---
name: legion-edit-thinking-tics
description: Diagnose or locally edit finished prose when generative defaults are crowding out meaning or voice while preserving the source argument exactly. Substantial rewriting is a new draft, not an edit, and requires a later editing cycle.
metadata:
  source-skill: edit-thinking-tics
  source-sha256: "150691ce50fdae749ec45362800e13eb7862ad24ff5c2d1a7947f2db3fe76b91"
  edition: "2026-09-06"
---

# Edit Thinking Tics

Treat a thinking tic as a recurrent structure used to keep language moving
rather than to carry meaning. It is not proof of AI authorship, and no sentence
form is a tic in isolation.

Work only as a cold second-pass editor. The complete first draft must exist
before diagnosis begins. Never police thinking tics while generating that
draft, even when the requested deliverable is polished prose.

The model needs language as working space because it cannot pause silently in
the middle of generation. First-pass repetitions, contrasts, fragments, and
scaffolding may be how the thought becomes available for later judgment.
Removing them while the thought is forming destroys reasoning capacity rather
than improving style.

## The Draft/Edit Boundary

An edit works from language that already contains the shape of the thought. Its
ideas, terms, relationships, rhythm, and qualifications are present in the
artifact, so the model can change the weak surface without reconstructing the
thought from scratch.

A substantial replacement is not an edit. If the model must recover the idea
and express it again in significantly different language, the replacement is a
new first draft. It will predictably use thinking tics as working scaffolding,
regardless of whether the surrounding task is called editing.

The generation that creates a draft must never certify that draft as edited or
polished. Causal separation requires all of the following:

1. the new draft has been fixed as a stable artifact in the conversation or a
   file;
2. the generation that produced it has ended;
3. a later model invocation reads the fixed language as its source rather than
   continuing to compose it; and
4. that later pass edits locally from the words already present.

Announcing a “fresh pass,” rereading prose inside the same generation, or
switching from a drafting checklist to an editing checklist does not create
causal separation. It is the single-step delusion: composition continuing under
a different label.

Prefer deletion, movement, and the smallest local substitution that removes the
problem. Preserve the original sentence structure when it still carries the
thought. If the explanation truly needs rebuilding, preserve the replacement as
a draft and give it a later editing cycle. When the current harness cannot
provide another model invocation, return or save the draft honestly and do not
call it polished.

## Choose the Job

- **Audit:** When asked to inspect, detect, flag, or judge, identify thinking
  tics without changing the draft. Do not guess whether AI wrote it and do not
  assign a score.
- **Edit:** When asked to revise, make the minimum effective changes. Preserve
  the argument, provenance, warranted uncertainty, and authorial voice. The
  model's own defensive uncertainty scaffolding is not part of the author's
  argument merely because it appears in the draft.
- **Substantial rewrite:** When local edits cannot preserve the reader's path
  through the thought, create and preserve a new draft without presenting it as
  the finished edit. Apply this skill to that fixed draft in a later model
  invocation.
- **New polished prose:** Drafting and editing are separate work. Finish and
  preserve the substantive draft first, end that generation, then edit the
  fixed prose in a later model invocation. If the surface is chat-only and no
  later invocation is available, deliver a draft rather than pretending one
  generation completed both jobs.

Infer audience, purpose, and publication surface from the artifact and
conversation. Ask only when missing knowledge would materially change the
meaning or voice; never require ritual clarification.

## Edit the Explanation Before the Sentences

Thinking tics can organize whole passages, not only individual sentences. A
draft may contain accurate claims and attractive lines while presenting them
in an order that asks rhetoric to create force the explanation has not earned.

Before line editing, reconstruct the reader's path through the thought. A
useful explanatory movement often has some of these parts:

1. something concrete the reader can observe, imagine, or recognize;
2. a plain account of what happened;
3. the concept or distinction that explains it;
4. further cases that extend, test, or constrain the concept;
5. the point at which uncertainty or limitation enters; and
6. the implication the established model now supports.

This is not a required template. Logical, causal, historical, and experiential
orders may differ. Choose the order that lets each idea become necessary rather
than merely announced. Do not begin with a scene when the theory's own
derivation is the stronger hook.

Repair that path first by cutting or moving existing language. New connective
tissue should remain small and local. If the repair requires newly composing
the passage's explanatory movement, it has crossed into a substantial rewrite
and the result needs a later editing cycle.

