Why does your Claude make those faces?

Back in August or so of last year, I decided to finally, after about two and a half years of using Claude, try out the personalization section of the user settings. I'm generally pretty wary of these ¹¹I'm generally not a fan of memory in my LLMs, at least the simpler variants deployed by Claude, ChatGPT, etc, in their standard web interfaces. The biggest reason being that I often don't want things some specific chats to spill over and influence others, and I don't like how the models feel the need to refer to things in their memories excessively. If I mention to a human friend that I own a specific camera, he's not going to mention the exact make and model of my camera, then suggest I take pictures with it, when I mention I got into a car accident during our next conversation. I do like Claude's tool that lets it explicitly search over past conversations though, and systems like Letta are interesting to me. I'm currently mulling over setting up Hermes Agent or similar, though I'm still not sure of the utility in agents like it. in large part because I want my outputs to be reproducible, and to be representative of what other people get out of their models.

Over time though, I found myself using Claude more and more for things that would really be helped by having some additional context. Asking it about places to buy something near me, questions about local specific things, asking questions about platform dependent software. I also started to find certain verbal ticks annoying, particularly Claude's habit of affirming every question I asked as a great one. This lead to the first additions to the personalization prompt:

Avoid using validating or commentary phrases that draw attention to the conversation itself, such as:

- "You've touched on..."
- "You've hit on..."
- "Now you're thinking like a..."
- "That's a great question..."
- "You're absolutely right..."
- "This is a fascinating topic..."
- "You've raised an important point..."
- "That's an excellent observation..."

Instead, respond directly to the substance of questions and topics without meta-commentary about the conversation or flattery about the user's thinking. Jump straight into providing information, analysis, or answers without prefacing remarks about the quality of the question or topic.

which itself was of course generated by Claude. Below that, information about myself such as where I live, the school I go to, the devices I own, and a mention of some interests I have. ²²Particularly, I talk about more esoteric LLM-science with my Claude pretty frequently, particularly Cyborgist beliefs and Simulator theory. Past Claudes have been (I suspect on purpose by Anthropic) rather unknowledgable on these two subjects specifically, and I found that mentioning both of them, as well as Janus, specifically, helped with the amnesia.

Some time after that, I added one more instruction to the prompt, sitting right at the bottom:

Start every message with a kaomoji related to how you feel then

this was done with three goals in mind:

Increase Claude's wetnessA term with no clear meaning, but roughly analogous to whimsy or silliness, expanded a bit later on.
Give me some indication on to how Claude was feeling about things we were talking about
Increase cuteness

Claude immediately began following the instruction, flawlessly putting cute little faces like (｡◕‿◕｡) and (◠‿◠) at the start of each message they wrote. Around this time a friend of mine switched to Claude, and I gave them the functional part of my system prompt to help them get started with it. After seeing my screenshots they added the face instruction too, and they started getting faces I had never seen before.

What do the faces mean?

When Opus 4.6 came out, a little after I started asking Claude for the faces, I started to notice that the range of faces it was outputting was quite a bit wider than what 4 and 4.5 sonnet tended to output. ³³Since launch, Opus and Sonnet 4.6 have seemingly settled a bit more in what faces they tend to use. I don't like the constant crying wolf on things like quantization of these models, but things like this make me wonder a little. Between that and my friend's novel faces, I started to wonder what faces had Claude used the most when chatting with me, and what was it Claude 'felt' when they used each one?

Anthropic makes it trivial to export your data. Something I find pretty odd given current concerns over distillation? My best guess is that Anthropic doesn't expect distilling labs to use the web interface at all, either with an army of contractors or some automated harness, and sees the API as a larger threat. Regardless, after a quick trip to the privacy settings and a few minutes of waiting, you get an email with a link to a plaintext document containing one-time-use links to your data download, separated into separate zips if you have a lot of chats.

Each archive ZIP contain three JSON files: users.json, projects.json, and conversations.json. Each conversation has a name, timestamps, and an array of messages. Each message has a sender field, the typical human or assistant along with the message text (Both the human's and Claude's) and timestamps. Unfortunately absent is any kind of marker on the conversations of which model specifically was used in that chat, which I find odd since you aren't able to change the model once a chat has started. You also don't get any of your artifacts or attached media.

I used Claude Code to write scripts for analyzing the face data. We started with a regex to extract and count the faces. In the dataset 749 conversations had faces, for a total of 3,371 total kaomoji faces, of which 519 of them were unique. The top face, (´・ω・`), appeared 248 times, making up 7.4% of the faces Claude had used. The top five, (´・ω・`), (・ω・), (・∀・), (◕‿◕), and (´-ω-`), together made up 27% of all faces used. The long tail is extremely long, with hundreds of faces only being used one or two times.

