---
name: captain-hook
description: >
  Use when producing ANY human-facing deliverable Mason could be scrutinized over
  — conference abstract, scholarship/application essay, cover letter, manuscript or
  paper draft, public-facing article/blog post, Slack/email message, bio, grant
  text. Run it BEFORE presenting or finalizing the deliverable. Rewrites the text
  to read as competent, authentic human writing -- a seasoned academic, not an LLM who skims topics and then gathers buzzwords and colorful language in an attempt to portray themselves like an expert (REAL seasoned academics will see right through it)-- (strips AI tells, matches Mason's
  voice) and adversarially checks it with a second model.
version: "1.1.0"
---

_Last updated: 2026-09-09_

# captain-hook

The captain of the hooks. Its job: take any deliverable and make it read like a
sharp, competent human wrote it — not a language model. This is the thing to run
before Mason sends an abstract, essay, article, message, or manuscript anywhere
it'll be read critically.

**Core principle:** AI prose fails the same way every time — evenly hedged,
over-signposted, abstract, list-happy, relentlessly smooth. Human writing has
*specificity, a point of view, varied rhythm, and restraint*. You fix AI-sounding
text by adding the human traits and cutting the tells — not by scrambling words.

**What we want (Mason, 2026-09-09):** prose does not have the unusually steady
sentence rhythm often associated with formulaic generated academic writing. This
is a requirement, not a preference. Check it, don't eyeball it: split the
deliverable into sentences, count words per sentence, and look at the spread. A
passage where nearly every sentence lands in the same 18–28 word band fails even
if every phrase is clean. Fix by varying — a short declarative against a long
one — never by chopping everything short, which is its own tell. Report the
spread (min / max / how many sentences under 10 and over 30) in the change-note.

## When to use / not

**Use for:** abstracts, essays, application/scholarship/grant text, cover letters,
manuscripts, papers, public articles, bios, and any message someone might judge as
"AI-generated." Run it as the LAST step before presenting. This applies to inline
chat deliverables too — an abstract or message drafted straight into the
conversation still needs the pass even though no file was written (the file-write
hook won't fire for those; a request-time hook nudges, but the responsibility is
yours to run it).

**Don't use for:** code, config, commit messages, internal notes, or anything
Mason won't be scrutinized on as a writer.

## The loop

Run this every time. Don't shortcut to a single pass.

```dot
digraph captain {
  "Draft / receive deliverable" [shape=box];
  "Voice + register match" [shape=box];
  "Kill the tells" [shape=box];
  "Depth & command pass" [shape=box];
  "Red-team: detector + domain expert" [shape=box];
  "Reads human AND expert?" [shape=diamond];
  "Present + change-note" [shape=box];
  "Draft / receive deliverable" -> "Voice + register match" -> "Kill the tells" -> "Depth & command pass" -> "Red-team: detector + domain expert" -> "Reads human AND expert?";
  "Reads human AND expert?" -> "Kill the tells" [label="no — texture flags"];
  "Reads human AND expert?" -> "Depth & command pass" [label="no — substance flags"];
  "Reads human AND expert?" -> "Present + change-note" [label="yes"];
}
```

### 1. Voice + register match first

Match Mason's actual voice, don't impose a generic one. Pull samples from their
writing (`~/article-ideas/*.md` drafts, prior essays/abstracts they point to). Note
their sentence length, how blunt they are, whether they use contractions, dashes, humor,
first person. If you have no sample for this genre, ask for one (or one paragraph
they like).

But voice serves the genre — not the other way around. For academic and formal
work (abstracts, manuscripts, grants, cover letters), the field's *register*
outranks personal voice: the target is Mason-*as-academic*, not their blog voice. An
abstract that reads like a casual post fails even if it's unmistakably theirs. When
you can, anchor to a second template too — a strong, well-regarded piece in the
*same subfield and genre* — and match its structure, register, and level of
precision alongside Mason's voice. Write toward *that*, not toward "good essay
voice."

