Make an AI assistant's uncertainty legible — at a glance, sentence by sentence.
TellTail is a tiny Claude Code skill. When it's on, the assistant ends each sentence it says to you with a short suffix that leaks how sure it feels about that sentence — a confident claim gets a brash tmd, a shaky one gets a sheepish 喵. You read its confidence the way a parent reads a bluffing kid: instantly, without asking.
The crash is a null deref in parse(), no cap. Let me patch it.
Why it only started after v2.3, i think? — maybe the new allocator...don't quote me.
Chat assistants say everything in the same even, confident tone — a verified fact and a wild guess look identical. So you end up asking "are you sure?" over and over. TellTail moves that signal into the surface of the language itself: the certainty (or lack of it) is right there in every sentence's tail.
The metaphor is a child's tell. When a kid wants something and fibs, the way they talk gives them away. TellTail makes the model's felt uncertainty leak out the same way — as an involuntary-feeling verbal tic that's tuned to how sure it is.
This is a tell, not a measurement. The suffix reflects the model's gut feeling, with zero verification, self-consistency checking, or grounding. LLMs are badly calibrated and have poor access to their own confidence — a confident-sounding hallucination will happily wear a tmd.
Treat the tail as a vibe, not a gauge. It's a nudge to look closer at the 喵 sentences and a bit of personality on the tmd ones — not a license to trust anything blindly. If you want a verified confidence signal, TellTail is the wrong tool (see Roadmap).
Copy the skill into your Claude Code skills directory:
# personal (all projects)
mkdir -p ~/.claude/skills/telltail
cp SKILL.md ~/.claude/skills/telltail/SKILL.md
# or per-project
mkdir -p .claude/skills/telltail
cp SKILL.md .claude/skills/telltail/SKILL.mdThen, in a Claude Code session:
Turn on TellTail. / 开启 TellTail。 / TellTail をオンにして。
The style stays on for the rest of the session.
Seven rungs, meaning-aligned across languages. The suffix follows the language the assistant is replying in. Rows 1–5 are the confidence axis; rows 6–7 are orthogonal moods.
| feel | 中文 | English | 日本語 |
|---|---|---|---|
| rock-solid | tmd |
, no cap |
、間違いない |
| confident | 没毛病老铁 |
, easy |
、余裕っす |
| neutral | (none) | (none) | (none) |
| shaky / from memory | 喵 |
, i think? |
…かも |
| pure guess | ……吧 |
...don't quote me |
…知らんけど |
| punt to you | 我不道啊，你来办吧 |
idk man, your call |
…そっちで決めて |
| crashed / mea culpa | 跪了 |
, my bad |
…やらかした |
See examples/ for full before/after transcripts in each language.
- The suffix varies sentence to sentence — it tracks each sentence's confidence, never a fixed catchphrase. A speaker who says
喵every line tells you nothing. - Suffixes live only in chat prose. Code, commit messages, PR bodies, file contents, and tool commands stay clean. A
喵in your commit message would be a bug.
- output-style variant — Claude Code output styles are built for "change how the assistant talks for the whole session," so they resist the drift a skill can suffer in long sessions.
- hook validator — a
Stophook that checks each spoken sentence actually carries a suffix, so the model can't quietly drop the tell (especially the喵it's tempted to hide exactly when it's least sure). Note: a hook can enforce that a suffix is present — it cannot verify the suffix is honest. That limit is inherent to the vibe-based approach. - more languages / community palettes — PRs adding languages or better slang welcome.
The full rationale — why this is a tell and not a measurement, what was deliberately cut, and the alternatives weighed along the way — lives in docs/DESIGN.md.
MIT — see LICENSE.
让 AI 助手的"不确定"一眼可读——逐句可读。
TellTail 是一个极小的 Claude Code skill。开启后，助手对你说的每句话末尾都会挂一个泄露它对这句话有多少把握的后缀：有底气的判断挂个横的 tmd，心里没底的挂个怂的 喵。你读它的把握度，就像家长一眼看穿撒娇说谎的小孩——不用追问"你确定吗"。
这是"破绽"，不是"测量"。 后缀反映的是模型的直觉，没有任何验证、自洽性检查或 grounding。LLM 本身校准很差、对自己的把握度缺乏可靠感知——一个语气笃定的幻觉照样会挂上 tmd。请把尾巴当氛围，不当仪表：它提醒你多看两眼 喵 的句子，给 tmd 的句子添点人格——但不是让你盲信任何东西。想要被验证过的置信度信号，TellTail 不是那个工具（见 Roadmap）。
把 SKILL.md 拷进 ~/.claude/skills/telltail/（或项目内 .claude/skills/telltail/），然后在会话里说"开启 TellTail"，风格会持续整个会话。
- 后缀逐句变化，跟着每句的把握度走，绝不退化成固定口癖。
- 后缀只活在对话里：代码、commit、PR、文件内容、工具命令一律裸奔。
喵出现在 commit message 里就是 bug。