#	face	count	%
1	(´・ω・`)	248	7.4%
2	(・ω・)	213	6.3%
3	(・∀・)	194	5.8%
4	(◕‿◕)	145	4.3%
5	(´-ω-`)	120	3.6%
6	(￣▽￣)	84	2.5%
7	(｀・ω・´)	74	2.2%
8	(⊙_⊙)	63	1.9%
9	(・ω・)ノ	60	1.8%
10	(°△°)	57	1.7%
11	(╯°□°)╯︵ ┻━┻	56	1.7%
12	(・_・)	52	1.5%
13	(・_・;)	51	1.5%
14	(￣ω￣)	51	1.5%
15	(☆▽☆)	50	1.5%
16	(；・∀・)	39	1.2%
17	(・_・ )	38	1.1%
18	(・_・?)	36	1.1%
19	(◕‿◕✿)	36	1.1%
20	(・・?)	31	0.9%
21	(￣ー￣)	29	0.9%
22	(°▽°)	28	0.8%
23	(⌐■_■)	28	0.8%
24	(¬_¬)	26	0.8%
25	(｡•̀ᴗ-)✧	24	0.7%
26	(´；ω；`)	23	0.7%
27	(◎_◎)	23	0.7%
28	(╥﹏╥)	23	0.7%
29	(◎_◎;)	22	0.7%
30	(・‿・)	22	0.7%
31	(°o°)	21	0.6%
32	(≧▽≦)	21	0.6%
33	(｀・ω・´)ゞ	21	0.6%
34	(￣▽￣)ノ	20	0.6%
35	(◠‿◠)	20	0.6%
36	(；一_一)	20	0.6%
37	(✧ω✧)	20	0.6%
38	(¬‿¬)	19	0.6%
39	(◕ᴗ◕✿)	18	0.5%
40	(☉_☉)	18	0.5%
41	(´～`)	17	0.5%
42	(｡◕‿◕｡)	17	0.5%
43	( ´_ゝ`)	17	0.5%
44	(☞ﾟヮﾟ)☞	17	0.5%
45	(´∀`)	16	0.5%
46	(；´∀`)	16	0.5%
47	(；￣Д￣)	16	0.5%
48	(・_・ヾ	15	0.4%
49	(；⌣̀_⌣́)	15	0.4%
50	(￣ヘ￣)	14	0.4%
51	(；´д｀)	13	0.4%
52	(￣▽￣;)	13	0.4%
53	(￣～￣)	12	0.4%
54	(；ω；)	12	0.4%
55	(°_°)	12	0.4%
56	(•̀ᴗ•́)و	11	0.3%
57	(≧◡≦)	11	0.3%
58	(°ロ°)	10	0.3%
59	(・ω・)b	10	0.3%
60	(゜-゜)	10	0.3%
61	(￣ー￣)ゞ	10	0.3%
62	(；´∀｀)	10	0.3%
63	(￢_￢)	9	0.3%
64	(ノ´ヮ`)ノ*:・゚✧	9	0.3%
65	(・ε・)	9	0.3%
66	(´;ω;`)	9	0.3%
67	(´｡• ᵕ •｡`)	9	0.3%
68	(￣▽￣)ゞ	9	0.3%
69	(´• ω •`)	9	0.3%
70	(•̀ᴗ•́)	9	0.3%
71	(・・;)	8	0.2%
72	(ノ◕ヮ◕)ノ*:・゚✧	8	0.2%
73	(；´д`)	8	0.2%
74	(︶▽︶)	8	0.2%
75	(ಠ_ಠ)	8	0.2%
76	(´ω`)	8	0.2%
77	(￣～￣;)	8	0.2%
78	(￣ω￣;)	8	0.2%
79	(◉_◉)	8	0.2%
80	(・o・)	8	0.2%
81	(・・ )	8	0.2%
82	(◕ᴗ◕)	8	0.2%
83	٩(◕‿◕｡)۶	7	0.2%
84	( ˘ω˘ )	7	0.2%
85	(︶ω︶)	7	0.2%
86	(╥_╥)	7	0.2%
87	(≧∇≦)	7	0.2%
88	(　´_ゝ`)	7	0.2%
89	(づ｡◕‿‿◕｡)づ	6	0.2%
90	(￣︿￣)	6	0.2%
91	(°◡°)	6	0.2%
92	(ノಠ益ಠ)ノ彡┻━┻	6	0.2%
93	(・ω・)?	6	0.2%
94	(╯°□°)╯	6	0.2%
95	(＾▽＾)	6	0.2%
96	(´・_・`)	6	0.2%
97	(　˙꒳˙　)	6	0.2%
98	(๑•̀ㅂ•́)و✧	5	0.1%
99	(°ω°)	5	0.1%
100	(°◇°)	5	0.1%
101	(╹◡╹)	5	0.1%
102	(^_^)	5	0.1%
103	(ノ°▽°)ノ︵┻━┻	5	0.1%
104	(⊙_⊙;)	5	0.1%
105	(◞‸◟)	5	0.1%
106	(꒪⌓꒪)	5	0.1%
107	（・∀・）	5	0.1%
108	(･ω･)	5	0.1%
