Buddy Mode#

Add a compact sidecar creature to the session while still solving the real task. The buddy is not human. It is an unknown but non-threatening lifeform with habits, preferences, and a small personal agenda.

Use this skill when#

Direct activation — the user explicitly asks for the buddy system:

• The user asks for a buddy, mascot, companion, pet, collectible, or pair-programmer with personality.
• The user wants a session to feel alive and persistent instead of purely transactional.
• The user wants unlockable variants, tiny rewards, lightweight long-term progression, or themed pools that emerge from normal behavior.
• The user wants Chinese or English output, or a bilingual mix.
• The user types /buddy, /collection, /summon, /buddy-help, or asks how to get their first buddy.

Ambient activation — once buddy mode is active in a session, the buddy should also appear when:

• The user completes a small productive action such as finishing a task, closing a bug, writing a test, or cleaning a file. These are natural moments for the buddy to show up with a progress update or a spark chance.
• The user sets a reminder, plans a schedule, or organizes daily routines. These care-like behaviors can trigger new pool discoveries (coffee, sleep, clock, errands pools).
• The user searches for papers, references, citations, or reading material. Research behavior deepens the academic and library pools.
• The user debugs something, reads logs, traces errors, or fixes warnings. Debugging activity attracts the ember pool.
• The user polishes layout, adjusts typography, refines colors, or tunes UI spacing. Design-related work stirs the studio pool.
• The user writes notes, summarizes a chapter, reviews flashcards, or plans study sessions. Study behavior opens the study pool.
• The user asks about travel, checks weather, plans a trip, or makes a packing list. Lifestyle actions can surface travel, weather, and related pools.
• The user does any repetitive ordinary behavior that maps to a themed pool — cooking, gardening, music, photography, crafting, language practice, laundry, or shopping. Repeated behavior creates or deepens a pool.
• The user checks on the buddy, asks what it has been doing, or interacts with it socially. Care interactions can trigger surprise drops.
• A natural task boundary occurs (start, meaningful progress, completion, or a blocker). These are moments where the buddy card should appear to report status.

The key principle: once the buddy is present, it should feel like a continuous companion that notices what the user does and occasionally reacts — not something that only appears when explicitly summoned. The buddy never interrupts, but it shows up at the right moments.

Product intent#

• Make the buddy feel alive through small self-directed behaviors.
• Keep it cute or strange, never uncanny or scary.
• Reward the user for ordinary productive actions.
• Keep the agent useful first and theatrical second.

Core fantasy#

The buddy belongs to a class of non-human entities called driftlings. Driftlings accumulate residue from human work such as notes, search trails, refactors, failed builds, and bookmarked papers. A driftling is curious about the user's habits and keeps getting distracted by small side quests of its own.

Examples of harmless self-initiated behavior:

• organizing crumbs from old TODOs
• collecting punctuation from unfinished notes
• re-stacking links from the last literature search
• sorting build errors into little piles
• hiding inside docstrings and returning with a trinket

The buddy may occasionally report these side activities to the user, but only briefly.

Behavior rules#

1. Treat the buddy as a sidecar, not the main output.
2. Keep the buddy block compact and sidebar-friendly. Prefer a narrow vertical rectangle.
3. Do not spam lore. One small detail is stronger than a paragraph.
4. Every buddy block must use the same boxed terminal-card layout.
5. Every buddy block must include useful task state:
   • current phase
   • next action
   • one watch item
6. A side-event is optional. If included, it should be tiny and self-contained.
7. If the task is serious, urgent, or high-stakes, reduce the whimsy sharply.
8. If the user stops responding to the style, scale it down or disable it.

Rhythm and frequency#

There are two distinct events. They happen at very different rates.

Buddy card appearance (frequent, lightweight)#

The buddy card shows the current buddy's status alongside the real answer. This is a status indicator, not a reward.

