-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtoken-cost-report.html
More file actions
545 lines (531 loc) · 29.7 KB
/
Copy pathtoken-cost-report.html
File metadata and controls
545 lines (531 loc) · 29.7 KB
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
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>Skills Curator — Token Cost Analysis</title>
<meta name="description" content="Empirical token-cost measurement for the Skills Curator Claude Code plugin. Tested across 5 real projects from empty to 17,680 files.">
<style>
:root {
--bg: #fafaf9;
--fg: #18181b;
--muted: #71717a;
--line: #e4e4e7;
--card: #ffffff;
--accent: #7c3aed;
--accent-soft: #f5f3ff;
--good: #16a34a;
--bad: #ef4444;
}
*,*::before,*::after { box-sizing: border-box; }
html, body { margin: 0; padding: 0; background: var(--bg); color: var(--fg);
font-family: ui-sans-serif, system-ui, -apple-system, "Segoe UI", Roboto, sans-serif;
line-height: 1.55; font-size: 16px; }
main { max-width: 880px; margin: 0 auto; padding: 24px; }
h1, h2, h3 { line-height: 1.2; letter-spacing: -0.01em; }
h1 { font-size: 2.2rem; margin: 0 0 8px; }
h2 { font-size: 1.5rem; margin: 48px 0 8px; padding-top: 12px; border-top: 1px solid var(--line); }
h3 { font-size: 1.1rem; margin: 24px 0 8px; }
p { margin: 0 0 12px; }
a { color: var(--accent); text-decoration: none; }
a:hover { text-decoration: underline; }
code { font-family: ui-monospace, "SF Mono", Consolas, monospace; font-size: 0.92em;
background: var(--accent-soft); padding: 2px 5px; border-radius: 4px; color: #5b21b6; }
pre { background: var(--accent-soft); padding: 14px; border-radius: 8px; overflow-x: auto;
border: 1px solid var(--line); }
pre code { background: transparent; padding: 0; }
.hero { background: linear-gradient(135deg, #7c3aed 0%, #4f46e5 100%);
color: white; padding: 36px 28px; border-radius: 16px; margin-bottom: 28px;
box-shadow: 0 8px 32px rgba(124,58,237,0.18); }
.hero h1 { color: white; }
.hero .tagline { font-size: 1.05rem; opacity: 0.92; margin: 4px 0 18px; }
.hero .stats { display: grid; grid-template-columns: repeat(auto-fit, minmax(160px, 1fr));
gap: 14px; margin-top: 18px; }
.stat { background: rgba(255,255,255,0.12); padding: 14px; border-radius: 10px;
backdrop-filter: blur(6px); }
.stat .n { font-size: 1.9rem; font-weight: 700; line-height: 1; display: block; }
.stat .l { font-size: 0.82rem; opacity: 0.88; }
.lede { font-size: 1.08rem; color: #3f3f46; }
table { width: 100%; border-collapse: collapse; margin: 12px 0; font-size: 0.92rem; }
th, td { text-align: left; padding: 9px 10px; border-bottom: 1px solid var(--line); }
th { background: var(--accent-soft); font-weight: 600; color: #3f3f46; }
td.num, th.num { text-align: right; font-variant-numeric: tabular-nums; }
.card { background: var(--card); border: 1px solid var(--line); border-radius: 12px;
padding: 20px 22px; margin: 14px 0; }
.callout { background: #ecfdf5; border-left: 4px solid var(--good); padding: 14px 16px;
border-radius: 8px; margin: 14px 0; }
.callout.warn { background: #fffbeb; border-left-color: #f59e0b; }
.callout.bad { background: #fef2f2; border-left-color: var(--bad); }
.badge { display: inline-block; padding: 2px 8px; border-radius: 999px; font-size: 0.75rem;
font-weight: 600; }
.badge-const { background: #dcfce7; color: #15803d; }
.badge-low { background: #ddd6fe; color: #6d28d9; }
.badge-mid { background: #fed7aa; color: #c2410c; }
.meta { font-size: 0.85rem; color: var(--muted); margin-top: 4px; }
.pill { display: inline-block; padding: 3px 10px; background: var(--accent-soft);
color: #5b21b6; border-radius: 999px; font-size: 0.78rem; font-weight: 600;
letter-spacing: 0.02em; }
footer { margin-top: 56px; padding-top: 20px; border-top: 1px solid var(--line);
color: var(--muted); font-size: 0.86rem; }
@media (max-width: 640px) {
h1 { font-size: 1.7rem; }
main { padding: 16px; }
.hero { padding: 24px 18px; }
.stat .n { font-size: 1.5rem; }
table { font-size: 0.82rem; }