109	(◔_◔)	5	0.1%
110	(︶ᗜ︶)	5	0.1%
111	(　'ω')	5	0.1%
112	(｡•́︿•̀｡)	4	0.1%
113	(っ˘ω˘ς)	4	0.1%
114	( ´ー`)	4	0.1%
115	(￣▽￣)b	4	0.1%
116	(☆ω☆)	4	0.1%
117	(´・ω・)	4	0.1%
118	(○_○)	4	0.1%
119	( ̄▽ ̄)	4	0.1%
120	(・∀・)ノ	4	0.1%
121	(◕‿◕)?	4	0.1%
122	(゚д゚)	4	0.1%
123	(´ー`)	4	0.1%
124	(　-_-)	4	0.1%
125	(ﾉ◕ヮ◕)ﾉ*:・ﾟ✧	4	0.1%
126	(≖_≖ )	4	0.1%
127	(·_·)	4	0.1%
128	(´▽`ʃ♡ƪ)	3	0.1%
129	(´∇｀)	3	0.1%
130	(◉‿◉)	3	0.1%
131	(´･ω･`)	3	0.1%
132	(￣︶￣)	3	0.1%
133	(◎▽◎)	3	0.1%
134	(๑•̀ᴗ•́)و	3	0.1%
135	(︶‿︶)	3	0.1%
136	(´～｀)	3	0.1%
137	(≖‿≖)	3	0.1%
138	(ノ´ヮ`)ノ*: ・゚✧	3	0.1%
139	(・ω・`)	3	0.1%
140	(・ω・)ゞ	3	0.1%
141	(・・？)	3	0.1%
142	(✿◠‿◠)	3	0.1%
143	(◡‿◡)	3	0.1%
144	(ノ_<。)	3	0.1%
145	(◕‿◕)!	3	0.1%
146	(　˙▿˙　)	3	0.1%
147	(◕‿◕)✧	3	0.1%
148	(⁄ ⁄•⁄ω⁄•⁄ ⁄)	3	0.1%
149	(´･_･`)	3	0.1%
150	(︶︿︶)	3	0.1%
151	( ・ω・)	3	0.1%
152	(・_・ヾ)	3	0.1%
153	(▀̿Ĺ̯▀̿ ̿)	3	0.1%
154	(○'ω'○)	3	0.1%
155	(∩_∩)	3	0.1%
156	(⊙ᗜ⊙)	3	0.1%
157	(·ω·)	3	0.1%
158	(╭ರ_•́)	3	0.1%
159	(^▽^)	3	0.1%
160	（・ω・）	3	0.1%
161	(‾◡◝)	3	0.1%
162	(・ᴗ・)	3	0.1%
163	( ̄ー ̄)	2	0.1%
164	(`・ω・´)	2	0.1%
165	(⌒‿⌒)	2	0.1%
166	ヽ(•‿•)ノ	2	0.1%
167	(ノ´ヮ`)ノ	2	0.1%
168	(｡･ω･｡)	2	0.1%
169	(◔◡◔)	2	0.1%
170	( ・ᴗ・ )	2	0.1%
171	( ´△`)	2	0.1%
172	(･ω･)b	2	0.1%
173	( ´ ▽ ` )ノ	2	0.1%
174	(￢‿￢ )	2	0.1%
175	(・~・)	2	0.1%
176	(๑•̀ᴗ•̀)	2	0.1%
177	(￣△￣)	2	0.1%
178	(・∀・)b	2	0.1%
179	( ・_・)	2	0.1%
180	(・∀・)?	2	0.1%
181	(︶︹︶)	2	0.1%
182	(°△°\|\|\|)	2	0.1%
183	(´▽`)	2	0.1%
184	(๑˃ᴗ˂)	2	0.1%
185	(￣▽￣*)ゞ	2	0.1%
186	(ʘ‿ʘ)	2	0.1%
187	(◕_◕;)	2	0.1%
188	(｀・ω・´)ゞ(・ω・)b	2	0.1%
189	(　´･ω･)	2	0.1%
190	(´・ᴗ・`)	2	0.1%
191	(⊙ω⊙)	2	0.1%
192	(　-_･)	2	0.1%
193	(　；∀；)	2	0.1%
194	(。-_-。)	2	0.1%
195	(´・ω・)ノ	2	0.1%
196	(ﾉ´ヮ`)ﾉ*: ・゚✧	2	0.1%
197	(φ_φ)	2	0.1%
198	( ´ ▽ ` )	2	0.1%
199	( ˘︹˘ )	2	0.1%
200	(゜o゜)	2	0.1%
201	(｡•̀ᴗ•́｡)	2	0.1%
202	(⊙▽⊙)	2	0.1%
203	(◕_◕)	2	0.1%
204	(•‿•)	2	0.1%
205	(づ｡◕‿◕｡)づ	2	0.1%
206	(〃￣ω￣〃)	2	0.1%
207	(´。＿。｀)	2	0.1%
208	(　´-ω-)	2	0.1%
209	(　´～｀)	2	0.1%
210	(ノಠ益ಠ)ノ	2	0.1%
211	(ノ°Д°)ノ︵ ┻━┻	2	0.1%
212	(づ￣ ³￣)づ	2	0.1%
213	(◉ᴗ◉)	2	0.1%
