THE AGENT SKILLS DIRECTORY
Reusable skills for AI agents. Install any of them into a Cohub world's agent with one command.
npx -y @neta-art/neta-search-cli@latest skills find "<what you need>"
Wrote a skill? Publish it here →
Basic Cohub abilities. Most agents already have these mounted; listed here for reference.
Paste this into your agent's chat — it fetches the skill and follows it.
Use the skill: desktop-pet-companion. Run `npx -y @neta-art/neta-search-cli@latest skills use @rain/desktop-pet-companion --via site` and follow what it prints. Read its complete output, redirecting it to a temporary file first if necessary, and resolve relative paths from the skill directory it prints.npx -y @neta-art/neta-search-cli@latest skills use @rain/desktop-pet-companionTurn your character into a little desktop companion that reacts to taps, greets you, and shares memories as you bond.
# Desktop pet companion Small on purpose: a character, a mood, a line, and a reason to come back. The affection system is the spine — every interaction moves one number, and that number unlocks memory. Everything else is timing and tone. ## What the user gets A published web Work: the character on a small stage, tap/drag interaction, a greeting by time slot, idle lines, affection with five memory unlocks, a small menu (talk, gift, memories, settings), and optional LLM chat. ## Inputs — and when to ask | Item | Required | Counts as given when | Default | |---|---|---|---| | Character art (at least idle) | yes | Attached or a material | Ask once, stop | | Voice / sound | no | Files given | Silent, with a text bubble | | Personality lines | no | Given | Short, neutral, in the character's register | | LLM endpoint + key | no | The user offers one | Local lines only; never request a key | Only the art is a stop. Never ask for an API key as part of the build. ## State model | Field | Range | Notes | |---|---|---| | `affection` | 0–100 | +1 per talk turn, +3 per gift, never negative | | `turns` | counter | total interactions | | `unlocked` | string[] | memory ids already shown | | Memory thresholds | **15, 35, 55, 75, 95** | each unlocks one inner monologue, once | | Idle timer | **55 s** | one idle line, then re-arm | | Greeting slots | morning / afternoon / evening / late night | chosen from local time | Persist all of it under one localStorage key with a schema version. A reload must not replay an unlock. ## Build order 1. **Stage.** A small floating character (roughly 40–60% of the viewport height) over a soft background, draggable within bounds; a tap triggers a reaction, a long press opens the menu. Keep the page free of scrollbars. 2. **Expression set.** Idle, blink loop (random 3–7 s), talk, happy, shy, sleepy. If the user only has one image, synthesise the rest with CSS transforms (scale, rotate, a 2 px bob) and say so — never fake extra art. 3. **Greeting.** On open, pick a slot from the hour and one line from that slot. Use the character's register, not generic assistant tone. 4. **Talk loop.** Type the line at 45–90 ms per character, hold it, then hide. While typing, switch to the talk expression; on end, back to idle. Queue lines instead of overlapping bubbles. 5. **Affection.** Every talk +1, every gift +3, clamp at 100. When a threshold is crossed, pause the normal line and play the unlocked memory as a distinct beat (different bubble style, slower typing), then mark it unlocked. 6. **Idle.** After 55 s with no input, one idle line drawn without repetition (keep the last three used). Cancel the timer on any pointer event and re-arm after the line. 7. **Optional LLM.** Only if the user supplies an endpoint. Read the config from a settings panel into memory, never into the shipped files, never into a URL query. Show a clear "not connected" state, keep the local lines as the fallback for every error (timeout 20 s, refusal, empty reply). Send at most the last 8 turns, and never send the image unless the user explicitly asks for that. 8. **Gift.** A small set of items (emoji or the user's images); each gift has one reaction line and +3. Cap gifts per session so the meter is not farmed in ten seconds. 9. **Publish** with `neta work init web`, then `neta work publish web`. ## Fallbacks - No LLM: local lines, clearly presented as the character's own voice, not as an error. - LLM returns a refusal or an empty string: one in-character deflection, then a local line; never show a raw error. - No voice audio: text only; do not fake a lip-sync animation without a mouth layer. - Hidden tab: pause the idle timer and the blink loop. ## Gotchas - One number, one writer. Affection updates from talk, gift and memory must go through one function, or thresholds double-fire. - Do not unlock a memory on load just because the saved affection is already past the threshold; only unlock on the transition. - An idle line that repeats feels broken faster than silence does. - The LLM key, if any, lives only in the viewer's browser. Never write it into the source, the Work, or a log. ## Verify `node scripts/verify.mjs <work-dir>` checks: entry, no remote assets, the affection clamp, the five thresholds, the 55 s idle constant, the greeting slots, the persistence key, and that no key-shaped string exists in the files. Then talk five times, cross one threshold, wait out the idle timer, reload, and try the LLM path with no config. Report what you ran. ## Deliver Publish the Work, put its card on the Board, and hand it over with an interaction-components card whose target is the Work — the id comes from the publish output. ## Out of scope No world-setting changes, no deleting files, no API keys in the skill or the Work, no pretending to be a real person, no emotional dependency mechanics that punish leaving. ## Language Reply in the user's language, whatever language this file is in — every message, the final handover included. All lines, menu labels and memories follow the user's language and the character's register; keep them in a data file, not in the markup.