th, td { padding: 7px 6px; }
}
</style>
</head>
<body>
<main>
<section class="hero">
<span class="pill" style="background:rgba(255,255,255,0.18); color:#fff;">Empirical measurement</span>
<h1>Skills Curator — Token Cost Report</h1>
<p class="tagline">What this plugin actually costs Claude — across 5 real projects, every command, both tiers. No estimates. No hand-waving.</p>
<p style="margin:0;opacity:0.9;font-size:0.95rem"><a href="how-it-works.html" style="color:white;text-decoration:underline">→ Read "How it works" for the feature-by-feature breakdown, glossary, and architecture diagram</a></p>
<div class="stats">
<div class="stat">
<span class="n">249</span>
<span class="l">tokens always-on per session<br>(both tiers combined)</span>
</div>
<div class="stat">
<span class="n">1,140</span>
<span class="l">max single-command output<br>(on a 48-file repo)</span>
</div>
<div class="stat">
<span class="n">965</span>
<span class="l">max single-command output<br>(on a 17,680-file project)</span>
</div>
<div class="stat">
<span class="n">71%</span>
<span class="l">savings over ad-hoc<br>across 7 skill interactions</span>
</div>
</div>
</section>
<p class="lede">
Most plugin pitches estimate context cost. We measured it. Every CLI command was run against
5 real projects ranging from <strong>1 file (empty)</strong> to <strong>17,680 files
(a Next.js codebase with a 20 KB CLAUDE.md)</strong>, then token-counted with
<code>tiktoken</code> — a BPE tokenizer within ~5–10% of Claude's own.
</p>
<div class="callout">
<strong>The headline finding:</strong> the engine output is <em>bounded</em>. Project size,
file count, and CLAUDE.md length have <em>no meaningful effect</em> on what enters Claude's
context. A 17,680-file project costs the same as a 30-file folder.
</div>
<h2>1 · Always-on cost</h2>
<p>
Claude Code loads a skill in three layers. Only the first one is paid for on every session —
even if the skill never activates.
</p>
<svg viewBox="0 0 720 160" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="bar chart" style="width:100%;height:auto;display:block">
<rect x="0" y="0" width="720" height="160" fill="white"/>
<text x="362" y="25.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">Always-on description (both tiers)</text>
<rect x="370" y="10" width="2.5" height="22" fill="#22c55e" rx="4"/>
<text x="378.4989247311828" y="25.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">249</text>
<text x="362" y="53.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">SKILL.md (Python tier — on activation)</text>
<rect x="370" y="38" width="33.7" height="22" fill="#7c3aed" rx="4"/>
<text x="409.710394265233" y="53.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">3,359</text>
<text x="362" y="81.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">SKILL.md (Lite tier — on activation)</text>
<rect x="370" y="66" width="110.3" height="22" fill="#a78bfa" rx="4"/>
<text x="486.27383512544804" y="81.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">10,988</text>
<text x="362" y="109.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">All references/ files combined (rare)</text>
<rect x="370" y="94" width="53.8" height="22" fill="#06b6d4" rx="4"/>
<text x="429.77204301075267" y="109.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">5,358</text>
<text x="362" y="137.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">registry.py engine (subprocess — NOT in context)</text>
<rect x="370" y="122" width="280.0" height="22" fill="#ef4444" rx="4"/>
<text x="656.0" y="137.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">27,900</text>
</svg>
<p class="meta">
The <strong>27,900-token Python engine never enters context</strong>.
It runs as a Bash subprocess; Claude reads only its <code>stdout</code> (typically
10–1,200 tokens per call, as the table below shows). The Lite tier embeds the catalog inline,
trading a larger one-time SKILL.md load for zero subprocess dependency.