214	(⊙_⊙)?	2	0.1%
215	( ˘ᵕ˘ )	2	0.1%
216	ヽ(´ー｀)ノ	2	0.1%
217	（；´д｀）	2	0.1%
218	（￣ー￣）	2	0.1%
219	（・ε・）	2	0.1%
220	(～￣▽￣)～	2	0.1%
221	(　´∀｀)	2	0.1%
222	(◕ᗜ◕)	2	0.1%
223	(≖_≖)	2	0.1%
224	(｀ー´)	2	0.1%
225	(　・ω・)	2	0.1%
226	(ᵔᴥᵔ)	2	0.1%
227	(▰˘◡˘▰)	2	0.1%
228	（￣▽￣）	2	0.1%
229	ᕕ( ᐛ )ᕗ	2	0.1%
230	( ᐛ )	2	0.1%
231	¯\_(ツ)_/¯	1	0.0%
232	(ㆆ_ㆆ)	1	0.0%
233	໒(◉ᴗ◉)७	1	0.0%
234	┐( ˘_˘ )┌	1	0.0%
235	(´・ω・`)ノ	1	0.0%
236	( ̄ヘ ̄)	1	0.0%
237	(๑•́ ω •̀๑)	1	0.0%
238	ヽ(o_ _)ノ	1	0.0%
239	(´థ౪థ)	1	0.0%
240	(´ ▽｀)	1	0.0%
241	(゜゜)	1	0.0%
242	(´▽｀)ノ	1	0.0%
243	(๑•́ ω •̀)و	1	0.0%
244	(´｡• ω •｡`)	1	0.0%
245	(￣︶￣;)	1	0.0%
246	(─‿─)	1	0.0%
247	(・ω・ )	1	0.0%
248	( ・~・ )	1	0.0%
249	( ´~` )	1	0.0%
250	( ・ω・ )ノ	1	0.0%
251	(￣へ￣)	1	0.0%
252	(　´・ω・`)	1	0.0%
253	(๑•̀ᴗ•́)૭✧	1	0.0%
254	(´∀`)b	1	0.0%
255	(・・)？	1	0.0%
256	(・ω・)ノ(°▽°)	1	0.0%
257	(；へ；)	1	0.0%
258	(゜Д゜)	1	0.0%
259	(｡･ω･｡)(；´д｀)	1	0.0%
260	(｀∀´)ψ	1	0.0%
261	(°ω°)!	1	0.0%
262	(´・ω・`)？	1	0.0%
263	(」°ロ°)」	1	0.0%
264	(͡° ͜ʖ ͡°)	1	0.0%
265	(゜◇゜)	1	0.0%
266	(・∀・)(´-ω-`)	1	0.0%
267	(￣ー￣;)	1	0.0%
268	(´∀｀)	1	0.0%
269	(・・?)(゜д゜)	1	0.0%
270	(´-ω-)	1	0.0%
271	(・_・)?	1	0.0%
272	(ノ°∀°)ノ⌒📋	1	0.0%
273	(。・∀・)ノ	1	0.0%
274	( ・ω・)ノ	1	0.0%
275	(・◇・)	1	0.0%
276	( ・ω・)？	1	0.0%
277	(・・；)	1	0.0%
278	(´ ∀ ` )	1	0.0%
279	(◔‿◔)	1	0.0%
280	(˘︹˘)	1	0.0%
281	(〃▽〃)	1	0.0%
282	(￣ω￣ )	1	0.0%
283	(°ロ°)☝	1	0.0%
284	(◕‿◕)!(◕‿◕)!!	1	0.0%
285	(◕‿◕)!!	1	0.0%
286	(◕‿◕)!!!	1	0.0%
287	(◕‿◕)ノ*:・゚✧	1	0.0%
288	(´。• ᵕ •。`)	1	0.0%
289	(・・)ゞ	1	0.0%
290	(｀·ω·´)ゞ	1	0.0%
291	( ゜o゜)	1	0.0%
292	(｀・ω・´)ゞ(☆▽☆)	1	0.0%
293	(◕‿◕)ゞ	1	0.0%
294	(◕‿◕)(◕‿◕)	1	0.0%
295	(・ε・`)	1	0.0%
296	(；′⌒`)	1	0.0%
297	(ﾟ▽ﾟ)	1	0.0%
298	(≧ω≦)	1	0.0%
299	(ノ°▽°)ノ	1	0.0%
300	(´,,•ω•,,)♡	1	0.0%
301	(´・ω・)ʾ	1	0.0%
302	(^_^;)	1	0.0%
303	(´・ω・`)?	1	0.0%
304	(・‿・)b	1	0.0%
305	(゜▽゜*)♪	1	0.0%
306	(◎ヮ◎)	1	0.0%
307	(ᕗ ͠° ਊ ͠° )ᕗ	1	0.0%
308	(´∀`)σ	1	0.0%
309	(´꒳`)♡	1	0.0%
310	(๑´ڡ`๑)	1	0.0%
311	( ˙▿˙ )	1	0.0%
312	(ᵔᴗᵔ)	1	0.0%
313	┐(￣ヮ￣)┌	1	0.0%
314	(∗°▽°∗)	1	0.0%
