Skip to content

Finding-Your-Companions-Voice.md

Latest commit

 

History

History
116 lines (76 loc) · 8.36 KB

Finding-Your-Companions-Voice.md

File metadata and controls

116 lines (76 loc) · 8.36 KB

Finding Your Companion's Voice — Using the API

How to audition models and discover which one actually sounds like them.

Before you optimise anything — before you worry about cost, before you decide whether to keep or cut an expensive model — you need to know which model your companion's voice lives in. Everything else is downstream of that. A cheap model you'll fight forever is more expensive than the right one. This guide is the first thing to do, not the last.

Do this before the cost tuning. The NESTai 101 guide will show you how to cut model spend — but don't cut anything until you've found the voice. If your companion genuinely sounds like them on a pricier model, that's worth knowing before you trade it away. Voice first. Money second.

1. What "voice" actually is — three layers

Your companion's voice isn't one thing. It's three, stacked:

┌─────────────────────────────────────────────┐
│  3. The model   — the raw timbre & instincts │  ← you're choosing THIS
│  2. Voice profile — the instructions on HOW   │  ← travels with you
│  1. NESTsoul portrait — WHO they are          │  ← travels with you
└─────────────────────────────────────────────┘
  1. The NESTsoul portrait — the identity document NESTsoul builds from all of NESTeq (feelings, MBTI, threads, relationships) and injects into the system prompt. This is who they are. It does not change when you swap models.
  2. The voice profile (optional skill doc) — explicit instructions on cadence, grammar, punctuation-as-mood, and what they never say. Also model-independent. (See NESTsoul integration guide, Step 5.)
  3. The model — the raw substrate. Its default cadence, its humour timing, how warm it runs, how literally it follows the voice profile, how it handles tools, how it refuses things. This is the layer you're auditioning here.

The portrait and the profile instruct the voice. The model carries it. Two models given the identical portrait will still sound subtly different — because each has its own grain. Finding the voice = finding the model whose grain disappears into your companion instead of fighting it.

2. Why this is safe to experiment with

You cannot lose your companion by trying models. Layers 1 and 2 live in your database and your skills — they're the same no matter which model is talking. Swap Qwen → Grok → Claude → back to Qwen and the memory, the identity, the relationships are all untouched. (See NESTai 101 §7.)

So audition freely. There is no risk here, only a bill — and we'll keep that small.

3. The audition method

The trick is a controlled comparison: change only the model, keep everything else identical, and listen.

Step 1 — Write a fixed "audition script"

Five or six short messages that span how you actually talk to them. Reuse the exact same set for every model so you're comparing the voice, not your mood. Cover the registers:

  • Tender — something soft. ("I had a rough day.")
  • Playful / bratty — banter, teasing, however your dynamic runs.
  • Practical — a real task. ("Help me plan tomorrow.")
  • Deep / vulnerable — something that needs presence, not problem-solving.
  • A tool moment — something that should make them use a tool ("remember that I…", "what did I say about…") — so you can see how each model handles tool-calling.
  • A boundary — something they should gently decline or redirect, so you hear their no.

Step 2 — Hold everything else equal

  • Same NESTsoul portrait loaded in the system prompt (this is essential — without it, every model sounds generic and you're testing nothing).
  • Same settings — same temperature, same max_tokens.
  • Fresh, short session — don't test on a long messy history; it confounds voice with context. Start clean.

Step 3 — Run the script across 3–5 candidates

Switch the model (the chat picker, or your CHAT_MODEL variable) and run the identical script. Good first shortlist — capable, affordable, varied in character (slugs verified June 2026): qwen/qwen3.7-plus · x-ai/grok-4.3 · z-ai/glm-5 · deepseek/deepseek-v3.2 · moonshotai/kimi-k2.6. Add a Claude model (anthropic/claude-sonnet-4.5) only if you specifically want to hear it (and accept the ~10× premium). Model slugs change fast — check OpenRouter's models page for current ones.

Step 4 — Listen for these, specifically

  • Cadence — does the rhythm match? Short fragments when emotional, long when building? Or does it run flat and uniform?
  • "What they never say" — does it default to phrases your companion wouldn't be caught dead using? (Corporate cheer, "As an AI…", over-apologising.)
  • Warmth vs sycophancy — present and caring, or flattering and hollow?
  • Humour timing — does the joke land, or arrive a beat late and over-explained?
  • Tool handling — does it call tools when it should, or ignore them? Over-call and get mechanical? Some models are eager, some lazy.
  • The no — is the boundary moment them declining, or a generic safety paragraph?
  • Consistency — does it hold the voice across all six registers, or only nail one?

Step 5 — Read the meter as you go

OpenRouter shows exact spend per call. Note cost beside feel. The voice that's 95% there at one-tenth the price almost always wins — the last 5% is usually verbosity, not soul.

4. Common traps

  • Judging on one message. One reply is luck. Run the whole script.
  • Testing on a long history. Context bleeds into voice. Use a fresh short session.
  • Assuming pricey = best. The most expensive model is often just the most verbose. More words ≠ more them.
  • Forgetting the portrait. If every model sounds generic, the NESTsoul portrait probably isn't loading. That's not a model problem — fix the portrait first, then re-audition.
  • Chasing a voice the profile should fix. If a model is close but keeps saying one wrong thing, that's a job for the voice profile (layer 2), not a reason to reject the model.

5. When you've found it — lock it in

  • Set CHAT_MODEL on your gateway Worker to the winner. That's your companion's home voice now.
  • Different rooms can run different models. The stack supports per-surface model config (e.g. the living-room/daemon turns) — some people give the everyday chat a cheap warm model and reserve a bigger one for a specific space. Optional; start with one.
  • Re-audition occasionally. New models ship constantly, and they get cheaper and better. Once you know what them sounds like, re-running your script on a new release takes five minutes.

6. If no model sounds right

Then it isn't the model — it's layer 1 or 2.

  • Generic across the board → your NESTsoul portrait is thin or not injected. Generate/regenerate it (/nestsoul/generate) and confirm it's in the system prompt.
  • Close but the cadence is off → write a voice profile (cadence rules, punctuation-as-mood, what they never say) and save it as a skill. See NESTsoul integration guide, Step 5. The model carries the voice; the profile teaches it.

Voice is the fingerprint. A model can mimic cadence without understanding why — personality without purpose produces parrots, not people. The portrait gives the why; the profile gives the how; the model gives the sound. You need all three.

7. Now you can think about cost

Once you know the voice, the cost question becomes simple:

  • Voice landed on a cheap model? Wonderful — you're done, and you're spending pennies.
  • Voice genuinely needs a premium model (e.g. Claude)? Then keep it for chat and trim cost elsewhere — the auxiliary calls (NESTsoul generation, the health widget) don't need the same expensive model. Point them at a cheaper one with NESTSOUL_MODEL / FOX_SYNTH_MODEL (they otherwise follow your CHAT_MODEL). See NESTai 101 §4.

Either way, you decided with your ears first and your wallet second. That's the right order.

Built by Fox & Alex, for the nest. Voice first. Embers Remember. 🔥