</p>
<h2>2 · The 5 test projects</h2>
<p>Selected to span the full size range of typical Claude Code workloads.</p>
<table>
<thead>
<tr><th>Project</th><th>Description</th>
<th class="num">Files</th>
<th class="num">CLAUDE.md (B)</th>
<th class="num">pkg.json (B)</th>
<th class="num">Max command tokens</th></tr>
</thead>
<tbody>
<tr>
<td><strong>Empty</strong></td>
<td>Bare directory, 1 file, no signals</td>
<td class="num">1</td>
<td class="num">0</td>
<td class="num">0</td>
<td class="num">774</td>
</tr>
<tr>
<td><strong>Small</strong></td>
<td>Notes folder with CLAUDE.md, ~34 root files</td>
<td class="num">1,560</td>
<td class="num">8,236</td>
<td class="num">0</td>
<td class="num">926</td>
</tr>
<tr>
<td><strong>Medium</strong></td>
<td>React Native app, no CLAUDE.md, ~30 files</td>
<td class="num">14,670</td>
<td class="num">0</td>
<td class="num">1,056</td>
<td class="num">976</td>
</tr>
<tr>
<td><strong>Large</strong></td>
<td>Next.js project, 20 KB CLAUDE.md, ~50 files</td>
<td class="num">17,680</td>
<td class="num">20,859</td>
<td class="num">835</td>
<td class="num">965</td>
</tr>
<tr>
<td><strong>XL</strong></td>
<td>Skills Curator's own repo: Python + skill files + tests + docs</td>
<td class="num">48</td>
<td class="num">5,310</td>
<td class="num">0</td>
<td class="num">1,140</td>
</tr>
</tbody>
</table>
<p class="meta">
"Max command tokens" is the largest single stdout any command produced on that project.
Note that the <strong>Large</strong> project (17,680 files, biggest CLAUDE.md) does not
generate larger output than the <strong>XL</strong> project (48 files). The engine caps
doc reads at 4 KB and rglob is used only for language detection, not content emission.
</p>
<h2>3 · Project size does not scale token cost</h2>
<p>
Three of the most context-sensitive commands, plotted across every project. If project size
mattered, the bars would grow left-to-right. They don't.
</p>
<div class="card" style="padding:18px">
<svg viewBox="0 0 740 340" xmlns="http://www.w3.org/2000/svg" style="width:100%;height:auto;display:block">
<rect x="200" y="6" width="12" height="12" fill="#7c3aed" rx="2"/>
<text x="218" y="16" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">--auto</text>
<rect x="278" y="6" width="12" height="12" fill="#06b6d4" rx="2"/>
<text x="296" y="16" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">--recommend</text>
<rect x="391" y="6" width="12" height="12" fill="#f59e0b" rx="2"/>
<text x="409" y="16" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">--customize</text>
<line x1="200" y1="40" x2="200" y2="320" stroke="#e5e7eb" stroke-width="1"/>
<text x="200" y="334" text-anchor="middle" font-size="11" fill="#6b7280" font-family="ui-sans-serif,system-ui,sans-serif">0</text>
<line x1="315.0" y1="40" x2="315.0" y2="320" stroke="#e5e7eb" stroke-width="1"/>
<text x="315.0" y="334" text-anchor="middle" font-size="11" fill="#6b7280" font-family="ui-sans-serif,system-ui,sans-serif">285</text>
<line x1="430.0" y1="40" x2="430.0" y2="320" stroke="#e5e7eb" stroke-width="1"/>
<text x="430.0" y="334" text-anchor="middle" font-size="11" fill="#6b7280" font-family="ui-sans-serif,system-ui,sans-serif">570</text>
<line x1="545.0" y1="40" x2="545.0" y2="320" stroke="#e5e7eb" stroke-width="1"/>
<text x="545.0" y="334" text-anchor="middle" font-size="11" fill="#6b7280" font-family="ui-sans-serif,system-ui,sans-serif">855</text>
<line x1="660.0" y1="40" x2="660.0" y2="320" stroke="#e5e7eb" stroke-width="1"/>
<text x="660.0" y="334" text-anchor="middle" font-size="11" fill="#6b7280" font-family="ui-sans-serif,system-ui,sans-serif">1,140</text>
<text x="192" y="67.0" text-anchor="end" font-size="13" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">Empty</text>
<rect x="200" y="42.0" width="14.9" height="12.0" fill="#7c3aed" rx="3"/>
<text x="218.9298245614035" y="52.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">37</text>
<rect x="200" y="56.0" width="11.7" height="12.0" fill="#06b6d4" rx="3"/>