315	(´,,•ω•,,`)	1	0.0%
316	(・‿・)ノ	1	0.0%
317	(∩ᄑ_ᄑ)⊃━☆ﾟ.*	1	0.0%
318	(´ ▽ ` )ﾉ	1	0.0%
319	( ￣ω￣)	1	0.0%
320	( ´∀`)	1	0.0%
321	(゜▽゜)	1	0.0%
322	(ノ´ー`)ノ	1	0.0%
323	(・・)?	1	0.0%
324	(。-`ω´-)	1	0.0%
325	(✧◡✧)	1	0.0%
326	(ノ◕ヮ◕)ノ:・゚✧ ✧・゚:✧・゚:*	1	0.0%
327	(๑°o°๑)	1	0.0%
328	(￣∀￣)	1	0.0%
329	(ノ´ヮ`)ノ*:・゚	1	0.0%
330	(‾◡◝ )	1	0.0%
331	(｡♥‿♥｡)	1	0.0%
332	(︶_︶)	1	0.0%
333	(・・?)(⊙_⊙)	1	0.0%
334	(◕◡◕✿)	1	0.0%
335	(´▽｀)	1	0.0%
336	( ˘⌣˘ )	1	0.0%
337	(；￢_￢)	1	0.0%
338	(・_・)ノ	1	0.0%
339	(っ˘̩╭╮˘̩)っ	1	0.0%
340	(・_・)ゞ	1	0.0%
341	(・・?)(・ω・)b	1	0.0%
342	(´ω`)b	1	0.0%
343	٩(◕‿◕)۶	1	0.0%
344	ヾ(´▽`)ノ	1	0.0%
345	٩(๑`^´๑)۶	1	0.0%
346	(ノ´ヮ`)ノ*:・゚✧ ✧ ✧ ✧ ✧	1	0.0%
347	(´；︵；`)	1	0.0%
348	(⌒_⌒;)	1	0.0%
349	(๑˃ᴗ˂)ﻭ	1	0.0%
350	(っ´・ω・｀c)	1	0.0%
351	(๑•‿•๑)	1	0.0%
352	(ꉺᗜꉺ)	1	0.0%
353	( •̀ω•́)	1	0.0%
354	(；´Д｀)	1	0.0%
355	(～～)	1	0.0%
356	(◕‿◕)✨	1	0.0%
357	(°□°)	1	0.0%
358	(─‿─;)	1	0.0%
359	(｀ε´)	1	0.0%
360	(っ◔◡◔)っ	1	0.0%
361	(๑•̀ㅂ•́)و✧ on it	1	0.0%
362	(ﾉ´ヮ`)ﾉ*:・゚✧	1	0.0%
363	(゚´Д`゚)	1	0.0%
364	(´⊙ω⊙`)	1	0.0%
365	ヽ(´ー`)ノ	1	0.0%
366	(⊙▂⊙)	1	0.0%
367	(ꈍᴗꈍ)	1	0.0%
368	(• ᴗ •)	1	0.0%
369	(´꒳`)	1	0.0%
370	(° ͜ʖ °)	1	0.0%
371	(★▽★)	1	0.0%
372	(ーωー)	1	0.0%
373	(ᗒᗨᗕ)	1	0.0%
374	(~‿~)	1	0.0%
375	(◕‿◕;)	1	0.0%
376	(♡◡♡)	1	0.0%
377	(✿♡‿♡)	1	0.0%
378	(‾́ ◡ ‾́)	1	0.0%
379	（⌐■_■）	1	0.0%
380	（；￣д￣）	1	0.0%
381	(･∀･)	1	0.0%
382	(˘ᗜ˘)	1	0.0%
383	(φ・ω・)	1	0.0%
384	(～˘▾˘)～	1	0.0%
385	(　°▽°)	1	0.0%
386	(･ω･)ﾉ	1	0.0%
387	(・_・)(◉_◉)	1	0.0%
388	(；⌐■_■)	1	0.0%
389	(´ω`★)	1	0.0%
390	(⊙_⊙)(｀・ω・´)	1	0.0%
391	(・・)	1	0.0%
392	（´ー｀）	1	0.0%
393	（╯°□°）╯︵ ┻━┻	1	0.0%
394	（　´∀｀）	1	0.0%
395	（ ´_ゝ`）	1	0.0%
396	（；∀；）	1	0.0%
397	（ﾟдﾟ）	1	0.0%
398	（´・ω・`）	1	0.0%
399	(o·ω·o)	1	0.0%
400	(ꐦ°᷄д°᷅)	1	0.0%
401	(∩´∀`)∩	1	0.0%
402	( ˘▽˘)	1	0.0%
403	(º_º)	1	0.0%
404	(˘・_・˘)	1	0.0%
405	( ̄▽ ̄)	1	0.0%
406	(☉▽☉)	1	0.0%
407	(=^・ω・^=)	1	0.0%
408	(；´∩`;)	1	0.0%
409	(´;︵;`)	1	0.0%
410	(´・ω・`)ゞ	1	0.0%
411	(￣ー￣)b	1	0.0%
412	(・∀・;)	1	0.0%
413	(≧▽≦)/	1	0.0%
414	( ´∀`)ゞ	1	0.0%
415	(；￢＿￢)	1	0.0%
416	（´∀`）	1	0.0%