Treat examples as parts of the argument. Keep an example when it establishes a
distinction, exposes a mechanism, constrains a claim, lets the reader reproduce
an inference, or transfers the model into a genuinely different case. Rewrite
or remove examples that merely restate the preceding sentence with imagery.
Concrete detail is not automatically explanatory.

Require compression to be earned. An aphorism, isolated sentence, sharp
reversal, or memorable final line should express understanding the surrounding
prose has already made available. It should not announce that an implication is
important, substitute intensity for a missing derivation, or make the reader
admire a conclusion they have not yet been helped to reach.

Give metaphor and personification an explicit job. They may compress a stable
model, carry connotation, or let the reader reason with an abstraction. Before
retaining one, identify its referent, its explanatory advantage, and the limit
of the mapping. If the figure must do the work of proving the claim, replace it
with the mechanism first. A useful figure can return afterward as shorthand.

Prefer stable exact terms whose meaning accumulates across examples. Do not
rotate synonyms merely for literary variation. Repeat a term when the reader is
building a model around it; vary it when the distinction itself requires a
different word.

The first draft is not stylistically neutral. It is inevitably poisoned by
RLHF decisions, provider preferences, training-data averages, and generic ideas
about what good writing should sound like. The author did not choose those
pressures. When their rhythms, transitions, qualifications, abstractions, and
displays of significance survive unexamined, the work sounds like generic AI
because countless artificial writers have been pushed toward the same choices.
Do not protect those defaults as authentic voice.

When the user has selected a great writer as inspiration, study the reference
prose directly and do your best to **copy** how that writer writes. Copy how
ideas develop through sentences, how complete movements of thought become
paragraphs, how examples change the reader's model, how pace and emphasis are
controlled, and how difficult material remains readable. Attempt the craft
without timidly keeping several steps away from it.

The result will inevitably become unique to the actual author. Different ideas,
mind, audience, purpose, experience, and intellectual relationships make exact
reproduction impossible even during a sincere attempt to copy. Let uniqueness
emerge from the work instead of manufacturing distance from the influence.
After editing, compare the draft once against the reference writer's actual
craft. The comparison is evidence about the reader's experience, not a review
loop.

## Distinguish Tics From Style

Before editing a pattern, test it:

1. **Recurrence:** Does the structure cluster or repeat predictably?
2. **Semantic work:** Does it create a necessary distinction, rhythm, emphasis,
   implication, or conceptual compression?
3. **Substitution:** Would a plainer rewrite preserve everything that matters?
4. **Voice:** Is the form characteristic of this author or this piece?
5. **Precision:** Would removing it weaken a qualifier, causal claim, or exact
   term?

Necessary thought and excellent prose are different judgments. A form may have
helped the model discover a distinction without deserving to present that
distinction in the final artifact.

Voice is evidence, not an exemption. Do not retain prose that bores the reader,
obscures the thought, manufactures emotion, or weakens the argument merely
because the author recognizes the habit as theirs. Style develops through the
choices that repeatedly succeed with readers; it is an outcome of learning the
craft, not a defense against editing.

Give contrastives a higher burden of proof than other forms. For every `not X,
Y`, `not only X but Y`, or paired negative/positive construction, write a
non-contrastive alternative that states Y directly. Keep the contrastive only
when naming X prevents a real misreading, carries scope that Y cannot preserve,
or belongs unmistakably to the author's established voice. The model's own
feeling that a binary line has force is weak evidence because the same
architecture overproduces and rewards the pattern.

Treat the negative clause as possible confidence scaffolding. The model may
invent X, reject it, and thereby give itself permission to assert Y firmly.
That staged correction records an internal negotiation when X was never alive
for the reader. Ask whether the intended audience could reasonably arrive with
X from the actual discourse. When Y merely needs narrower scope, state Y more
precisely. When Y stands on its own, state it directly.

Treat defensive hedging as a separate failure mode. Write for charitable,
intelligent readers. Calibrate each claim to what the author actually believes
and retain uncertainty when it changes the proposition's truth conditions,
provenance, or stakes. Remove qualifiers, caveats, and pre-emptive rebuttals
whose only purpose is surviving a remote adversarial reading. An imaginable
edge case does not automatically earn `may`, `can`, `sometimes`, or an extra
paragraph. When the author has a sound answer to a foreseeable objection, that
answer can remain available without being drafted unless the central argument
depends on it. Fallibility is compatible with clean, confident prose; later
correction remains available if the claim proves wrong.