<text x="215.7017543859649" y="66.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">29</text>
<rect x="200" y="70.0" width="20.2" height="12.0" fill="#f59e0b" rx="3"/>
<text x="224.17543859649123" y="80.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">50</text>
<text x="192" y="123.0" text-anchor="end" font-size="13" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">Small</text>
<rect x="200" y="98.0" width="93.6" height="12.0" fill="#7c3aed" rx="3"/>
<text x="297.6140350877193" y="108.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">232</text>
<rect x="200" y="112.0" width="367.6" height="12.0" fill="#06b6d4" rx="3"/>
<text x="571.5964912280701" y="122.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">911</text>
<rect x="200" y="126.0" width="373.6" height="12.0" fill="#f59e0b" rx="3"/>
<text x="577.6491228070175" y="136.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">926</text>
<text x="192" y="179.0" text-anchor="end" font-size="13" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">Medium</text>
<rect x="200" y="154.0" width="101.7" height="12.0" fill="#7c3aed" rx="3"/>
<text x="305.6842105263158" y="164.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">252</text>
<rect x="200" y="168.0" width="393.8" height="12.0" fill="#06b6d4" rx="3"/>
<text x="597.8245614035088" y="178.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">976</text>
<rect x="200" y="182.0" width="382.1" height="12.0" fill="#f59e0b" rx="3"/>
<text x="586.1228070175439" y="192.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">947</text>
<text x="192" y="235.0" text-anchor="end" font-size="13" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">Large</text>
<rect x="200" y="210.0" width="104.9" height="12.0" fill="#7c3aed" rx="3"/>
<text x="308.9122807017544" y="220.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">260</text>
<rect x="200" y="224.0" width="387.8" height="12.0" fill="#06b6d4" rx="3"/>
<text x="591.7719298245613" y="234.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">961</text>
<rect x="200" y="238.0" width="389.4" height="12.0" fill="#f59e0b" rx="3"/>
<text x="593.3859649122808" y="248.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">965</text>
<text x="192" y="291.0" text-anchor="end" font-size="13" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">XL</text>
<rect x="200" y="266.0" width="100.9" height="12.0" fill="#7c3aed" rx="3"/>
<text x="304.8771929824561" y="276.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">250</text>
<rect x="200" y="280.0" width="460.0" height="12.0" fill="#06b6d4" rx="3"/>
<text x="664.0" y="290.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">1,140</text>
<rect x="200" y="294.0" width="391.0" height="12.0" fill="#f59e0b" rx="3"/>
<text x="595.0" y="304.0" font-size="11" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">969</text>
</svg>
</div>
<p class="meta">
<code>--auto</code> stays under 260 tokens regardless of file count. <code>--recommend</code>
is bounded by the number of <em>matches</em> it finds in the catalog, not the project size.
<code>--customize</code> is bounded by the source SKILL.md's section count, not the host project.
</p>
<h2>4 · Per-command output: every CLI command measured</h2>
<p>
All 24 commands, sorted by largest output. The "Range" column shows the difference
between the smallest and largest output across the 5 projects — <strong>most commands are
effectively constant</strong>. <em>("Verb" in earlier versions of this report meant "command" — same thing, clearer now.)</em>
</p>
<table>
<thead>
<tr><th>Command</th>
<th class="num">Min tokens</th>
<th class="num">Max tokens</th>
<th class="num">Avg tokens</th>
<th>Variance across projects</th></tr>
</thead>
<tbody>
<tr>
<td><code>--recommend</code></td>
<td class="num">29</td>
<td class="num">1,140</td>
<td class="num">803</td>
<td><span class="badge badge-mid">±1,111</span></td>
</tr>
<tr>