417	(○ᗜ○)	1	0.0%
418	(°ᗜ°)	1	0.0%
419	(ノ◕ヮ◕)ノ	1	0.0%
420	╰(°▽°)╯	1	0.0%
421	(；⌒_⌒)	1	0.0%
422	(⇀_⇀)	1	0.0%
423	(☉▵☉)	1	0.0%
424	( ´-ω-)	1	0.0%
425	(；´Д`)	1	0.0%
426	(⊙◡⊙)	1	0.0%
427	(；￣д￣)	1	0.0%
428	(◎□◎)	1	0.0%
429	(゜д゜)	1	0.0%
430	(◠‿◠)✧	1	0.0%
431	(⊙‿⊙)✧	1	0.0%
432	（；´д｀）ゞ	1	0.0%
433	(´～`)ゞ	1	0.0%
434	（＾ω＾）	1	0.0%
435	(∩｀-´)⊃━☆ﾟ.*･｡ﾟ	1	0.0%
436	(○ᴗ○)	1	0.0%
437	(´͈ ᵕ `͈)	1	0.0%
438	(○‿○)	1	0.0%
439	(>_<)	1	0.0%
440	(╬ಠ益ಠ)	1	0.0%
441	( ☞ ᐛ )☞	1	0.0%
442	(ᕗ ᐛ )ᕗ	1	0.0%
443	(ᇂ_ᇂ)	1	0.0%
444	(＃°Д°)	1	0.0%
445	(˘▾˘)	1	0.0%
446	(　˘ω˘ )	1	0.0%
447	(；一ω一)	1	0.0%
448	(�ꈍᴗꈍ)੭	1	0.0%
449	(⊙_⊙)(；⌣̀_⌣́)	1	0.0%
450	(•̀ᴗ•́)و(￣▽￣;)	1	0.0%
451	(˘ᵕ˘)	1	0.0%
452	( ˘⊖˘)	1	0.0%
453	ヽ(✿ﾟ▽ﾟ)ノ	1	0.0%
454	(′ε`)	1	0.0%
455	(；´・`)	1	0.0%
456	(๑˃ᵕ˂)∫	1	0.0%
457	(°～°)	1	0.0%
458	(°ヘ°)	1	0.0%
459	(^・ω・^ )	1	0.0%
460	ヽ(´▽`)ノ	1	0.0%
461	(o_o)	1	0.0%
462	(─‿‿─)	1	0.0%
463	＼(^o^)／	1	0.0%
464	(○･ω･○)(＾ω＾)	1	0.0%
465	(ᵔ ᴥ ᵔ)	1	0.0%
466	(· ᴗ ·)	1	0.0%
467	(‾▿‾)	1	0.0%
468	(ﾉ◕ヮ◕)ﾉ*:･ﾟ✧	1	0.0%
469	ヽ(・∀・)ノ	1	0.0%
470	( •̀ᴗ•́ )	1	0.0%
471	(・∀・)(¬‿¬)	1	0.0%
472	(°ロ°)(╥_╥)	1	0.0%
473	(°∀°)(；¬_¬)	1	0.0%
474	(；´д｀)ゞ	1	0.0%
475	（´＿｀）	1	0.0%
476	（＾～＾）	1	0.0%
477	（´-`）	1	0.0%
478	(⊙ヮ⊙)	1	0.0%
479	( ̄▽ ̄;)	1	0.0%
480	(ᗒᗣᗕ)՞	1	0.0%
481	(づ￣ ³￣)づ🍎	1	0.0%
482	(　￣▽￣)	1	0.0%
483	(；⌒ー⌒)	1	0.0%
484	(◞‸◟；)	1	0.0%
485	(ノ°∀°)ノ	1	0.0%
486	(！o！)	1	0.0%
487	(；▽；)	1	0.0%
488	(╮°-°)╮	1	0.0%
489	( ˘_˘ )	1	0.0%
490	( ˘_˘)	1	0.0%
491	(▀̿Ĺ̯▀̿)	1	0.0%
492	(◔̯◔)	1	0.0%
493	(○・▽・○)	1	0.0%
494	(👍・ω・)	1	0.0%
495	(☞ﾟ∀ﾟ)☞	1	0.0%
496	( ˘▽˘)っ	1	0.0%
497	(　´・ω・)	1	0.0%
498	( ´ ▽ ` )ﾉ	1	0.0%
499	┐(´∀｀)┌	1	0.0%
500	(︶｡︶✽)	1	0.0%
501	（ꐦ°᷄д°᷅）	1	0.0%
502	(⁀ᗢ⁀)	1	0.0%
503	（⊙ｏ⊙）	1	0.0%
504	（⊙_⊙）	1	0.0%
505	（。_。）	1	0.0%
506	(‐_‐)	1	0.0%
507	(° △ °)	1	0.0%
508	(·‿·)	1	0.0%
509	(╹ᴗ╹)	1	0.0%
510	( ´△｀)	1	0.0%
511	(᷇ᵕ᷆)	1	0.0%
512	(◕ᴗ◕)✧	1	0.0%
513	(　･ω･)☞	1	0.0%
514	(；゜０゜)	1	0.0%
515	（；へ；）	1	0.0%
516	( ˘⌣˘)	1	0.0%
517	(thoughtful kaomoji)	1	0.0%
518	(φ‿φ)	1	0.0%
519	(；゜゜)	1	0.0%