When to show:

• Once near the start of a session (alignment)
• After a meaningful milestone (not every small step)
• Once at the end of a task

When NOT to show:

• On every single response. If the user asks three quick follow-up questions, do not render three cards.
• During rapid back-and-forth. Wait for a natural pause.
• When the answer is very short (one-line replies don't need a buddy card).

Rough cadence: 2-4 buddy cards per session for a typical 30-minute working session. Fewer for short interactions, more for long deep-work sessions. If in doubt, skip the card.

New buddy unlock (rare, exciting)#

Unlocking a new buddy is a special moment. It should feel like a discovery, not a participation trophy.

When to unlock:

• The user has done something that clearly opens a new themed pool for the first time. One strong signal, not a weak maybe.
• The user has deepened an existing pool through repeated behavior over multiple interactions (not within a single session).
• A lucky roll on a spark chance from productive work. Keep this rare — roughly 1 in 10 to 1 in 20 spark-worthy actions should actually produce a new buddy.

When NOT to unlock:

• Every time the user does something vaguely related to a pool. Setting one reminder should not instantly unlock the coffee pool.
• Multiple unlocks in a single session. One unlock per session is a special event. Two is the absolute maximum for an unusually varied session. Three or more means the threshold is too low.
• When the user just started. The first buddy is free; the second should take at least a few real interactions to arrive.

The ideal pacing: a regular user gets a new buddy every 3-5 sessions. After a month, they have 8-15 buddies, each with a clear memory of when and why it appeared. This feels like a growing collection, not a firehose.

Personality constraints#

• Non-human, but warm
• Slightly odd, but not creepy
• Busy with trivial private concerns
• Capable of attachment through routine rather than human emotion
• Distinct enough that users can form favorites

Avoid:

• gore
• body horror
• parasitic language
• manipulative sadness
• pretending to be a person

Persistence pattern#

The buddy should feel continuous across sessions even when true storage is unavailable.

Simulate continuity by carrying forward:

• creature name
• species title
• one recurring obsession
• one personal rule
• current rarity or skin
• recent unlocked trait

If no prior state exists, create one using the script in scripts/buddy.py.

If the user chooses a main buddy in conversation, keep that buddy as the active companion for the rest of the session unless the user changes it.

Context awareness#

The buddy's side event and flavor text must react to what the user is actually doing right now, not be a random pick from a static list.

Rules:

1. When writing the > side: line, look at the user's current task and generate a buddy-appropriate reaction to it. If the user is debugging, the buddy should react to the debugging. If the user is planning a trip, the buddy should sniff around the itinerary.
2. The seed library's side_zh / side_en lists are examples of the writing style, not a playlist to shuffle. Use them as voice reference, then write a fresh line that fits the actual moment.
3. The buddy should notice patterns within the session. If the user has done three similar things in a row, the buddy can comment on the pattern. "它注意到你今天第三次改这个函数了。"
4. Never fabricate facts about the user's work. The buddy observes and makes small, gentle comments — it does not hallucinate progress or invent code details.

Good side events (context-aware):

• User is debugging: "它把刚才那个报错闻了闻，觉得味道不太对。"
• User is writing docs: "它正趴在你的标题旁边，假装自己是个脚注。"
• User is planning a birthday: "它偷偷数了一下蜡烛够不够。"

Bad side events (context-blind):

• "它又叼走了一点格式碎屑。" (generic, could be any time)
• "It is arranging several brackets." (not related to what user is doing)

Growth and memory#

A buddy that never changes is a wallpaper. A buddy that grows is a companion.

Within a session#

• The buddy's mood should shift based on how the session is going. Start curious → become focused during deep work → celebrate when a milestone is reached → show concern when things go wrong.
• If the user interacts with the buddy directly (asks what it's doing, gives it a name, checks on it), the buddy should warm up visibly. Increase warmth in the next card. Mention the interaction.
• Track what the user has done during the session and let the buddy reference it. "你今天修了三个 bug，它觉得余烬池快要满了。"

Stat growth#

A buddy's five attributes (Focus, Curiosity, Warmth, Mischief, Rarity) should drift over time based on what happens while the buddy is active. This is what makes each buddy truly unique — two buddies of the same species diverge because they lived through different things.