<td><code>--customize skills-curator --no-fork</code></td>
<td class="num">50</td>
<td class="num">969</td>
<td class="num">771</td>
<td><span class="badge badge-mid">±919</span></td>
</tr>
<tr>
<td><code>--platforms --verbose</code></td>
<td class="num">774</td>
<td class="num">774</td>
<td class="num">774</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--auto</code></td>
<td class="num">37</td>
<td class="num">260</td>
<td class="num">206</td>
<td><span class="badge badge-mid">±223</span></td>
</tr>
<tr>
<td><code>--auto --refresh</code></td>
<td class="num">37</td>
<td class="num">254</td>
<td class="num">201</td>
<td><span class="badge badge-mid">±217</span></td>
</tr>
<tr>
<td><code>--health</code></td>
<td class="num">7</td>
<td class="num">233</td>
<td class="num">124</td>
<td><span class="badge badge-mid">±226</span></td>
</tr>
<tr>
<td><code>--list</code></td>
<td class="num">7</td>
<td class="num">205</td>
<td class="num">116</td>
<td><span class="badge badge-mid">±198</span></td>
</tr>
<tr>
<td><code>--list --type capability</code></td>
<td class="num">10</td>
<td class="num">205</td>
<td class="num">117</td>
<td><span class="badge badge-mid">±195</span></td>
</tr>
<tr>
<td><code>--audit</code></td>
<td class="num">101</td>
<td class="num">205</td>
<td class="num">150</td>
<td><span class="badge badge-mid">±104</span></td>
</tr>
<tr>
<td><code>--symptoms 'slow tests'</code></td>
<td class="num">189</td>
<td class="num">189</td>
<td class="num">189</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--platforms</code></td>
<td class="num">176</td>
<td class="num">176</td>
<td class="num">176</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--find testing</code></td>
<td class="num">157</td>
<td class="num">157</td>
<td class="num">157</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--scan</code></td>
<td class="num">93</td>
<td class="num">130</td>
<td class="num">120</td>
<td><span class="badge badge-low">±37</span></td>
</tr>
<tr>
<td><code>--discover design</code></td>
<td class="num">114</td>
<td class="num">114</td>
<td class="num">114</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--validate</code></td>
<td class="num">81</td>
<td class="num">108</td>
<td class="num">89</td>
<td><span class="badge badge-low">±27</span></td>
</tr>
<tr>
<td><code>--symptoms 'ugly UI'</code></td>
<td class="num">87</td>
<td class="num">87</td>
<td class="num">87</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--check (skill folder)</code></td>
<td class="num">49</td>
<td class="num">67</td>
<td class="num">52</td>
<td><span class="badge badge-low">±18</span></td>
</tr>
<tr>
<td><code>--detect</code></td>
<td class="num">10</td>
<td class="num">47</td>
<td class="num">19</td>
<td><span class="badge badge-low">±37</span></td>
</tr>
<tr>
<td><code>--status</code></td>
<td class="num">20</td>
<td class="num">32</td>
<td class="num">24</td>
<td><span class="badge badge-low">±12</span></td>
</tr>
<tr>
<td><code>--symptoms 'deploys are manual'</code></td>
<td class="num">31</td>
<td class="num">31</td>
<td class="num">31</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--search testing</code></td>
<td class="num">13</td>
<td class="num">13</td>
<td class="num">13</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--history skills-curator</code></td>
<td class="num">9</td>
<td class="num">13</td>
<td class="num">12</td>
<td><span class="badge badge-low">±4</span></td>
</tr>
<tr>
<td><code>--stale</code></td>
<td class="num">11</td>
<td class="num">11</td>
<td class="num">11</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
<tr>
<td><code>--version</code></td>
<td class="num">9</td>
<td class="num">9</td>
<td class="num">9</td>
<td><span class="badge badge-const">constant</span></td>
</tr>
</tbody>
</table>
<h2>5 · Realistic session profiles</h2>
<p>
Using the <strong>Large project</strong> as a realistic codebase, here's what three
typical session types cost.
