not-but

name	not-but
description	Find and defuse "not X but Y" / "isn't X; it's Y" / "X — not Y" constructions. Deterministic grep for detection (no LLM blind spot), the not-but subagent for triage. Split-or-cut tail negations ("X, not Y") aggressively — promote the point to its own sentence or delete it; recast in place only mid-clause "not X but Y" that genuinely can't be split; keep just the rare iconic thesis line literal.
argument-hint	<file_path>
allowed-tools	Read, Edit, Grep, Bash, Agent

Not-but: Negative-Parallelism Hunter

LLMs have a structural blind spot for their own patterns. "Not X but Y" is the most common AI tic and the slipperiest one — pattern-matching by another LLM misses it as often as the original write did. This skill uses deterministic grep for detection (no blind spot) and an LLM only for triage (where judgment is needed).

Sister skill to /humanize. Where humanize scans broadly across many patterns, this drills on the one pattern that slips past humanize most often.

The construction is a tell even when the contrast is real. Human readers are turned off by the shape of "not X but Y" regardless of whether the X-versus-Y distinction is load-bearing. This is the insight that makes the skill more than a filler-cutter: a triage that stops at "earned vs dead-weight" leaves a dense paper riddled with earned negations that still read as machine-written. So triage has three outcomes, not two:

• Dead-weight — the negation is filler; Y stands alone. Cut it. Two sharpeners for tail-attached negations (X, not Y / X — not Y), which are often remnants of an earlier not-but recast: (a) judge redundancy at paragraph scope, not sentence scope — a tail can parse fine in its own sentence and still be dead because the paragraph already planted the positive; (b) the own-sentence test — if the negated point is important enough to keep, it earns its own sentence; if it cannot stand as one, cut it. A tail that is neither paragraph-novel nor sentence-worthy is dead-weight.
• Iconic — a thesis line whose force is the antithesis ("True and false are not opposites but siblings"). Keep the literal construction. These are rare; protect them.
• Earned-but-varyable — the contrast is real and must survive, but the syntax should change. Keep the contrast, recast the construction. Never delete the distinction. For mid-clause not X but Y this is the common case; for tail forms it mostly collapses into split-or-cut (next paragraph).

The earned-but-varyable bucket is the skill's main yield on already-tight prose, where almost everything triages as "earned" under a two-way rubric and nothing gets fixed. The density of the construction is the tell, not any single instance.

Tail negations get the aggressive default: split or cut, not recast. AI tics are easy enough to spot now that the shape costs more than any cleverness it carries, so overcompensate toward removal. For tail-attached forms (X, not Y, X — not Y, X, not just Y), do not swap one tail for another contrastive (instead of, where X, Y) and call it fixed. Decide each by the own-sentence test: if the negated point earns keeping, promote it to its own sentence (X. It is not Y. / X. Y already holds.); if it cannot stand as a sentence, cut it. Recast-in-place via the palette is reserved for mid-clause not X but Y that genuinely cannot be split. And the iconic bucket is tiny — often a single line in a long document, rarely more than three. A 16k-word paper holding exactly one (Merit attaches to the work, not the doer) is the calibration, not six-to-ten.

Process

1. Read the file at the given path. Note the word count.

2. Detect (deterministic). Run grep with these patterns. Each captures a distinct shape of negation. Run them all; collect every hit with line number and matched line:

   FILE="<file_path>"
   
   # Same-sentence: subject + not + complement, then pivot
   grep -nE '\b(is|are|was|were|am|be|been)\s+not\s+\w+[;,—]\s*(it|they|that|this|but)\b' "$FILE"
   
   # Cross-sentence opener: "It's not...", "This isn't...", "That's not..."
   grep -nE "\b(It'?s not|This isn'?t|That'?s not|There is no|There are no)\b" "$FILE"
   
   # Tail-attached negation: "X — not Y" or "X, not Y"
   grep -nE '[—,]\s*not\s+\w+' "$FILE"
   
   # "Not just / only / merely / simply" intensifiers
   grep -nE '\bnot (just|only|merely|simply)\b' "$FILE"
   
   # Contracted negative copulas (often pivot cross-sentence)
   grep -nE "\b(isn'?t|aren'?t|wasn'?t|weren'?t)\b" "$FILE"
   

   Skip hits inside fenced code blocks, tables (lines starting with |), HTML tags, and frontmatter. Deduplicate by line number.

3. Triage (subagent). Dispatch to the not-but subagent (defined in ~/agents/not-but.md, or fallback to general-purpose if not registered). Pass:

   • The file path
   • The deduplicated list of hits with line numbers and matched lines
   • The rubric, in two questions: (1) Can the sentence be replaced by just stating Y, with no loss of meaning? If yes → dead-weight. Judge this at paragraph scope, not just the sentence; for tail-attached negations also apply the own-sentence test (see the three outcomes above). (2) If no (the contrast is real), is this a thesis line whose whole force is the antithesis? If yes → iconic (keep literal); if no → earned-but-varyable (keep the contrast). Most real contrasts in dense prose are varyable, far from iconic. For tail-attached hits (X, not Y / X — not Y), a varyable verdict means split-or-cut, not recast: promote the negated point to its own sentence if it earns keeping, else delete the tail. Reserve in-place palette recasts for mid-clause not X but Y that cannot be split into two clean sentences.