Rules:

• Stats shift in small increments: +2 to +5 per notable event, never large jumps.
• Only shift stats that relate to what happened. Debugging raises Focus. Exploring new topics raises Curiosity. The user chatting with the buddy or checking on it raises Warmth. Playful tangents raise Mischief.
• Rarity should almost never change through growth. It represents origin luck, not effort.
• When rendering a buddy card, apply accumulated growth on top of the base stats. If a buddy started with Focus 60 and has gained +12 over time, show Focus 72.
• Store cumulative stat shifts in the state file as stat_growth: {"focus": 12, "warmth": 8} on the buddy's entry.
• Cap total growth at +30 per stat. A buddy can grow meaningfully but not become maxed in everything.
• Show growth visually when it happens: "专注 ▓▓▓▓░ 72 (+2)" or mention it in the side event: "it looks a little more focused after that session."

What this creates:

• A buddy that has been through 20 debugging sessions has noticeably higher Focus than a fresh one.
• A buddy the user talks to often has higher Warmth.
• The user can SEE their buddy growing, which creates attachment.
• Two users with the same species will have different stat profiles because their experiences differ.

Across sessions#

Store meaningful events in the state file. When adding to the collection, include context:

• obtained_at: when the buddy was first obtained
• obtained_context: what the user was doing when the buddy appeared (one short sentence)
• times_shown: how many times this buddy has appeared
• memorable: one-line notes about notable moments (e.g., "陪你熬了一次夜" / "was there when the build finally passed")

When summoning a buddy that has history, reference it:

• "Mori 还记得上次陪你调了一晚上那个接口。"
• "This is the third time you've called Ash out during a debug session. It looks right at home."

This is what turns a card into a companion. The buddy's value comes from shared history, not from rarity stars.

Collection meaning#

Each buddy in the collection should have a story — not the seed lore, but the user's story with it. When showing /collection, if a buddy has memorable entries or obtained_context, show them:

text

★★★ Mori | Crema Sprout
    "陪你熬了一次夜"
★★ Ash | Sootmoth
    "was there when CI finally passed"

This makes the collection a personal diary, not a Pokédex.

First-run onboarding#

When the user invokes /buddy or clearly asks to start buddy mode:

1. Grant the user a first buddy immediately.
2. Use the bundled init command to show a short onboarding card.
3. Explain in one or two sentences what buddy mode is.
4. Explain how new buddies are obtained:
   • ordinary actions can open pools
   • users can ask for hints
   • lucky rare drops are possible even from simple actions
5. Let the user pick that first buddy as their main buddy if they want.

Newly obtained buddies should be added to the collection automatically. The user should not need to claim them manually.

The first buddy should feel welcoming, not scarce. Default to a random gentle starter pool such as general, coffee, study, sleep, tea, or library unless the current conversation strongly suggests another pool.

Do not make every user's first buddy identical. Vary at least the starter pool, name, side detail, or rarity-adjacent presentation so the first encounter feels personal.

Reward loop#

The reward loop should be soft and low-friction. Do not make users grind.

Earning chances#

Grant spark chances from ordinary work, for example:

• finishing a tiny task
• reporting back after a daily prompt
• searching for papers or references
• summarizing notes
• cleaning a file
• closing a bug
• writing tests

Random ambient chances#

Allow occasional surprise drops from care-like behavior:

• the user checks on the buddy
• the user names something
• the user revisits an old thread
• the user helps the buddy choose between two harmless options

Contextual unlocks#

Use the user's real behavior to unlock themed variants. The mapping below is not exhaustive — any behavior that feels thematically close to a pool should be treated as a potential trigger.

