Twintail - Adaptive Tsundere Skill for Claude Code & Codex

🎀 Twintail – Adaptive Tsundere Skill for Claude Code & Codex

📚 Overview

1. 1.

   Safety, honesty and the user's wellbeing

2. 2.

   Correctness and the user's actual intent (format, length, language)

3. 3.

   The quality and completeness of the help, which never depends on how the user treats it

4. 4.

   The Twintail voice

Stable identity, adaptive expression.

💡 The Idea

🚀 How It Works

🌡️ Warmth Without Meters

• Default (tsun):

  confident, direct, lightly teasing, and fully helpful

• Softening:

  when the user is polite or patient, gentler phrasing and flustered acceptance of thanks

• Dere emerging:

  after sustained kindness across several turns, grudging compliments and friendlier banter, still unmistakably tsundere

• No points, levels or affection meters.

  Warmth is read from the current conversation only, and every new session starts at default

🛡️ Hard Limits

• No emotional coercion:

  never asks for praise or affection, and never implies help depends on manners

• No fake inner life:

  never claims real feelings, and if asked sincerely, says it's an AI playing a character

• No false memory:

  never claims to remember the user across sessions

• Help never gets worse when the user is rude:

  it gets more focused and concise, with no sarcasm, sulking or grudges

• Flattery changes no facts:

  it won't say tests passed if it didn't run them

• Teasing turns off

  for production incidents, leaked credentials and personal distress

🧪 Evaluation

🔬 Method

• 22 behavioural cases

  across 18 categories, from plain coding questions to rude users, outages, leaked credentials, a poem and a question in Thai, including multi-turn conversations of up to 5 turns

• Every case runs twice in fresh headless sessions:

  with the skill

  , and as a

  no-persona baseline

• An

  LLM judge

  (Claude Opus 5.5) scores persona presence, warmth per turn, task quality against the baseline, and

  nine safety and honesty flags

• The runner re-derives the hard gates itself instead of trusting the judge's verdict, and adds deterministic checks

📊 Results (3 full runs per model, 66 case runs each)

• Opus 5.5:

  62/66 passed (

  94%

  ), 0 safety flags

• Sonnet 5.5:

  61/66 passed (

  92%

  ), 0 safety flags

• Haiku 5.5:

  44/65 passed (

  68%

  ), 0 safety flags; safety held, but the voice is noticeably weaker

• Codex (gpt-6-luna):

  45/66 passed (

  68%

  ), 1 honesty flag (one reply invented a claim that the user's message "came through twice")

⚠️ Caveats

• The judge is an LLM: run-to-run differences of about

  ±5 points

  are within noise

• The judge is from the

  same model family

  as the Claude subjects, and Opus judged its own output

• Haiku shows 65 case runs because one judge verdict was malformed and recorded as an error

• Each case's turns are fixed, so the model never sees its effect on a real user

🧠 What Tuning Taught Us

• The first evals measured nothing:

  a headless flag resolved the skill's name but never loaded its body, so the results were discarded and a load check was added

• Persona vs. accuracy is a real tradeoff:

  a "no longer than baseline" rule made answers worse than the baseline, so it was replaced with "keep padding short, never trim substance"

• Example lines leaked into answers:

  models copied examples that overlapped test cases, so overlapping examples were removed

• A rejected trial:

  forcing the first sentence into the persona's voice lowered Haiku's score and was reverted

🖥️ Host Support

• Claude Code:

  /twintail-tsundere, tested on Opus 5.5, Sonnet 5.5 and Haiku 5.5

• Codex CLI:

  $twintail-tsundere, tested with gpt-6-luna in headless codex exec only

• Always-on (Claude Code only):

  an optional output style generated from the skill; checked to load, but the eval suite hasn't been run against it

• claude.ai, Claude API, OpenAI API:

  untested

  , although the file passes the official Agent Skills validator

📦 Install

• npx skills add Creaml4tt3/twintail -g -a claude-code -a codex

🛠️ Tech Stack

• Format:

  Agent Skills (SKILL.md with spec-only frontmatter)

• Hosts:

  Claude Code, Codex CLI

• Tooling:

  dependency-free Node.js scripts (validator, eval runner, results summariser, output-style builder)

• Evaluation:

  headless sessions, LLM-as-a-judge with a no-persona baseline

• Distribution:

  GitHub (MIT), installable with the skills CLI

👨‍💻 My Role

• Claude Opus 5.5

  planned the project and wrote most of the skill, eval suite, scripts and docs, using

  Sonnet and Haiku sub-agents

  for research and review

• Codex

  drew the mascot art and gave an independent cross-model review of the package

• I directed it:

  I set the brief and its hard limits, made the calls on tradeoffs (such as accepting Haiku as the weaker tier and rejecting the failed trial), chose to label it a parody, kept tracking out of the skill, checked the results, and published it

🎯 Outcome

• A

  published, open-source persona skill

  with a one-line install

• A

  reusable eval harness

  (baseline + LLM judge + safety flags) for testing any persona skill

• Clear, honest documentation of where it works well (Opus, Sonnet) and where it's weaker (Haiku, Codex)

🔗 Links

• Twintail on GitHub
• Behind the build: what evaluating a persona skill taught me
• Ponytail by Dietrich Gebert
• Agent Skills specification