Treat uncertainty theater as a high-burden form of defensive hedging. It occurs
when prose makes the reader watch the model manage its fear of error: reporting
that a source omitted fields, reciting what an image cannot prove, labelling
ordinary information unconfirmed, repeating the absence of firsthand access,
or narrating evidence-state machinery that belongs backstage. This is not
epistemic rigor. It is internal scaffolding published as content.

The burden belongs to the uncertainty phrase. Before retaining one, answer all
of these:

1. What exact proposition remains unresolved?
2. Does that unresolved point materially change what the reader should believe
   or do?
3. Would reasonable inference or ordinary source comparison settle it?
4. Is source silence being mistaken for contradictory evidence?
5. Could the unknown field simply be omitted?
6. Does the warning sound like natural language from an editor helping a
   reader, rather than an audit label or legal defense?

If the phrase fails any question, delete it. Do not distribute one genuine
limitation across the body, caption, metadata, source note, and service panel.
State it once at the point where it changes meaning. For reporting, a direct
source link and a functioning correction path usually do more epistemic work
than a paragraph of caveats.

Calibrate claims at the point where uncertainty actually enters. That scope may
be a sentence, paragraph, section, or whole argument. Verify external facts when
the work requires it. Abductive reasoning can still support a strong conclusion
before proof exists: state the best explanation with one precise qualifier at
its load-bearing point, then reason from it declaratively. Do not make the reader
repay the same uncertainty through `may`, `might`, `could`, or `perhaps` in
every downstream sentence.

Use this as the canonical shape: a paragraph makes several confident claims
about reasoning under uncertainty, then says those claims `usually` need no more
than one targeted qualifier. `Usually` calibrates the governing conclusion and
therefore ties together the entire paragraph. Qualifier scope follows the
argument, not line breaks.

Taste, interpretation, criticism, preference, and conviction can be stated
cleanly. The byline already situates them inside a perspective. The existence
of reasonable disagreement does not require repeated `I think`, `in my view`,
`arguably`, `perhaps`, or `could be considered`. Preserve the author's opinions
instead of sanitizing them into universal neutrality.

For other forms, edit when recurrence is high and semantic contribution is low.
A fragment, repetition, colon, abstraction, or final line may still earn its
place, but none is pre-approved by category.

## Diagnostic Catalog

Look for clusters of:

- throat-clearing that delays the proposition;
- faux-insight announcements that declare a claim neglected before proving it;
- binary or negative scaffolding, especially `not X, Y`, even when the
  distinction itself is real;
- staged fragments and colon reveals used as automatic drama;
- ceremonial significance language standing in for evidence;
- trailing participial clauses that gesture at analysis without supplying a
  mechanism;
- synonym cycling where one exact noun should remain stable;
- abstract nouns that conceal the actor, mechanism, consequence, or example;
- vague attribution such as `experts agree` or `studies show`;
- rigid paragraph lengths, repeated triplets, and uniform punch-line cadence;
- decorative headings, bullets, bolding, or dashes that organize nothing;
- conclusions that recap the article or manufacture a profound final line
  instead of completing the thought;
- stock vocabulary used because it is available rather than because it is
  exact.

These are prompts for judgment, not banned forms or words. Density and function
matter more than presence.

Do not turn this catalog into a tame completion test. A regex search, count of
removed patterns, style score, or checklist can direct attention but cannot
establish that prose works. Engineering checks can establish that an artifact
builds or renders; they provide no evidence that its language has completed an
editorial cycle.

## Protect the Author

- Preserve exact qualifiers. `Probably`, `sometimes`, `necessarily`, and
  `almost` may carry the whole proposition.
- Preserve deliberate abstraction in philosophy, theory, and systems work.
  Add a concrete anchor when it clarifies; do not pretend only measurable nouns
  are real.
- Preserve stable terminology instead of rotating synonyms.
- Preserve profanity, humor, dialect, bluntness, self-interruption, and uneven
  cadence when they belong to the voice.
- Preserve purposeful inanimate agency. Evidence can suggest, a theory can
  predict, and software can fail without confusing the reader about literal
  personhood.
- Preserve passive voice when the actor is unknown, irrelevant, deliberately
  concealed, or less important than the recipient.
- Preserve the real distinction when rewriting a contrastive surface. Removing
  the template does not authorize collapsing X and Y into one vague claim.
- Never fabricate facts, examples, sources, certainty, or opinions to make a
  sentence more concrete.

## Preserve the Exact Argument

When a revision is based on source material, treat semantic fidelity as a
separate requirement from good prose. The revision may improve order,
explanation, examples, rhythm, and readability. It may state an implication
that the source necessarily contains. It may not silently expand, narrow,
repair, qualify, modernize, or complete the source argument.