   The subagent reads context, applies the rubric, and returns a per-hit verdict of dead-weight / iconic / earned-but-varyable (see the three outcomes above). Pass it two more things:

   • The iconic-keep list for this document, if known (the handful of thesis lines whose antithesis is the point). When unknown, instruct the subagent to nominate them conservatively and default everything else to varyable.
   • The substitute palette (below), with the standing instruction: for every dead-weight AND every earned-but-varyable hit, propose a recast; vary the device across hits so the fix does not migrate the tic onto one replacement.

   The subagent returns, per hit: the verdict, and for dead-weight/varyable hits a proposed recast naming which palette device it used. The subagent does NOT apply edits.

   The substitute palette (keep the contrast, change the construction; the point is variety, so no single device should dominate):

   • positive recast that implies the contrast ("X is no new metaphysics. It specifies...")
   • period split / "X. It is Y" / "X; it is Y"
   • "instead of", "as opposed to", "in contrast to"
   • subordinating clause: "while X, Y" / "X while leaving Y untouched" / "where X, Y"
   • "without Ying", "with no Y", "never Y" (single negated tail, milder than the full antithesis)
   • "though no more Y", "less X than Y", "more than X; it is Y"
   • full negated clause: "It does not mean Y" / "None attests Y"
   • AVOID defaulting to "rather than" — it is itself an overused contrastive in argumentative prose, so leaning on it just moves the tell. Check the document's existing "rather than" count first.

4. Apply. Three different decisions, three different sources of authority. The skill enforces this separation because it is the source of the skill's value.

   • Change decision (mechanical). Once the subagent verdicts a hit as dead-weight or earned-but-varyable, the literal "not X but Y" construction goes. Dead-weight → cut the negation; varyable tail → split into its own sentence or cut; varyable mid-clause → recast via the palette. No re-litigation by the orchestrator. Letting the orchestrator re-judge re-introduces the LLM blind spot the deterministic detection was designed to defeat — the orchestrator has read the full document, built attachment to the prose, and will rationalize "actually this one reads fine" more often than the rubric warrants.
   • Phrasing of the recast (orchestrator discretion). The subagent's proposed rewrite is a draft. The orchestrator has the full document context the subagent didn't and can refine wording, merge adjacent fixes, restructure sentence boundaries, match surrounding rhythm, and — for varyable hits — swap one palette device for another to keep the document's overall device mix diverse. The one rule: do not restore the bare "not X but Y" shape. Taste lives here. (A recast that would be clunkier than the original is the one exception: kick it back to KEEP and note why. A varyable verdict is not a mandate to make the prose worse.)
   • Re-classification as iconic (user only). Not the orchestrator's call. If the orchestrator genuinely thinks a varyable verdict should be left literal (a thesis line the subagent missed), surface it in the report — not in the apply — so the user makes the final call.

   Then Edit the file. Dead-weight hits lose the negation; varyable hits get the recast; iconic hits stand. Across the varyable set, watch the aggregate: if three recasts in a row used "; it is", vary the third.

5. Report (designed to feed /humanize's earn-back surface):

   • Before/after word count
   • Count of hits: dead-weight, earned-but-varyable, iconic, false-positive
   • Applied changes as a numbered list. Each entry: line number, original text, applied rewrite, and (for varyable hits) the palette device used. Changes where the orchestrator had reservations get marked with ⚠ to draw the eye first.
   • A one-line device tally for the varyable set (e.g. "period-split ×3, in-contrast-to ×2, never ×2, ...") so over-reliance on one substitute is visible at a glance.
   • Iconic (kept-literal) cases, with line numbers, so the user can override.

   When /not-but runs under /humanize, /humanize uses the applied-cuts list to build its post-apply earn-back report and offer undo N (or undo 2,4,7) to restore specific cuts. When /not-but runs standalone, present the same list as the final output and accept the same undo affordance directly.

6. Verify: re-grep after edits. Every remaining tail negation should be on the tiny iconic kept-literal list — varyable tails are split or cut and varyable mid-clause hits are recast, so both drop out of the grep. A surviving X, not Y that is not iconic means a false-negative in the apply step (or a hit the original grep missed — re-run the full pattern set, since clustered tails hide each other on a long line). Note either for skill iteration.

Composability

• Standalone: /not-but <file> for any file.
• Inside /copyedit: invoke after /humanize as a forced check. Negative parallelism slips past humanize often enough that a deterministic, dedicated step earns its slot in the loop.
• Batch mode: pass a directory; greps run across every prose file at once, subagent triages each file independently.

Why deterministic detection

The patterns are aligned with the LLM's training distribution, so they read as natural to the LLM scanning for them. A regex has no aesthetic preference. False positives are fine — the triage step filters them. False negatives are not — and grep doesn't have any.

The /cord vs /cord-human contrast (referenced in /humanize) is the calibration set: re-running this skill on src/content/blog/2026-02-18-cord.md should converge it toward src/content/blog/2026-02-21-cord-human.md on the negation axis.

Pattern catalog (what the regexes target)

• is/are not X; it's Y (same-sentence)
• isn't X. It's Y (cross-sentence)
• not X but Y
• X — not Y (tail-attached em dash)
• X, not Y (tail-attached comma)
• not just X — Y too
• not only X but also Y
• not merely X / not simply X
• there is no X. The Y is... (paragraph-scope negation-then-assertion)