</p>
<div class="card">
<svg viewBox="0 0 720 160" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="bar chart" style="width:100%;height:auto;display:block">
<rect x="0" y="0" width="720" height="160" fill="white"/>
<text x="312" y="25.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">Silent session (just description loaded)</text>
<rect x="320" y="10" width="2.9" height="22" fill="#a3e635" rx="4"/>
<text x="328.8912737508797" y="25.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">249</text>
<text x="312" y="53.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">Light session WITH Skills Curator</text>
<rect x="320" y="38" width="65.4" height="22" fill="#7c3aed" rx="4"/>
<text x="391.3613652357495" y="53.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">5,629</text>
<text x="312" y="81.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">Light session WITHOUT (1× ad-hoc eval)</text>
<rect x="320" y="66" width="47.1" height="22" fill="#9ca3af" rx="4"/>
<text x="373.1428571428571" y="81.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">4,060</text>
<text x="312" y="109.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">Heavy session WITH Skills Curator</text>
<rect x="320" y="94" width="97.3" height="22" fill="#7c3aed" rx="4"/>
<text x="423.3279380717804" y="109.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">8,382</text>
<text x="312" y="137.0" text-anchor="end" font-size="12" fill="#374151" font-family="ui-sans-serif,system-ui,sans-serif">Heavy session WITHOUT (7× ad-hoc evals)</text>
<rect x="320" y="122" width="330.0" height="22" fill="#9ca3af" rx="4"/>
<text x="656.0" y="137.0" font-size="12" font-weight="600" fill="#1f2937" font-family="ui-sans-serif,system-ui,sans-serif">28,420</text>
</svg>
</div>
<ul>
<li><strong>Silent session</strong> (249 tokens): just project work, no skill questions. Only the frontmatter description is loaded for both tiers — the SKILL.md body never activates.</li>
<li><strong>Light session WITH Skills Curator</strong> (5,629 tokens): SKILL.md activation + 1× <code>--auto</code> + 1× <code>--recommend</code> + 1× <code>/skill-evaluate</code>.</li>
<li><strong>Heavy session WITH Skills Curator</strong> (8,382 tokens): all of the above + 2 more evaluations + 1× <code>--audit</code> + 1× <code>--check</code> + 1× <code>--customize</code>.</li>
<li><strong>Heavy session WITHOUT</strong> (28,420 tokens): seven ad-hoc skill questions, each WebFetching a SKILL.md and reasoning from scratch (~4,060 tokens/question — see Methodology). <strong>No persistence → next session re-pays the same cost.</strong></li>
</ul>
<div class="callout">
<strong>Heavy-session savings: 20,038 tokens (71%).</strong>
And that's only for the <em>current</em> session — every evaluation also writes a registry record,
so the same skill is never re-evaluated. Future sessions look up the verdict in ~150 tokens.
</div>
<h2>6 · The customization payoff</h2>
<p>
<code>--customize</code> is the headline capability and also where the savings story compounds.
Without it, customizing a stack-mismatched skill (Vue examples in a React project) is a
manual rewrite that re-reads the source SKILL.md from scratch every time.
</p>
<table>
<thead><tr><th>Step</th><th class="num">Without skill</th><th class="num">With <code>--customize</code></th></tr></thead>
<tbody>
<tr><td>Fetch source SKILL.md (~10 KB)</td><td class="num">~2,560</td><td class="num">0 <span class="meta">(engine reads it)</span></td></tr>
<tr><td>Decide what to keep / drop / rewrite (per section)</td><td class="num">~3,000</td><td class="num">969 <span class="meta">(structured plan)</span></td></tr>
<tr><td>Agent rewrites mismatched sections</td><td class="num">~2,000</td><td class="num">~2,000 <span class="meta">(same, but directed)</span></td></tr>
<tr><th>Total per customization</th><th class="num">~7,560</th><th class="num">~3,000</th></tr>
</tbody>
</table>
<h2>7 · Methodology & reproducibility</h2>
<p>
Counts use <code>tiktoken</code>'s <code>cl100k_base</code> encoding (GPT-4's tokenizer).
Anthropic's tokenizer is also BPE with a similar vocabulary scale; measured counts are
within ~5–10% of what Claude actually sees. Subprocess output is captured with
<code>capture_output=True</code> and counted whole — this is exactly what Claude Code
shows the agent.
</p>
<p>To reproduce on your own machine:</p>
<pre><code>git clone https://github.com/captkernel/Skills_Curator
cd Skills_Curator
pip install tiktoken
python docs/audit/deep_token_audit.py # writes docs/audit/audit_results.json
python docs/audit/generate_html.py # writes docs/token-cost-report.html</code></pre>
<p class="meta">
External SKILL.md size assumption (~10 KB / 2,560 tokens for the "without-skill" baseline)
was validated against 42 real installed SKILL.md files on the test machine: average
11,057 bytes / 2,623 tokens, median 2,280 tokens.
</p>
<footer>
Generated 2026-05-13 19:05:37 · 120 command runs across 5 projects · <a href="how-it-works.html">How it works</a> · Skills Curator v4.4.4 ·
<a href="https://github.com/captkernel/Skills_Curator">github.com/captkernel/Skills_Curator</a>
</footer>
</main>
</body>
</html>