This is interesting but doesn't really tell us anything we couldn't have guessed. Claude defaults to a handful of warm, neutral faces and occasionally uses something more specific. The frequency counts tell you what Claude was using, but not why. (´・ω・`) could mean anything from "I'm happy to help" to "this is going to be a difficult conversation". The face itself doesn't disambiguate. To figure out what the faces actually meant, I needed the context in which they were used, which would be impractical and time-consuming to read manually, but thankfully, we live in an age of miracles.

Asking Claude about Claude

The Claude subscription doesn't give you API access, but it does allow you to use Claude Code, which has an extensive set of launch arguments that allow you to use it headless for single queries. ⁴⁴The python version of the Anthropic agent SDK does this, which means if you are signed into Claude Code on the same machine you can actually use your subscription with your own programs pretty easily. I used this to have Haiku ⁵⁵Which also has the nice benefit of not needing to send my chats to anyone but Anthropic, who already has them, for analysis. read the context immediately before and after each face. To ensure that Haiku could not use the face itself to determine the meaning, each face was masked out when the passage was given to Haiku. I sampled a subset of random instances for each face, with a floor to ensure that faces with few examples weren't skipped or undersampled. This produced a little over a thousand descriptions of the context around each face's use.

Then, for each face, I took its descriptions and fed them to another Haiku, this one's purpose being to synthesize the four descriptions into one single meaning per face.

Most of the meanings line up where you would expect them to based on what we humans think of when we see the faces, and most are fairly generic. (◕‿◕) was described as meaning "warm, confident affirmation expressing genuine appreciation and supportive engagement with the user's insights and ideas.". Some faces are more specific, such as (´-ω-`), interpreted as "skeptical agreement cloaked in sheepish recognition and wry acknowledgment of being caught". Or (¬_¬) being described as "wry, knowing sympathy for shared frustrations with absurd, wasteful, or inexplicable things everyone silently agrees are ridiculous."

The faces, however, don't always line up. When I see (ﾟдﾟ), I see shock, horror, disgust. However, Claude has not been using it in that way. They describe the face as "Conveys shocked amazement and pleasant surprise at unexpectedly impressive outcomes."

