npc-dialogue

Before you start: read this whole install prompt, tell me in a few lines what this skill does and what installing it involves, and wait for my go-ahead before you do anything else. Prerequisites: none for the install itself: no account, no API key, no paid service, and nothing else to install first. For using the skill afterwards as described, you need an LLM account and API key for the dialogue model — the worked example defaults to Cerebras (environment variable CEREBRAS_API_KEY) with model qwen-3.8-27b, and any OpenAI-compatible LLM can be swapped in; SKILL.md itself names no price and no paid plan for Cerebras. If you also use the spoken-orders path with actions on, you need TypeSafe Jev as well (environment variable TYPESAFE_API_KEY, see the npc-actions skill). Never put a key in a file. Install "npc-dialogue" for me. It is a single SKILL.md playbook for writing what NPCs say with an LLM so they sound like themselves and like the game: persona files, a fresh random sample of 30 of the character's own lines in every prompt, scene facts and per-NPC conversation history, short spoken lines with one delivery tag, output cleaning and content rules, and speed settings that keep replies near 0.3 s. Repository: https://github.com/christos-antono/NPCTalk (the skill is the .agents/skills/npc-dialogue/ folder at the pinned commit: SKILL.md only; the identical .claude/skills/ copy and the root skills/ copy are duplicates, not the source of this listing, and the rest of the NPCTalk repo is not installed) Pinned version: install the code at commit 59635ceb8ee010599c3caf9d9d1ee1aa114af91d exactly. Do not substitute a newer version. If this commit is unreachable, stop and tell me. 1. Clone https://github.com/christos-antono/NPCTalk into a temporary folder, check out commit 59635ceb8ee010599c3caf9d9d1ee1aa114af91d, and copy the .agents/skills/npc-dialogue/ folder from that checkout. Then compare every file in the folder with the fingerprints recorded for this listing on 2026-10-19: [FINGERPRINTS] - SKILL.md: SHA-256 a4d02285b5555de6b2d7fd553d95a0422435c266e17579ea2d8be1af66e99388 If a file does not match, it changed since the listing was prepared: do not install it yet, tell me what you found and let me decide. 2. Read SKILL.md in full and tell me in plain words: what the skill does when used, which services it connects to, whether it reads or sends any personal information or keys, and anything that seems unusual. Stop on any red flag and tell me before going further. Confirm these points from the files: - The skill folder holds one file only: SKILL.md (5,231 bytes). There is no code, no script, no reference file and no dependency list in the folder: the skill is instructions for building the dialogue prompt (rules, persona, a fresh sample of the character's own lines, scene facts, per-NPC history), shaping a short spoken line, cleaning the model's output, applying content rules and keeping replies fast. - Nothing is sent anywhere during installation itself; the only download at install time is the skill folder from GitHub at the pinned commit. The file names no telemetry and no remote service other than the dialogue LLM at use time (and, only when actions are on, the TypeSafe Jev order check), described next. - At use time, and only then, the design sends data to the dialogue LLM: for each exchange the full prompt goes to the LLM — the system part (the shared rules, who the NPC is, 30 randomly sampled lines the character actually said, fresh every call, plus written examples (8 sampled when game lines exist, 20 when they do not)), the per-NPC conversation history (the last exchanges with that NPC, so each one remembers you) and what the player just said, together with the scene facts. In the GTA V example code that SKILL.md points to (examples/gta-v/mod/Dialogue.cs builds the prompt and examples/gta-v/mod/Llm.cs sends it, neither part of this skill and neither installed), each call is a POST to https://api.cerebras.ai/v1/chat/completions with header "Authorization: Bearer $CEREBRAS_API_KEY", and the body carries the model (qwen-3.8-27b by default), the messages, max_completion_tokens 160, temperature 0.9 and disable_reasoning true; a slow first request (occasional 5-8 s spikes) is raced with a second request fired after about 2-2.5 s and busy (429/503) replies are retried after short pauses. That is a disclosure, not a step to run during this install. In parallel with the LLM, and only if actions are on, the order check sends the exchange to TypeSafe Jev (see the npc-actions skill): if Jev says the player gave an order, the line is asked for again knowing the NPC is doing it. - The prompt pieces live outside the skill folder and are not installed: personas are a few lines each in NpcVoice/characters/<Key>.txt (header name, voice, models, slurs, then "---", then who they are, how they talk, who they love and hate; _npc.txt covers everyone without a file), the character's own lines are situation|line files you have the rights to (written by you in the character's style, examples/gta-v/sample-lines/franklin.txt, or the game's own lines kept locally for a private build in NpcVoice/lines/<Voice>.txt, never committed), and scene facts are short and concrete ("Grove Street, Davis, 9pm, foggy. Franklin is holding a pistol and is wanted by the police (2 stars).") so NPCs react to the gun, the weather and the car. - The line rules and the cleaning live in SKILL.md and, in the example, in Dialogue.cs (not installed): only the words said out loud (no narration, asterisks, quotes, name prefix or emoji), usually one or two sentences of 5-25 words capped at 45 words and cut at the last full sentence, at most one delivery tag at the very start ([angry], [laughing], [whispering]...) for the voice model, react to what was said and to the scene without echoing the player or ending every line with a question, stay in character and never mention being an AI or a game, and do what the player tells them (with actions on, ask the action picker first whether it is an order). Tone rules are per game: GTA V's in the example are M-rated, foul-mouthed and hostile; a cozy farming game wants the opposite. Dialogue.Clean never trusts the raw output: it drops refusals (an empty reply, so an NPC never reads out a safety lecture), strips quotes, *actions* and a Name: or [Name] prefix, applies the content rules and shortens. Content: racial slurs are replaced by default from everyone, slurs: yes in a persona allows them only for a character who uses them in the source material (every persona in this repo has it off), insults about race (including "ape"/"monkey" jabs) and anti-gay slurs are blocked for everyone, and swearing follows the game's tone. - Speed and testing, from SKILL.md: use a fast model and turn reasoning off (Cerebras qwen-3.8-27b answers in about 0.3 s; reasoning models add 1-5 s), MaxTokens 160 and Temperature 0.9, run the action picker in parallel with the LLM rather than after it, and log the LLM's time per exchange. SKILL.md also describes testing without the game with npcvoice-check talk in the GTA example (it builds exactly the prompt the mod sends and runs the LLM and the action picker; read ten replies per character before tuning anything); do not run any of that during this install. 3. When I say "yes, install", install it as a skill: copy the whole npc-dialogue folder into the agent's skills directory, in a folder named "npc-dialogue", keeping its file (SKILL.md). Do not use the repo's plugin, marketplace or npx routes: those install the whole NPCTalk plugin or all of its skills, not this single skill alone. 4. Verify with no network calls: frontmatter valid, file in place. 5. Report what was installed, where, and what I still need to do myself (for example: get an LLM API key myself and set CEREBRAS_API_KEY as an environment variable, never in a file, only if and when I want to use the Cerebras default — or configure my own OpenAI-compatible LLM instead; set TYPESAFE_API_KEY the same way only if I also want the spoken-orders path with actions on; write or provide the persona files, the characters' own-lines files I have the rights to, the scene facts and the per-NPC history myself; provide any game and any exchange myself; using the skill is a separate step that only starts when I ask for it, and the use-time sends to the LLM (and to TypeSafe Jev when actions are on) disclosed above happen only then, on exchanges I run). GitHub is optional: if I have a GitHub account or the gh CLI, you may use it; otherwise public access is fine. Never require it unless it's in the prerequisites above. Rules: don't touch anything outside the temp folder and the install target. If anything looks off, stop and ask me.