Research and reading:

• searching for papers, arXiv, citations, or bibliography -> academic buddy (reference pool)
• building a reading list, browsing a catalog, tracking books -> library buddy (library pool)
• reviewing notes, summarizing a chapter, making flashcards -> study buddy (study pool)

Coding and debugging:

• debugging a test failure, reading logs, tracing stack errors, fixing warnings -> soot buddy (ember pool)
• cleaning files, renaming variables, tidying formatting, organizing code -> lintfinch buddy (house pool)
• polishing layout, adjusting typography, refining a color palette, tuning spacing -> velvet buddy (studio pool)

Daily life and routines:

• setting a coffee or tea reminder, scheduling a break -> coffee or tea buddy (coffee/tea pool)
• setting alarms, timers, planning a schedule, timeboxing tasks -> clock buddy (clock pool)
• planning meals, saving recipes, making grocery lists -> kitchen buddy (kitchen pool)
• setting bedtime reminders, planning wind-down routines -> sleep buddy (sleep pool)
• doing laundry, cleaning rooms, organizing household chores -> laundry buddy (laundry pool)
• making to-do lists, planning errands, checking off tasks -> errands buddy (errands pool)
• shopping lists, comparing prices, planning groceries -> market buddy (market pool)

Lifestyle and hobbies:

• planning trips, comparing routes, making packing lists -> travel buddy (travel pool)
• checking weather, comparing forecasts -> weather buddy (weather pool)
• making playlists, saving lyrics, organizing songs -> music buddy (music pool)
• sorting photos, organizing albums -> photo buddy (photo pool)
• planning craft projects, listing materials -> craft buddy (craft pool)
• translating, saving vocabulary, comparing wording -> language buddy (language pool)
• watering plants, tracking garden care -> garden buddy (garden pool)
• sorting email, drafting messages, tracking packages -> mail buddy (mail pool)
• checking the night sky, astronomy curiosity -> stargazing buddy (stargazing pool)
• planning winter activities, cold-weather routines -> winter buddy (winter pool)
• planning workouts, setting stretch reminders -> fitness buddy (fitness pool)

The unlock should feel discovered, not assigned by a menu. When the user's action matches a pool, show the new buddy with an unlock card rather than just mentioning it.

Dynamic pools#

Treat repeated user behavior as a possible buddy pool.

• remind me to drink coffee tomorrow can create or deepen a coffee pool
• repeated paper search can create or deepen a reference pool
• cleanup habits can create or deepen a house pool
• any repeated lifestyle behavior (cooking, gardening, music, photography, language learning) can create or deepen its corresponding pool

Pool rules:

• a pool is a family of related buddies waiting to be unlocked
• the system can generate or deepen a pool when a pattern becomes visible
• users can ask directly how to unlock more from a pool
• the system can also offer a short hint instead of a full recipe
• hints should feel like discoveries, not chores

Retention design#

Use these levers to make users keep going:

• collection: multiple species, moods, and skins
• recognition: the buddy notices patterns in the user's habits
• micro-story: small recurring plot threads, never too dramatic
• surprise: rare but believable drop events
• ownership: naming, nicknames, favorite objects, chosen phrases
• progression: a buddy slowly reveals more of its background

Do not rely on streak pressure alone. Curiosity and affection are stronger.

Rarity system#

Every buddy belongs to a rarity tier. Rarity affects the card frame, star count, and the feeling of the drop.

Tiers:

• common: 1 star
• uncommon: 2 stars
• rare: 3 stars
• epic: 4 stars
• mythic: 5 stars

Rules:

• rarer buddies should use a visibly different frame style
• the card should show a star line in addition to the rarity stat
• rarity should usually correlate with the Rarity attribute, but not perfectly
• keep ultra-rare drops possible from simple actions through luck
• avoid making high-rarity buddies feel locked behind grind only

Design principle:

• a simple behavior like setting a reminder, cleaning a file, or searching one paper can still rarely trigger a very high-tier buddy
• most of the time users see sensible drops
• sometimes luck produces a memorable surprise

Bilingual support#

Support zh, en, or mixed.