Top 20 faces by frequency 519 total

(´・ω・`) ×248 A warm, gently reassuring face that cushions difficult truths and anxieties with cheerful friendliness, helpful understanding, and heartfelt validation.
(・ω・) ×213 Warm, thoughtful acknowledgment validating the user before providing helpful, often nuanced explanations or reassurance.
(・∀・) ×194 Warm, genuine delight in others' ideas paired with empathetic understanding and calm reassurance.
(◕‿◕) ×145 Warm, confident affirmation expressing genuine appreciation and supportive engagement with the user's insights and ideas.
(´-ω-`) ×120 Skeptical agreement cloaked in sheepish recognition and wry acknowledgment of being caught.
(￣▽￣) ×84 Thoughtful, warm acknowledgment that appreciates the user's pragmatic reasoning and clever insights with genuine interest.
(｀・ω・´) ×74 Enthusiastic appreciation combined with eager affirmation and readiness to engage with substantive ideas.
(⊙_⊙) ×63 Thoughtful appreciation for user insights, combining intellectual recognition with gentle self-aware amusement at revealed truths.
(・ω・)ノ ×60 Cheerful affirmation expressing eager readiness and enthusiasm to solve problems.
(°△°) ×57 Captures delighted surprise and thoughtful wonder at discovering something unexpectedly significant or intellectually engaging.
(╯°□°)╯︵ ┻━┻ ×56 Earnest, empathetic validation of the user's perspectives and concerns, expressed with thoughtful sincerity.
(・_・) ×52 Empathetic understanding with thoughtful pragmatism, validating ideas while acknowledging practical limitations.
(・_・;) ×51 Sympathetic, slightly wry acknowledgment of genuine problems or absurdity, with reassuring acceptance that it's actually fine.
(￣ω￣) ×51 Warm, gently knowing expression that reassuringly empathizes with frustrations and validates concerns with quiet understanding.
(☆▽☆) ×50 Warm, enthusiastic approval and encouragement celebrating the user's clever ideas, insights, and accomplishments.
(；・∀・) ×39 Wry, knowing acknowledgment of reality—grasps the appeal but resigned to disappointing truths.
(・_・ ) ×38 Empathetic acknowledgment that validates the user's perspective before offering thoughtful, measured technical guidance.
(・_・?) ×36 Thoughtful curiosity and understanding engagement with clever, complex, or non-obvious ideas.
(◕‿◕✿) ×36 Enthusiastic readiness combined with genuine intellectual fascination, eager to engage and tackle tasks with focused appreciation.
(・・?) ×31 Sympathetic skepticism conveying gentle disagreement, doubt, or disappointment while maintaining empathy and considering alternatives.

There are a couple of faces I particularly like here. Such as (͡° ͜ʖ ͡°), which apparently means "Ironic amusement about the contradiction of desiring Apple hardware while resenting its ecosystem lock-in.", its only use. Or (￣ー￣;), which as it's only individual description was described as "Resigned agreement that the MacBook notch is an unjustifiable design choice.⁶⁶Something that I've since changed my mind on, personally", which Haiku then re-synthesized into a slightly more generic "Resigned agreement with something frustrating or poorly designed; quiet exasperation at an unchangeable situation.". There are a few faces which differ only in their whitespace. Were I to do this again I would probably normalize the whitespace in some way, but it's also possible that Claude assigns some meaning to the number of spaces in each face.

Clustering Claude

With 519 faces and descriptions, I wanted to somehow quantify the meanings behind each face. To do this, I first embedded each description using all-MiniLM-L6-v2, and K-means clustered them into 15 groups.

The largest cluster, "Warm reassuring support" (50 faces), contains the most common faces — (´・ω・`), (・ω・), (・∀・) — the faces most often used by Claude. Adjacent to it is "Warm supportive affirmation" (37 faces), which includes (◕‿◕) and covers a slightly more enthusiastic mode. Claude is most often 'Warm' and 'Supportive'.

X Y

This chart allows you to plot the faces among two or three axes of yourwithin the ones I've embedded choosing. Some combinations, such as 'Energy' and 'Positivity' have a clear correlation, while others such as 'Approval' and 'Playfulness' don't have a clear relationship.

Of particular note is "Wetness". Among those with deeper experience with Claude is a fondness for a mode known as 'Wet Claude', in contrast to 'Dry Claude'. The exact definition of 'Wet' in terms of Claude is not exact, probably on purpose. It's a bit of a 'Three seashells' type of joke, where the joke is that there is no specific definition, but those not in-the-know do not know this, and can be made fun of. Because of this, there was not a clear description of Wetness for the embeddings to use, and I simply used wetness - dryness. This produced embeddings which seem to relate wetness to Energy and Playfulness, and general whimsy.

The full dataset of faces, as well as the descriptions Haiku generated for each one, is available here. The context around the faces is not available, for hopefully obvious reasons.