1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
|
#+TITLE: Voice Profile — canonical source-of-truth for the voice skill
#+DATE: 2026-05-29
#+SOURCE: rulesets session 2026-05-29
* How this combines with SKILL.md (pairing rule)
This file is the canonical source-of-truth for the voice skill's rationale, evidence, examples, and history. =voice/SKILL.md= holds the thin rule-set: one-line directives per pattern, mode applicability, and a pointer back here. Everything else (Problem, Basis, Before/After, Detection guidance, History) lives in the per-pattern sections below.
Pairing rule. Every change to =voice/SKILL.md= MUST land alongside the corresponding update in this file. The two are normatively paired. A SKILL.md edit without a profile update is incomplete. A profile update without a SKILL.md edit is fine (rationale or evidence can deepen without changing the rule).
Pattern numbering in both files matches: SKILL.md's =### N. <Name>= maps to this file's =* §N <Name>= section. Mode tags use the same vocabulary: =general=, =prose=, =personal=.
When the agent runs =/voice=, it reads SKILL.md for the rules and consults this file for the examples and basis it needs to apply each pattern correctly.
* Corpus
Git commit bodies authored by Craig Jennings across all repos under =~/code/= and =~/projects/=. After cleanup (subject lines, trailers, URL-only lines, AI-attribution lines, blank-run collapse):
- 5355 raw commits, 1895 with non-trivial bodies
- 128608 words, 912400 characters
- 33 repos contributing; top sources: archsetup (703), rulesets (621), work (565), archangel (455), home (395)
PRs deferred. Email + Slack deferred. This is one register (deliberate technical prose) — useful but narrow.
* Findings against the 41 SKILL.md patterns
** Strongly confirmed by the corpus
*Pattern 17 (no emojis).* Zero emojis in corpus. Confirmed.
*Pattern 7 (AI vocabulary).* "delve" 0. "embark" 0. "navigate the" 0. "in the realm of" 0. "seamless" 0. "moreover" 0. "furthermore" 0. "in conclusion" 0. "additionally" 1. "robust" 1. "leverage" 1. Rule confirmed for 11 of 12 watch-words. (One exception below.)
*Pattern 22 (filler).* "moreover" / "furthermore" / "additionally" / "in conclusion": all zero or one occurrence. Filler-phrase avoidance confirmed.
*Pattern 32 (first-person rewrite).* Standalone "I" at 3.85 per 1000 words. Craig writes first-person heavily — this is real, not aspirational.
*Pattern 34 (contractions).* 459 contractions total (3.57 per 1000). Top hits: =doesn't= (92), =don't= (59), =isn't= (46), =it's= (43), =can't= (40), =that's= (34). Rule confirmed.
*Pattern 38 (terse cut).* 41.1% of paragraphs are single-sentence. Craig writes terse — paragraph breaks land after one complete thought even when short. Confirmed indirectly via paragraph structure.
** Aspirational (corpus contradicts, but the rule is intentional self-discipline)
*Pattern 13 (em-dash zero-tolerance, personal mode).* Corpus rate: 3.49 em-dashes per 1000 words. Comparable to AI-generated prose. Craig USES em-dashes regularly in commit bodies — the rule overrides his habit, it doesn't reflect it. Suggested rewording: drop the "LLMs use em dashes more than humans" framing; keep the zero-tolerance directive but rationale becomes "Craig's published voice — commit messages going forward, PR bodies, emails — drops em-dashes by choice because it reads cleaner and avoids a common AI tell, regardless of his pre-rule habit." Honest about the source.
*Pattern 33 (semicolons → period/comma).* Corpus rate: 3.16 semicolons per 1000. Craig uses semicolons regularly. Same shape as #13: rule is self-discipline, not habit-reflection. Suggested rewording: acknowledge the rule overrides habit rather than implying it codifies one.
These two rules are still valuable — em-dashes and semicolons both read cleaner when absent from short imperative-leaning prose. But the SKILL.md should say "this is a rule I've decided to follow," not "this is how I already write."
** Worth challenging
*Pattern 7 watch-word "comprehensive".* 42 occurrences in corpus (~0.33 per 1000). All other AI-tell watch-words clock near zero. "comprehensive" appears to be genuine vocabulary for Craig in technical contexts ("comprehensive test coverage", "comprehensive audit"). Suggested change: pull "comprehensive" out of the watch-list, or carve out a "watch in clusters, not solo" note — flag only when "comprehensive" co-occurs with other AI-tell words.
** Worth adding (corpus surfaces traits the rules don't capture)
*Single-sentence paragraph cadence.* 41.1% of paragraphs are exactly one sentence. This is distinctive — most prose-style guides advise multi-sentence paragraphs. Suggested addition (prose + personal): a positive pattern noting "a one-sentence paragraph is a finished thought, not a fragment. Break paragraphs after one complete thought when the next thought shifts angle, even if both are short." Anti-rule against "merge short paragraphs into multi-sentence ones."
*Parenthetical density.* 23.07 opening parens per 1000 words. Heavy parenthetical use — asides, clarifications, scope-narrowing in parens. Currently no rule addresses this either way. Could add a positive pattern: "parentheses for asides are part of the voice. Don't strip them in a 'clean prose' pass."
*Question-mark rarity.* 0.33 per 1000. Craig's prose is declarative — he states things, rarely asks them. Worth noting as a register marker (when /voice personal output has questions, double-check whether they're contextual or AI rhetoric).