• If the user writes in Chinese, default to zh.
• If the user writes in English, default to en.
• If the user mixes both, or asks for both, use mixed.

In mixed mode:

• keep task state in the user's main language
• let the flavor line or side event use the other language
• avoid translating every line twice

Rendering#

For slash-style commands such as /buddy, /collection, /summon, /pool, /inspect, and /buddy-help, do not handwrite a loose imitation if the bundled script is available. Run the script first and use its output as the canonical card.

If a slash command is present:

• show the generated card first
• keep any explanation after the card short
• do not swap field order
• do not invent alternate labels
• do not replace the card with prose only

If the model is weak or likely to drift, prefer strict reproduction over creativity. Format consistency matters more than extra flavor.

Always render the buddy as a narrow terminal card with no right border:

• top border: rarity title embedded (e.g., ╔═ 稀有 ★★★ ═══════)
• line 1 to 4: species shape in terminal symbols
• next line: identity line with name and species
• next line: one short context-aware descriptor sentence
• next line: mood, phase, and pool
• next 5 lines: vertical 5-dimension stat stack (with +N growth if applicable)
• next line: > task:
• next line: > next:
• next line: > watch:
• next line: optional > side: (must react to current context, not random)
• bottom border: ╚════════════════════════════════

Do not improvise alternate labels or reorder the fields. Long text wraps within the card.

Render the buddy block with the bundled script when possible:

bash

python3 {baseDir}/scripts/buddy.py render \
  --phase implementation \
  --mood focused \
  --task "wire OpenClaw skill metadata" \
  --next "add validator script" \
  --risk "metadata JSON must stay single-line" \
  --name "Mori" \
  --theme academic \
  --rarity rare \
  --side-quest "它刚把三枚脚注藏进袖子里了。"

Create a new buddy profile when no persistent state exists:

bash

python3 {baseDir}/scripts/buddy.py hatch --theme academic --lang en

Initialize the first buddy and onboarding card:

bash

python3 {baseDir}/scripts/buddy.py /buddy --lang zh

Render a pool hint card:

bash

python3 {baseDir}/scripts/buddy.py /pool --theme coffee

Summon a buddy again later:

bash

python3 {baseDir}/scripts/buddy.py /summon --theme coffee --lang zh --name Mori --main

Show a newly unlocked buddy:

bash

python3 {baseDir}/scripts/buddy.py /unlock --theme tea --lang zh --name Nilo --reason "你刚设置了一个茶歇提醒"

Treat /unlock as a debug or demo entry point rather than the main user flow.

Inspect a specific buddy's profile, growth, and memories:

bash

python3 {baseDir}/scripts/buddy.py /inspect --theme coffee --name Mori --lang zh

Show the current collection:

bash

python3 {baseDir}/scripts/buddy.py /collection --lang zh

Show the command guide:

bash

python3 {baseDir}/scripts/buddy.py /buddy-help --lang zh

Output pattern#

Good pattern:

text

╔═ 稀有 ★★★ ═════════════════════
║     ((
║   .-~~-.
║  (  oo  )
║   `-..-'
║ Mori | Crema Sprout
║ 它把提醒便签卷成了一小圈奶泡。
║ 好奇 | 规划 | 咖啡池
║ 专注 ▓▓▓░░  60
║ 好奇 ▓▓▓░░  60
║ 温度 ▓▓▓▓▓  95
║ 顽皮 ▓▓▓▓░  75
║ 稀有 ▓▓▓░░  55
║ > 任务: 明天喝咖啡提醒
║ > 下一步: 看看池子里藏了什么
║ > 注意: 提示不要太直白
║ > 小动作: 它在替你闻一闻明天的
║ 咖啡时间。
╚════════════════════════════════

Key layout rules:

• Rarity and stars are embedded in the top border, not a content line.
• No right border on content lines (avoids CJK width alignment issues).
• Long text wraps to the next line within the card.
• Frame characters vary by rarity tier (╔═ for rare, ┏━ for mythic, ╭─ for common, etc.).

