.

**Which tier a phrase belongs to follows one rule:** a bare single word lives in the whiff tier, because one correct use proves nothing; a multi-word construction lives in the tell tier, because the construction itself is the fingerprint. "Robust" is a whiff; "underscores the importance of" is a tell. The rule holds except where a tell names its own single words: #5 counts its verbs as a ratio, and #8 flags its words because the register, not the word, is the fingerprint.

### Tell tier — flag on sight

**1. Unearned significance**
Inflates a mundane fact into grand importance. The subject always "stands," "serves," or "plays a role" rather than simply doing something.

> "stands as a testament to" · "plays a crucial/pivotal/vital role in" · "underscores the importance of" · "marks a key turning point" · "reflects a broader shift toward" · "has become synonymous with"

*Fix:* cut, or replace the claim with the specific thing the subject actually does.

**2. Participle padding**
A trailing `-ing` clause bolted to a finished sentence, adding cadence and no information. The single most reliable tell in technical prose.

> "..., highlighting its importance in modern workflows" · "..., ensuring scalability and reliability" · "..., reflecting the team's commitment to quality" · "..., contributing to overall performance"

*Fix:* cut. If the clause carries a real claim, promote it to its own sentence with a subject.

**3. The formulaic closer**
A wrap-up that concedes a vague weakness and pivots to vague optimism. Often arrives unbidden at the end of a section that needed no conclusion.

> "Despite these challenges, [subject] remains well positioned to..." · "While not without limitations, ..." · "In conclusion," · "Overall," · "As the landscape continues to evolve, ..."

*Fix:* delete the paragraph. If a real limitation exists, state it plainly and stop.

**4. Phantom attribution**
Opinion attributed to an unnamed consensus, inflating one source — or zero — into a field.

> "Experts argue that" · "Observers have noted" · "Industry reports suggest" · "It is widely regarded as" · "Many developers find" · "Critics have pointed out"

*Fix:* name the source, or cut the claim. Flag this as **needs a fact you cannot supply** — never invent the citation.

**5. Copulative avoidance**
Systematic refusal of plain *is* and *has*. One instance is style; a pattern across a passage is a fingerprint.

| Tell | Plain form |
|------|-----------|
| serves as / functions as / acts as | is |
| represents / constitutes / embodies | is |
| marks / stands as | is |
| features / boasts / offers / maintains | has |
| enables / allows for | lets, or the actual verb |

*Fix:* use *is* and *has*. Count occurrences and report the ratio — that's the evidence. Report the ratio only; do not also file each avoidance as an individual finding unless it is the sole tell in the passage. (Bare "facilitate" as a fancy word belongs to `eos`, not here.)

**6. Negative parallelism**
Defines by contrast with a strawman, for rhythm rather than meaning.

> "not just X, but Y" · "not only X but also Y" · "it isn't about X — it's about Y" · "not a X, but a Y" · "X rather than Y" (when nobody proposed X)

*Fix:* state Y. Drop the contrast unless a reader would genuinely have assumed X.

**7. Rule-of-three padding**
Three coordinate items where one carries the meaning and two pad the cadence. Especially adjectives.

> "fast, reliable, and scalable" · "clear, concise, and actionable" · "designing, building, and maintaining"

*Fix:* keep the item that does work; cut the rest. Legitimate triads name three distinct things — flag only when the members are near-synonyms or the third is filler.

From pmvanev/phil-claude-plugin, MIT.

More from pmvanev/phil-claude-plugin