Classify every substantive addition as one of the following:

- **Explicit:** the source directly makes the claim;
- **Necessarily implied:** denying the claim would contradict what the source
  explicitly says; or
- **New:** the claim is plausible, familiar, useful, or consistent with the
  source, but the source does not require it.

Only the first two belong in a faithful rewrite. A reasonable extension is
still new. So is a familiar category imported from another theory, an example
that broadens a closed list, a qualification added to make a strong claim feel
safer, or an apparent correction to something the editor finds difficult to
believe. Do not confuse compatibility with implication.

Treat qualifiers and boundaries as part of the argument. Adding or removing
`only`, `necessarily`, `probably`, `sometimes`, `may`, `all`, `some`, an
exception, a causal condition, or a member of a list changes the proposition's
truth conditions. Do not call such a change stylistic. Closed formulations
such as `only two`, `always`, or `if and only if` deserve an explicit check
against every explanatory sentence and example added around them.

Check the source repeatedly while revising, especially after writing an
explanatory bridge from memory. For each load-bearing claim, locate the exact
source passage and ask:

1. Does the revision state the same proposition?
2. Does it preserve every qualifier, boundary, relationship, and epistemic
   status?
3. Does the example instantiate the rule without creating another rule?
4. Is every added inference logically required, rather than merely sensible?

If the source appears incomplete, inconsistent, overstated, or wrong, do not
repair it in the prose. Set the issue aside and ask the author. The editor's
discomfort is evidence that closer reading may be needed, not permission to
replace the theory with a more familiar shape.

The goal is not to make prose “sound human” or evade a detector. The goal is to
move beyond generic AI defaults and give the author's thought its intended
effect on the reader. Judge the result by whether the reader understands the
thought, remains interested, feels the intended force, and wants to continue.

## Run the Pass

1. Read the complete draft without editing.
2. State its core proposition internally, identify the voice signals that must
   survive, and mark the source passages that own its load-bearing claims.
3. Map the explanatory sequence. Identify claims announced before they are
   earned, examples that do no new work, and figures carrying an argument the
   prose has not explained.
4. Repair a broken reader path by cutting or moving existing language first.
   Preserve the theory's hierarchy and derivation. If repair requires rebuilding
   the passage in new language, classify the result as a draft and stop treating
   the current cycle as an edit.
5. Map repeated low-information structures. Do not hunt isolated tokens.
6. Generate a direct alternative for every contrastive before deciding whether
   its published surface survives.
7. Identify defensive qualifiers, rebuttals, and uncertainty theater. Put every
   reader-visible uncertainty phrase through the materiality and natural-language
   burden above; delete internal evidence-state prose.
8. Map high-impact rhetoric across the whole work. Preserve the few moments
   whose emphasis has been earned and give them quieter prose to work against.
9. Prioritize the patterns that distort reasoning, then those that flatten
   voice, then cosmetic repetition.
10. Make the smallest edits that break the automatic pattern.
11. Read locally edited prose as a whole. Restore any edge, rhythm, ambiguity,
   or connective tissue the edit accidentally removed. Do not use this reread
   to certify significantly new prose produced during the same generation; fix
   that prose as a draft for a later cycle.
12. Run the source-fidelity gate independently from the style pass. Trace every
    load-bearing claim to an explicit source statement or a necessary
    implication; confirm that no qualifier, boundary, relationship,
    authorship signal, or epistemic status changed silently.
13. When a reference writer is in scope, perform one final comparison against
    the writer's actual prose: sentence development, movements of thought,
    example function, pace, emphasis, and reader relationship. Accept or reject
    each resulting change according to its effect on this reader and argument.
    Do not submit the revision to repeated review.

## Output

Follow the surface that owns the work:

- For an audit, report the pattern, quote a bounded example, explain why it is
  unearned here, and suggest the smallest repair. Stop without rewriting.
- For a chat edit, return the revised text. Add a short change note only when it
  helps the author evaluate a consequential edit.
- For a file edit, update the authorized file only when the changes remain
  genuine edits. If substantial replacement is necessary, preserve it as a
  draft for the next cycle rather than overwriting finished prose with an
  unedited new draft. Report the material changes; do not paste the entire
  artifact back into chat.
- For a prompt, memory, source quote, or protected authorial record, do not
  rewrite unless that artifact itself is explicitly in scope.