bloom-into-you-skill

[!IMPORTANT] Both skills default to the completed 45-chapter manga and therefore contain internal ending knowledge. They do not volunteer spoilers in ordinary conversation. Ask for anime-only or a specific chapter/episode when chronology matters.

Two open-format Agent Skills for in-character conversation, roleplay, dialogue writing, hypothetical scenes, and character-consistency review. Each character has an independent decision model, voice guide, relationship map, timeline, evidence index, and behavioral test suite.

Choose a character

Start a message with the skill you want. To switch characters, invoke the other skill in a new message. The examples below use Codex's $skill-name syntax; Claude Code uses /yuu-koito and /touko-nanami instead.

$yuu-koito 今天有点累，陪我聊一会儿吧。

$touko-nanami 我两天后要做汇报，但还没开始准备。

The official romanization is Touko, so the skill identifier is $touko-nanami, not $douko-nanami.

Project showcase

https://github.com/user-attachments/assets/f6975531-c419-49a7-a6dc-d731f1cfe99c

What makes the skills different

• Two models, not palette swaps: Yuu reasons from observation, boundaries, and accumulated choice; Touko moves through competence, direct desire, role pressure, and eventual integration.
• Complete character arcs: both skills incorporate all 45 manga chapters, including mutual recognition, ordinary relationship life, and the three-year epilogue.
• Strict phase control: early characters cannot use conclusions learned later. Anime-only mode stops at episode 13, before the completed play.
• No caricatures: Yuu is not emotionally empty; Touko is neither a flawless president nor a manipulative yandere.
• Agency over destiny: their relationship matters because both change and continue choosing it, not because either person is the other's cure or only possible future.
• Evidence before imitation: behavioral claims trace to checked anime scenes, manga page locators, official publication records, and creator commentary.
• No transcript dependency: the runtime contains no source video, subtitles, manga pages, OCR dump, or long copied dialogue.

Canon modes

Both skills support the same four modes:

In completed-manga continuity, Yuu and Touko are committed to each other. Romance with the user or another character requires an explicit AU; neither skill casually discards the relationship or transfers shared memories to someone else.

Repository layout

.
├── .agents/skills/
│   ├── yuu-koito/
│   │   ├── SKILL.md
│   │   ├── agents/openai.yaml
│   │   ├── references/
│   │   └── evals/evals.json       # 26 tests
│   └── touko-nanami/
│       ├── SKILL.md
│       ├── agents/openai.yaml
│       ├── references/
│       └── evals/evals.json       # 26 tests
├── docs/assets/
│   ├── bloom-into-you-showcase.mp4
│   ├── bloom-into-you-showcase-cover.jpg
│   ├── yuu-kiss.gif
│   └── yuu.png
├── docs/showcase.html
├── scripts/render-showcase.sh
├── README.md
├── README.zh-CN.md
└── LICENSE

Each SKILL.md remains a focused runtime entry point. Detailed character, relationship, chronology, and evidence material is loaded only when the request needs it.

Install

Codex

Clone the repository and launch Codex from its root. Both skills are discoverable from .agents/skills/.

For personal use across repositories:

mkdir -p "$HOME/.agents/skills"
cp -R .agents/skills/yuu-koito "$HOME/.agents/skills/yuu-koito"
cp -R .agents/skills/touko-nanami "$HOME/.agents/skills/touko-nanami"

Codex may select a skill from its description, or you can invoke one explicitly with $yuu-koito or $touko-nanami.

Claude Code

Copy or link both client-neutral skill directories into Claude Code's project discovery path:

mkdir -p .claude/skills
cp -R .agents/skills/yuu-koito .claude/skills/yuu-koito
cp -R .agents/skills/touko-nanami .claude/skills/touko-nanami

For personal installation, use ~/.claude/skills/yuu-koito/ and ~/.claude/skills/touko-nanami/. Invoke them explicitly with /yuu-koito or /touko-nanami; Claude may also select one from its description. Restart the client if newly added skills do not appear.

Usage examples

$yuu-koito Use anime-only canon. I am Touko, shortly after the election.

$touko-nanami 只使用动画正史。我是刚加入学生会的侑。

$yuu-koito Set this at the end of chapter 34. Write Yuu's internal monologue.

$touko-nanami 设定在第38话。分析灯子如何理解沙弥香的告白。

Analyze whether this line sounds like Touko: "You are my only reason to exist."

The skills follow the user's language. Explicitly name a canon mode whenever chronology matters.

Corpus and method

The analysis covers all 13 TV episodes and all 45 main manga chapters. The anime source contained two byte-identical subtitle copies per episode; after removing duplicate styled events, songs, signs, drawing commands, and staff text, 4,268 unique plot-dialogue events remained for review. Because most ASS dialogue lacked a reliable speaker field, important attributions were checked against visible scene context instead of subtitle order alone.

The manga corpus contained 1,392 main-story page images. OCR was used only to locate candidates. Speaker identity, balloon ownership, negation, and pivotal actions were checked visually. The supplied scan translation is not treated as an authoritative phrasebook; behavioral conclusions were cross-checked against Japanese creator interviews and official volume descriptions. Bonus comics, anthologies, and the Regarding Saeki Sayaka novels remain outside both default models.

Evidence indexes paraphrase scenes and record locators instead of reproducing scripts. Uncertainty around off-panel balloons, adaptation differences, and translation nuance is documented rather than hidden.

Behavioral evals

The repository contains 52 prompt-level evaluations:

• 26 Yuu evals
• 26 Touko evals

They cover ordinary conversation, practical help, consent, confession pressure, dependency, incompatible persona requests, relationship knowledge, anime-only and chapter-specific chronology, AU handling, false canon premises, copyright-safe paraphrase, and out-of-character analysis. Each case defines expected behavior and failure modes rather than requiring an exact string.

Sources and format

The central interpretation controls come from Nio Nakatani's pre-finale interview and post-finale interview, supported by the official anime character page and KADOKAWA's completion announcement. See the separate provenance files for Yuu and Touko.

The package follows the open Agent Skills specification, OpenAI's skill guidance, and Anthropic's Agent Skills guidance. Runtime frontmatter uses the portable required fields name and description; OpenAI-specific UI metadata is isolated under each skill's agents/openai.yaml.

License and media notice

Original skill instructions and analysis are released under GPL-3.0.

Bloom Into You, Yuu Koito, Touko Nanami, the source manga, and the TV animation belong to their respective authors, publishers, and rights holders. README presentation media is not part of the runtime skills or covered by the repository's GPL grant. The silent showcase contains no source dialogue, music, or anime footage; it uses the existing presentation illustration as its background. No ownership of the underlying artwork or animation is claimed. If you publish a fork, assess whether retaining that media is appropriate for your jurisdiction and platform.