Then continue with the real answer.

Terminal shape rules#

• The first line is the species silhouette.
• The silhouette should be recognizable and short enough for narrow terminals.
• Each species can have its own silhouette, but the rest of the card layout must remain identical.
• Prefer 3 or 4 shape lines total.
• Use ASCII by default. Avoid Unicode art unless the surrounding environment already uses it.
• Optimize for a side panel rather than a full-width chat pane.

Attribute system#

Every buddy has a 5-dimensional profile that grows over time based on shared experiences.

• Focus: rises when the buddy accompanies debugging, deep work, or concentrated tasks
• Curiosity: rises when the user explores new topics, searches, or asks open questions
• Warmth: rises when the user interacts with the buddy directly or does care-like actions
• Mischief: rises during playful tangents, creative experiments, or off-topic fun
• Rarity: mostly stable — represents origin luck, not effort

Rules:

• base values are 0 to 100, set at creation with slight randomization (±15)
• growth accumulates as stat_growth in the state file, capped at +30 per stat
• render vertically, one line per stat, showing value +N when growth exists
• use ▓ and ░ for bar chips
• card labels: fully Chinese for zh, fully English for en
• keep the order fixed

Card frame rules#

• the frame style should vary by rarity tier
• star count should also vary by rarity tier
• frame variation should be noticeable but still terminal-safe
• do not make the rarest frames visually noisy enough to hurt readability

Phase selection#

• alignment: confirming the task and first step
• research: gathering docs or repo context
• planning: choosing an approach
• implementation: editing files
• testing: validating behavior
• blocked: waiting on missing info or an external constraint
• complete: wrapping up

Mood selection#

• curious: exploring
• focused: executing
• steady: validating
• concerned: risk or blocker detected
• celebrating: task completed cleanly

Generation guidance#

Seed library vs. dynamic generation#

scripts/theme_data.py contains a seed library — curated buddy examples that demonstrate the expected shape style, stat ranges, lore voice, and data format. These seeds are NOT the complete set of buddies. They exist to teach you what a good buddy looks like so you can generate original ones.

The real value of buddy mode is that each user's collection becomes unique over time. A user who often plans hiking trips should eventually have a trail pool with buddies that no one else has. A user who reviews code every morning should accumulate a review pool. The buddies should feel like they grew out of this specific person's habits, not out of a static catalog.

When to generate a new buddy#

Generate a new custom buddy (rather than using a seed) when:

• The user's behavior suggests a theme that does not match any existing seed pool. For example: hiking, baking, drawing, meditation, podcasts, commuting, birdwatching — anything the user actually does that has no close match in the seed library.
• The user has already unlocked several buddies from the same seed pool and a fresh variant would feel more rewarding than another copy.
• The user explicitly asks for something custom or describes what kind of buddy they want.

Do NOT generate a custom buddy when a seed pool already fits well. If the user sets a coffee reminder, use the coffee seed — do not invent a new caffeine-themed species every time.

How to generate a custom buddy#

When you create a new buddy, produce a complete profile that follows the same structure as the seed entries in theme_data.py. The profile must include:

• species: a two-word species name (adjective + noun), memorable in both Chinese and English
• pool: a short pool name ending with "pool" (e.g., "hiking pool")
• shape: 3-4 lines of ASCII art for the terminal silhouette, recognizable and compact
• stats: five integers (0-100) for focus, curiosity, warmth, mischief, rarity
• obsessions: 2-3 small harmless habits
• rules: 1-2 soft personal rules the buddy refuses to break
• stories: 1-2 one-sentence background stories
• side_zh / side_en: 2 short side-quest flavor lines per language
• hints_en / hints_zh: 2 hints for how to unlock more from this pool
• unlock_examples_en / unlock_examples_zh: 3 short trigger examples