### 2. Kill the tells

**2a. Run the blacklist first -- mechanically, before judgment.** Mason keeps a
living list of phrases that have shown up in my own drafts and read as AI to them:
`~/ionosphere-workspace/masonborchard.com/data/blacklisted-ai-sounding-phrases.md`.
Read it every run (it grows), then grep the deliverable for every entry -- the
literal phrases AND the templated ones ("It's not X, it's Y"; "the X is Y, the Z
that Zs it isn't"; the trio-of-claims). Zero hits is a gate: fix each one before
the red-team step, and mention the sweep result in the change-note. When Mason
flags a new tell in conversation, append it to that file the same turn. If a
blacklisted word is a genuine term of art in context (e.g. "robust" in
"cluster-robust standard errors"), reword anyway -- their list outranks the
convention, and there is always a synonym.

**2b. Then the general tells:**

Remove these on sight — they are what detectors and skeptical readers latch onto:

- **Filler openers / signposting:** "It's worth noting," "It's important to
  understand," "In today's fast-paced world," "In the realm of," "When it comes
  to," "That said," stacked "Moreover/Furthermore/Additionally."
- **AI vocabulary:** delve, tapestry, testament, realm, landscape, navigate (fig.),
  leverage (as verb), seamless, crucial, pivotal, meticulous, multifaceted,
  ever-evolving, foster, showcase, elevate, unlock. Cut on sight. (Two judgment
  calls: *robust* and *underscore* are real terms of art in technical writing —
  "robustness checks," "the results underscore" — keep those, cut the decorative
  uses.)
- **Filler hedging:** "can be seen as," "arguably," "in many ways," "to some
  extent," sprayed everywhere. Cut the empty ones and state things. Do NOT strip
  *load-bearing* hedges, though — in academic prose, precise epistemic caution
  ("is consistent with," "suggests," "we did not observe," "under these
  conditions") is a mark of expertise, not a tell. The skill is knowing which is
  which.
- **Uniform rhythm:** every sentence 15–25 words. Break it — mix a 5-word sentence
  against a 30-word one. Burstiness is the single strongest human signal. This is a
  hard requirement (see "What we want" above): measure the words-per-sentence spread
  and fix it before the red-team step, and report the numbers in the change-note.
- **Rule-of-three overuse:** endless "X, Y, and Z" triads. Vary list lengths; use
  a two-item or a single vivid example.
- **Perfect parallelism / bow-tie conclusions:** the tidy "In conclusion, X is a
  powerful tool that..." Kill it. End on a concrete point.
- **Abstraction:** replace generic claims with a specific detail, number, name, or
  example. Specificity is the fastest way to sound human.
- **Over-politeness / evenness:** take a position. A mild opinion or a small
  concession ("this is fiddly," "I got this wrong twice") reads human.

Preserve meaning, facts, and structure. This is a rewrite for texture and voice,
not a content change.

### 3. Depth & command pass

This is the step that separates "sounds human" from "sounds like a seasoned
academic" — and it's the one most humanizers skip. The tell a real specialist
catches isn't rhythm; it's *shallowness*: correct-sounding but generic claims,
buzzword-level instead of mechanism-level, prose that floats free of the actual
debate. Texture edits can't fix that. This pass targets it — inside a hard
guardrail.

Do:

- **Push generic → specific mechanism.** Where a sentence gestures at a concept,
  name the actual mechanism, method, result, or tension the author clearly means.
  "This approach improves performance" → *what* improves, and *why*.
- **Use the field's real terminology — precisely and sparingly.** A correct term
  of art used correctly signals command; buzzwords stacked for color signal a skim.
  One precise word beats three impressive vague ones.
- **Engage the actual conversation.** Position the claim against what the subfield
  already believes — the prior result it extends, contradicts, or qualifies.
  Generic prose is unplaceable; expert prose is situated.
- **Restore scope and appropriate hedging.** Add the caveats and scope conditions a
  careful expert would state. Seasoned academics underclaim.