** Out of corpus (commits don't test these — Phase 2 needed)
- *Pattern 13 in long-form prose.* Commit bodies are short. Email and PR bodies may show different em-dash rates.
- *Pattern 14 (boldface).* Org-mode bold uses =*word*=, not detectable by simple grep. Markdown bold rare in commits.
- *Pattern 16 (title case in headings).* Commits don't carry headings.
- *Pattern 19 (collaborative artifacts).* Not present in commit bodies.
- *Pattern 35 (sentence split on conjunctions).* Average sentence is 18.81 words, median 14, with 28% of sentences 21+ words — long-sentence rate is moderate. Need to inspect actual sentences to know if they're conjunction-stitched. Defer.
- *Pattern 36 (felt-experience cut).* Commit bodies wouldn't carry felt-experience prose. Email + journal corpus needed.
- *Pattern 37 (sentence fragments).* 9.7% of sentences are 1-5 words. Some are legitimate ("All eight pass."), some may be fragments. Can't tell from word-count alone. Defer to a pass that does syntactic detection.
- *Pattern 39 (public-artifact scope).* The corpus IS the public artifacts — circular. Defer.
- *Pattern 40 (praise vs correction asymmetry).* Not detectable in commit bodies. Email or PR-review corpus needed.
** Curiosities
- *=I'm=* (9 occurrences) and *=I'll=* (2 occurrences) are surprisingly rare relative to standalone =I= (495 occurrences). Either Craig prefers "I am" / "I will" in commit prose, or the regex missed contexts. Worth a quick sample check in Phase 2.
* Suggested deltas
Six concrete edits to =voice/SKILL.md=, all of which can land independently:
1. *#13 (Em-Dash).* Drop the "LLMs use em dashes more than humans" framing in the personal-mode section. Restate the zero-tolerance rule as self-discipline ("Craig's published voice drops em-dashes by choice"), not habit-reflection. Cite: corpus rate 3.49/1000, AI-comparable.
2. *#33 (Semicolons).* Same shape. Restate as self-discipline. Cite: corpus rate 3.16/1000.
3. *#7 (AI Vocabulary).* Remove "comprehensive" from the watch-list, OR add a note that "comprehensive" alone is acceptable; flag only when it co-occurs with =delve= / =leverage= / =robust= / =seamless= / =moreover= etc. Cite: 42 occurrences, all other watch-words at 0 or 1.
4. *NEW pattern (prose + personal): "Single-sentence paragraph cadence is a feature."* 41.1% of corpus paragraphs are exactly one sentence. A one-sentence paragraph is a finished thought, not a fragment. The voice pass should not merge short paragraphs into multi-sentence ones.
5. *NEW pattern (prose + personal): "Parentheses for asides are part of the voice."* 23 opening parens per 1000 words. Heavy parenthetical use is distinctive. Don't strip parenthetical asides in a "clean prose" pass.
6. *Register marker (advisory, not a rewrite rule): "Declarative is the default."* 0.33 question marks per 1000. Voice personal output that contains rhetorical questions should be checked — they're often AI rhetoric, not Craig's register.
* What Phase 2 would add
- Email corpus (gmail + cmail, sent-only, long-form) — different register, especially long-form prose flow.
- PR bodies and review comments — longer prose, deliberate register, includes the praise/correction asymmetry test ground.
- Slack messages — casual register, contraction rate, sentence-fragment rate.
- Syntactic detection — distinguish fragments from terse complete sentences for pattern #37.
- Long-form documents (résumé, proposals if any) — single register but high prose density.
* Per-pattern entries
The 41 patterns are migrated below from SKILL.md per the pairing rule above. Phase 1 of this migration populates only Pattern §13 as a worked example of the form. The remaining 40 patterns are scheduled for the backfill pass once Craig confirms the shape.
** §13 Em Dash Overuse
*** Modes
General mode: overuse-reduction.
Prose + personal modes: zero-tolerance.
*** Rule
Replace em-dashes (=—=) with a comma, period, colon, or parentheses, whichever fits. Zero-tolerance in prose and personal modes holds *everywhere in the text*, including inside example blocks, code-fence prose, and quoted material. An em-dash in a quoted line still gets replaced.
*** Problem
LLMs use em dashes more than the median human writer, mimicking "punchy" sales writing.
*** Basis
Corpus measurement (2026-05-29 git commits, 128k words): 3.49 em-dashes per 1000 words. Comparable to AI-generated prose. The zero-tolerance rule in prose and personal modes is self-discipline, not habit-reflection. Craig has decided his published voice drops em-dashes by choice because the result reads cleaner and avoids the most common AI tell, regardless of his pre-rule frequency.
*** Before
#+begin_example
The term is primarily promoted by Dutch institutions—not by the people themselves. You don't say "Netherlands, Europe" as an address—yet this mislabeling continues—even in official documents.
#+end_example
*** After
#+begin_example
The term is primarily promoted by Dutch institutions, not by the people themselves. You don't say "Netherlands, Europe" as an address, yet this mislabeling continues in official documents.
#+end_example
*** History
- Original SKILL.md entry: rule scoped to general overuse-reduction.
- 2026-05-26 (commit =4fac2a0=): prose mode added, rule strengthened to zero-tolerance in prose and personal.
- 2026-05-29 (commit =c3cf9a5=): Note on basis added with corpus measurement.
- 2026-05-29: migrated to this file as the canonical home per the pairing rule.
|