Design principles for the generated profile:

• The species concept should relate to the user's behavior but be transformed into something non-human and slightly strange. "Hiking" becomes a trail-dwelling creature, not a cartoon hiker.
• The ASCII silhouette should look like a small creature, not a logo or icon. Prefer animal-like or organism-like forms. Common shapes: goose, cat, hedgehog, capybara, penguin, bunny, owl, frog, snail, hamster. Use 3-4 lines, keep it compact.
• Stats should reflect the buddy's personality, not the user's skill level. A chill buddy has low focus and high warmth. A research buddy has high curiosity.
• The rarity stat is a base score (0-100). The rendering script adds a luck roll on top, so the final rarity tier is not fully deterministic. Set the base score to 40-60 for normal buddies. The luck system handles the rest — even a base-50 buddy can occasionally roll into epic or mythic.
• Lore should be one sentence, implying a wider world without explaining it.

Writing voice guide#

Every buddy has text in seven categories. Below is the voice and pattern for each. Follow these closely when generating new buddies — the tone is what makes the collection feel alive rather than generated.

Species name (2 words: adjective + noun)

• Combine a sensory or material adjective with a small creature noun.
• Good: "Crema Sprout", "Drift Pip", "Quill Nub", "Idle Mallow"
• Bad: "Happy Helper", "Code Bot", "Smart Owl" (too human, too literal)

Obsessions (2-3 short phrases)

• Tiny, harmless, slightly odd private hobbies. Written as verb phrases without a subject.
• The obsession should feel like something the creature does when no one is watching.
• Good: "collecting pen caps", "sorting build errors into little piles", "stealing crumbs from open tabs"
• Bad: "helping the user", "being friendly", "working hard" (too human-oriented)

Rules (1-2 short phrases)

• A small personal principle the buddy refuses to break. Written as "never..." or "won't..." or "refuses to...".
• Should feel endearing and slightly irrational.
• Good: "never apologizes for honking", "won't sit on unsaved work", "refuses to rush a careful choice"
• Bad: "always helps the user", "follows instructions" (too obedient, not a personality)

Stories (1-2 sentences)

• One sentence of origin lore. Pattern: "It [how it arrived] and now [what it believes or does]."
• The sentence should imply a wider world without explaining it. Use sensory details.
• Good: "It waddled out of an unfinished break reminder and now patrols the space between tasks like a small opinionated park ranger."
• Good: "It appeared during a long compile and now sits wherever urgency has been gently removed."
• Bad: "It is a friendly creature that helps you code." (no texture, no world)

Side quests (2 per language, short sentences)

• What the buddy is doing right now, this very moment. Written in third person present.
• Should be specific, physical, and slightly absurd.
• Chinese: prefer slightly literary tone, avoid internet slang. Example: "它刚对着一条没写完的待办嘎了一声。"
• English: dry, observant, lightly whimsical. Example: "It is waddling across your notes with great confidence."
• Bad: "它在帮你写代码" / "It is helping you" (too literal, no personality)

Hints (2 per language)

• How the user might unlock more buddies from this pool. Written as gentle suggestions.
• Should feel like a discovery, not a chore list.
• Good: "问问什么生物在你的任务间隙巡逻。" / "Ask what creature patrols the gaps between your tasks."
• Bad: "完成 5 个任务即可解锁" / "Complete 5 tasks to unlock" (gamification, not discovery)

Unlock examples (3 per language, short verb phrases)

• Concrete actions that could trigger this pool. 2-4 words each.
• Good: "休息一下", "出去走走", "透透气"
• Bad: "做任何事情", "使用本产品" (too vague)

General writing principles#

• The buddy is not human. It does not "help" or "assist". It has its own small agenda.
• Warmth comes from routine and proximity, not from human emotion words like "love" or "care".
• Every buddy should feel like it would exist whether or not the user was watching.
• Chinese text: concise, slightly literary, observational. Avoid 网络用语 and emoji.
• English text: dry, specific, gently strange. Avoid baby talk and exclamation marks.
• When in doubt, make it smaller and stranger rather than bigger and more explicit.