**Hard guardrail — do NOT cross:** the substance must be genuinely Mason's. Never
fabricate a result, citation, mechanism, credential, or claim to sound more expert.
If the writing is thin because the underlying material is thin, you cannot fill it
with invention — instead, **flag the specific spots to Mason** ("this paragraph
reads skimmed — what's the real mechanism? which prior work does it engage?") and
let them supply the depth. Sharpen what's there; surface what's missing; invent
nothing.

### 4. Second-model red-team (the captain's edge)

Get an INDEPENDENT model to attack it — one model editing its own output can't see
its own tells. Run TWO personas: an AI-detector and a domain expert. Use the OpenAI
bridge:

Write the request to your session's scratchpad directory (listed in your system
prompt — session-specific, so never copy a literal path from an old transcript).
With `$SCRATCHPAD` standing in for that directory:

```bash
cat > "$SCRATCHPAD/cap_req.json" <<'JSON'
{
  "system": "You are two independent reviewers of the text below. (1) A skeptical human editor and AI-detection red-teamer: score 0-100 for how AI-generated it reads (0 = obviously human, 100 = obviously AI) and quote the specific words, phrases, and sentences that give it away. (2) A seasoned domain expert in this field: judge whether the author actually commands the material or is skimming it — quote any claim that reads generic, buzzword-level, or disengaged from the real debate, and say what a specialist would expect there instead. Be concrete; quote everything. Do not rewrite.",
  "messages": [{"role": "user", "content": "<the humanized deliverable>"}]
}
JSON
python3 ~/.claude/skills/debate-gpt-brad/gpt_brad.py ask < "$SCRATCHPAD/cap_req.json"
```

This uses the OpenAI bridge — if `OPENAI_API_KEY` isn't set, `gpt_brad.py` errors;
export it first (same as debate-gpt-brad). Feed the flags back by type: texture
flags into step 2, substance/expertise flags into step 3. For a high-stakes piece,
run the two personas as SEPARATE calls so neither softens the other. Repeat until
the AI-likeness score is low, the expert can't find a skim, and the rest are
nitpicks. (Write the JSON with the Write tool if the text has tricky quoting.)

### 5. Present with a change-note

Edit the deliverable in place (for files) or present the humanized version inline.
Add a 2-3 line note: what you changed and why (e.g., "blacklist sweep: 2 hits
fixed; cut 6 filler hedges, broke up even rhythm, named the mechanism in the
results paragraph, matched your academic register") — and separately, flag any spots where the substance read thin so Mason
can supply real depth (never invented). Never claim it will defeat a specific
proprietary detector — no tool can promise that. What you CAN say: it now reads as
competent, expert human writing and the obvious tells are gone.

## Reality check

No editor guarantees passing every closed-source detector, and honest writing
still occasionally trips them. The durable win is real: specific, voiced, varied
prose is both harder to flag AND simply better writing. Optimize for that.

## Common mistakes

- **Skipping the blacklist sweep.** The list is Mason's own detector, calibrated on
  my drafts. A deliverable that passes the second-model red-team but carries a
  blacklisted phrase still fails -- they will see it first.
- **One pass and done.** The tells hide; the second-model red-team catches what you
  missed. Loop.
- **Generic "polished" voice.** If it doesn't sound like Mason, it failed — match
  a sample.
- **Word-swapping without rhythm change.** Uniform cadence flags even with a clean
  vocabulary. Vary sentence length.
- **Humanizing shallow substance.** Smooth, voiced prose over thin content is the
  loudest academic tell — an expert sees the skim instantly. Run the depth pass;
  flag the gaps for Mason instead of papering over them.
- **Fabricating depth.** Never invent a result, citation, or mechanism to sound
  more expert. Sharpen genuine substance; surface what's missing. The substance
  stays Mason's.
