Voice Guide System Audit — 2026-07-07

Created 2026-07-07

Status findings-for-review

Tags voicewritingauditgovernance

Requested by Patrick. Trigger: after several read-throughs of What’s Between the Notes, the tone still doesn’t fully match his voice despite 14 documented edit rounds. This audit uses that article as the use case, then examines the voice guide system itself: the guides, the process around them, and the published posts written with Claude.

Part 1 — Use case: “What’s Between the Notes”

The measurable finding: contrastive scaffolding at 5–10x the density of any other post

A pattern scan of all seven Claude-era blog posts (body prose only, editorial notes excluded) counted the constructions the unified guide’s AI Generation Warning names as failure mode #1: “X isn’t Y. It’s Z” openers, “not X, but Y” pivots, and “stopped being A and became B” transitions.

Post	Em dashes	”isn’t X. It’s Z"	"stopped being"	"not X, but Y”
When Bodies Think Like Neural Networks (2025-10)	24	0	0	0
What Movement Training Misses (2026-01)	14	0	0	0
Perception Gap (2026-03)	8	2	0	1
Winter Study Group Recap (2026-03)	70	1	0	1
Sensitivity as a Trainable Skill (2026-04)	27	1	0	0
The Triple Irony (2026-05)	46	0	1	1
What’s Between the Notes (2026-06)	0	8	4	4

The article is fully compliant on the mechanical rules that get checked (zero em dashes, contractions used, no “lands on”). The drift Patrick feels is concentrated in one place: roughly sixteen contrastive constructions in ~2,000 words. Each one is defensible alone. At this density the contrast move becomes the article’s dominant rhetorical gesture, and that gesture is the AI default the guide warns about, not Patrick’s “state what it is” voice.

Instance inventory, sorted by treatment

Scaffolding — rewrite candidates. These use negation as a setup device where the positive statement stands alone:

“That’s not a story about a mistake. It’s a story about what a constraint can do.” (The discovery inside the constraint)
“That’s not a consolation. It’s the capacity the practice is really building.” (What the practice is for)
“And the difference between them isn’t a talent you either have or you don’t. It’s skilled perception.” (The same chord, heard five ways)
“You’re not performing it anymore. You’re inside it.” (On stage)
“They aren’t adding more. They’re hearing more in the same space.” (Between the notes)
“The limitation isn’t in the way of the practice. A lot of the time, the limitation is the practice.” (Angine section; the second half is strong, the first half is the scaffold)
“it was never about building something that would impress people” (section opener defining the motive by what it wasn’t)
“There’s no agenda to push anyone into discomfort.” (defensive framing; the guide’s “It’s not a sales pitch” case)

Load-bearing negation — keep. These carry real information and are the argument itself:

“The song isn’t in the silence. You aren’t cycling when you’re off the bike.” (the intermittency point, Asia’s comment 8)
“The limit didn’t go away. My relationship to it changed.” (narrative fact)

The “stopped being X and became Y” tic. Four instances: trees/scenery (bike), board/riding + space/obstacle course (board), notes/sounds (chord). It’s a good move once. Four times in one essay reads as a template.

Direct evidence of context drift across rounds

Claude reply (3) in the file states the “isn’t a story about… it’s a story about” construction was fixed. The paragraph directly above that reply still opens with “That’s not a story about a mistake. It’s a story about what a constraint can do.” Reply (1) likewise reports dropping the “isn’t X, it’s Y” construction; eight instances remain in the body. The fixes were either applied to different sentences or regressed in later rounds. This confirms the suspicion that prompted the audit: rules acknowledged in one round stop being enforced a few rounds later.

Secondary findings in the article

Verbatim repetition across adjacent paragraphs. “It’s a place, with a pull toward somewhere else” (end of the Collier film paragraph) recurs as “become directions, a place with a pull toward somewhere else” two sentences later. The repetition-scan practice exists as session feedback but is not in any guide.
Meta-narration. “Here is where it gets interesting, and where the bike comes back.” / “That’s the whole thing, sitting right there on the surface.” / “There’s a nuance here I don’t want to skip.” Patrick’s guide: respects the reader’s intelligence, does not over-explain. These lines narrate the essay instead of advancing it.
Unapproved metaphors, borderline. “the pain stopped announcing itself,” “the mishearing was the gift,” “that impression is usually the gap talking,” “assemble a practice against the whole texture of their own life.” None are on the Allowed Metaphors list. Some may be deliberate Patrick choices from the drafting sessions; each needs his call rather than silent removal.
Residual intensifiers. “could really hurt you,” “the capacity the practice is really building,” plus several “actually”s. The guide (v2.7 draft) flags exactly this.
Clarity. “a good ride to time ratio” doesn’t parse cleanly on first read.

What this means

The article’s problem is not missing rules. Every pattern above is already named in VOICE-GUIDE-UNIFIED, most with examples. The problem is that enforcement is attention-based: each round focused on the sentences under discussion, and the patterns re-entered everywhere else. Fixing the article is a single targeted pass over the inventory above, with Patrick deciding keep/rewrite per instance.