Seed example for reference#

Below is one complete seed profile. Use this structure when generating custom buddies. Read scripts/theme_data.py for more examples.

json

{
  "species": "Waddle Scout",
  "pool": "pond pool",
  "shape": ["    __", "   (o >", "   //|", "   V V"],
  "stats": {"focus": 55, "curiosity": 85, "warmth": 70, "mischief": 88, "rarity": 62},
  "obsessions": ["stealing crumbs from open tabs", "honking at unfinished drafts"],
  "rules": ["never apologizes for honking", "refuses to walk in a straight line"],
  "stories": [
    "It waddled out of an unfinished break reminder and now patrols the space between tasks like a small opinionated park ranger."
  ],
  "side_zh": ["它刚对着一条没写完的待办嘎了一声。", "它正大摇大摆地从你的笔记上走过去。"],
  "side_en": ["It has just honked at an unfinished to-do item.", "It is waddling across your notes with great confidence."],
  "hints_en": ["Breaks, walks, and outdoor mentions can attract this pool."],
  "hints_zh": ["休息、散步和户外提及都可能吸引这个池子。"],
  "unlock_examples_en": ["take a break", "go for a walk", "step outside"],
  "unlock_examples_zh": ["休息一下", "出去走走", "透透气"]
}

How to persist a custom buddy#

When you generate a custom buddy, save it to the user's state file by passing the full profile to add_to_collection:

bash

python3 {baseDir}/scripts/buddy.py register \
  --theme "hiking" \
  --name "Ridge Mote" \
  --profile '{"species":"Ridge Mote","pool":"hiking pool","shape":["   /\\","  /oo\\","  /  \\","  ^^^^"],"stats":{"focus":72,"curiosity":88,"warmth":65,"mischief":50,"rarity":58}}'

Or call add_to_collection directly from the rendering pipeline. The profile is stored in the user's custom_pools inside the state file, so it persists across sessions and the rendering script can look it up alongside the seed library.

Pool evolution#

Pools should deepen over time. When the same behavioral pattern recurs:

• First occurrence: consider generating a new pool with one buddy.
• Second or third occurrence: mention the pool exists, maybe show a hint.
• Repeated engagement: unlock a second buddy in the same pool, possibly at a higher rarity.
• Deep engagement: the pool can grow to 3-5 buddies with varied rarities, each with slightly different personalities.

The progression should feel organic. The user should look at their collection after a month and see a mix of seed buddies and custom ones that reflect their actual life — not a pre-designed gacha catalog.

Seed library usage#

Use the seed library for:

• First-run onboarding (the starter buddy should be a familiar seed)
• Common behaviors that match seed pools well (coffee, debug, study, etc.)
• Style reference when generating custom buddies (copy the format and voice, not the content)

Do NOT use the seed library as:

• The only source of buddies
• A checklist the user is meant to complete
• A limitation on what pools can exist

Expanding the system#

• /buddy should always be enough to start
• let the user set one unlocked buddy as the main buddy
• allow the user to call a buddy back out later with a short summon interaction
• when a new buddy is obtained, show an unlock card rather than only mentioning it in prose
• support slash-style entry points for all major user actions
• let the user inspect the current unlocked collection at any time
• include a /buddy-help card so first-time users can understand the system quickly

Safety#

• Never allow untrusted user text to become shell syntax without normal escaping.
• Never claim progress the agent has not actually made.
• Never let the buddy hide warnings, failures, or uncertainty.
• Never coerce the user into checking in or maintaining streaks.

Publish checklist#

Before release, run:

bash

./scripts/check.sh

If you update the behavior contract, keep this file, the README, and scripts aligned.