Part 2 — The voice guide system

Finding 1: Rules without enforcement

The guide’s #1 named failure mode survived 14 review rounds on a single article. The three filters in WRITING-PROCESS are mandatory reading, but nothing checks output mechanically. A large subset of the rules is greppable: em dashes, “land on,” “reaching out,” “X isn’t Y. It’s Z,” “not X, but Y,” ”, one that,” intensifier phrases, “walk away with,” banned wellness vocabulary, verbatim phrase repetition.

Recommendation: a scripts/voice-lint.py that scans a draft’s body prose and reports hits with line numbers, plus a required final “voice pass” round added to WRITING-PROCESS (run the lint, then one human read purely for voice, separate from content review). The lint can’t judge tone, but it would have caught the contrast density in this article at round 1.

Finding 2: Published posts violate the em-dash ban

The ban became absolute on 2026-04-19 (v2.6, confirmed). Two posts published after that date carry em dashes in their body prose: Sensitivity as a Trainable Skill (2026-04-22, 27 instances) and The Triple Irony (2026-05-19, 46 instances). The live site versions presumably match. Decision needed: clean up the live posts, or grandfather anything published before today and apply the ban strictly going forward.

Finding 3: Confirmation backlog and stale metadata

VOICE-GUIDE-UNIFIED’s frontmatter says version: 1.1-draft, updated: 2026-03-05. The version history runs to 2.7 (2026-04-22). The header has been stale for four months.
Pending confirmations: v2.0–2.5 and 2.7 are all marked draft (mostly pending Ksenia; 2.5 pending Patrick since 2026-04-05; 2.7 pending Patrick since 2026-04-22). The “Replacing fitness” and “modalities” sections say “draft for Ksenia’s review” since March.
VOICE-GUIDE-GOVERNANCE has two Pending Review blocks from 2026-03-31 with no recorded response.
Practical effect: nobody can tell which rules are actually in force. The em-dash rule shows drafts get treated as binding anyway, so the draft/confirmed distinction currently signals nothing.

Recommendation: one joint session (or async inbox round) to bulk-confirm or reject everything pending, fix the frontmatter, and adopt a rule that the frontmatter version must match the history table.

Finding 4: Patrick’s personal guide has not kept up with Patrick

VOICE-GUIDE-PATRICK was last updated 2026-03-16 and is written almost entirely for emails, participant communications, and session summaries. Blog essays are now the highest-stakes writing in his name, and the guide has no essay/long-form section. Meanwhile roughly twenty Patrick-specific feedback rulings from March–June live only in Claude Code’s local memory on this machine. That memory is not shared: Asia’s machines and the VPS instance never see it, and a fresh session here only sees the one-line index. Rules that live in memory instead of the guides are exactly what produces “context drift when it comes to remembering what to dial in.”

Candidates to codify (source: accumulated session feedback):

For VOICE-GUIDE-UNIFIED:

Repetition scan before delivery (phrases, framings, and words repeated across a piece)
Avoid the generic AI register overall; metaphors only when earned, sparingly
Group/participant copy: neutral register, no praise adjectives (“strong session”)
Neutral professional tone for difficult points (no “worth flagging,” “I bring this up because”)
Vary framing of recurring facts across pieces (e.g., the development-timeframe phrasing)
“Instructors/facilitators,” never “teachers”

For VOICE-GUIDE-PATRICK:

A “First-person essay” section: what carries from the copy rules (metaphor discipline, no em dashes, intensifier restraint, state-what-it-is) and what relaxes (narrative negation that carries information is allowed; “I” voice; texture from concrete specifics). Include the contrast-density lesson from this audit: one contrastive move per section at most, and never as a paragraph opener twice in a row.
The tone target in Patrick’s own words: sophisticated and simple at the same time; anyone can understand it, but it is never oversimplified and never sensational.
No coaching language (guide/instruct/recommend, never coach)
Don’t lead with anatomical terms the other person didn’t use
Don’t open with “X isn’t Y” (lead positive)

Finding 5: Working well, for the record

The Ksenia guide is confirmed at v1.0 and current (updated 2026-03-27).
The governance loop itself functions when used: the 2026-03-02 seven-change review is a model of the process working.
The em-dash and contractions rules have fully taken in new drafting (the new article: zero em dashes).
Editorial-notes-in-file (the article’s 14-round log) gives exactly the audit trail that made Part 1 possible. Keep doing this.

Part 3 — Proposed actions (for Patrick’s decision)

Article: one targeted revision pass over the Part 1 inventory. Patrick rules keep/rewrite per instance; nothing else in the article is touched.
Guides: codify the Finding 4 items; add the essay section to Patrick’s guide; fix the stale frontmatter.
Backlog: joint confirm/reject round with Asia on everything pending (inbox item).
Tooling: build scripts/voice-lint.py and add the final voice pass to WRITING-PROCESS.
Live site: decide grandfather vs. cleanup for em dashes in the two post-ban published posts.