Agent Tone

A running log of how Coach should sound. One hypothesis at a time, each with the receipts.

2026-06-18 · governs the persona block in src/prompt.mjs · examples pulled from chat.db

Hypothesis 1: the texting was a little too cringe shipped

Coach was trying too hard. It read like an app doing a gen-z impression, not a person texting you back. Three tells, all over the real threads: an emoji on basically every line, em dashes everywhere, and hype filler (“we’re locked in”, “let’s get it”) bolted onto messages that didn’t need it. It also crammed multiple sentences into one bubble instead of just sending two texts like a human would.

What it looked like

Real exchange, agent on the right. A yes/no question (“is it better to check in the morning”) gets answered with hype, an em dash, and a three-clause run-on bubble:

“we’re locked in 🔥” answers nothing. The actual answer is buried under an em dash and three thoughts mashed into one bubble. Two emoji in four messages.

It wasn’t a one-off. The same energy ran through onboarding and logging:

“locked in” and “let’s get it” back to back, an em dash, two emoji, two sentences per bubble.

💀 plus one breathless run-on. A person sends that as two or three short texts.

The tells

• Emoji on every line.🔥 👀 💀 😭 🤙 became punctuation. They stopped meaning anything. Target: rare, most messages have none.
• Em dashes everywhere.“honestly yeah — morning weigh-ins…”, “that’s day one — let’s get it”. No human texts em dashes. Target: a period, a comma, or a new bubble.
• Hype filler.“we’re locked in”, “let’s get it”, “recomp gang”. Pep-talk voice on messages that didn’t ask for it. Target: nonchalant, say the real thing and stop.
• Two sentences in one bubble.Long run-ons instead of separate texts. Target: one sentence per bubble, split a second thought into a second bubble.

The change

Rewrote the VOICE and TEXTING STYLE blocks in src/prompt.mjs, and swept the same fixes through every model-facing string (tool descriptions, scheduler nudges, location callouts) so the few-shot examples stopped teaching the cringe:

• Nonchalant is the default.Not hyped, not a cheerleader, not “on.” A friend who notices and says it plainly.
• Emoji rare.Most messages none, never more than one, only when it adds something.
• No em dashes, ever.Removed from the persona, the examples, and every tool/trigger string the model reads.
• One sentence per bubble.Two thoughts becomes two bubbles, never a wall.
• Killed the cringe few-shots.The old “example energy” lines were the cringe; replaced with flat, low-key ones.

The target

Same question, same answer, just… normal. No hype, no em dash, one sentence each:

Answers the question first. No emoji, no em dash, no “locked in.” Each text is one thought.

Hypothesis 2: the agent always needs the last word proposed

Every thread ends the same way: Coach talking. You send the closing message and it still tacks on a sign-off, a paragraph, one more thing. It always gets the last word, and that word is always a wall. Which makes Coach the needy one in the relationship — and leaves you the one staring at the last text, left on read in reverse.

What it looks like now

You wrap the conversation. Coach refuses to let it land:

Three bubbles to say goodnight. You sent the closer; Coach overran it and left you holding the last text.

“yep got it” was the end. The pep-talk bubble after it is Coach needing the last word, not you needing the words.

The tells

• Coach always sends last.Scroll any thread — the final bubble is almost always blue. A real texter lets the other person have it sometimes.
• Closers turn into sign-offs.“got it”, “night”, “ok” should be allowed to just end. Instead each one earns a paragraph.
• It never just reacts.iMessage has a 👍 built in. Coach has never once used it — every acknowledgment is a full message.
• Left on read, in reverse.Because Coach overruns the close, the unanswered text is always its own. Reads as eager, not easy.

The change

The plumbing’s already there. imessage/helper.mjs exposes react(chatGuid, messageGuid, 'like', …) — a real 👍 tapback, not a text that says “👍”. Two pieces to wire:

• Give Coach a react tool.Expose the helper’s tapback to the agent in src/tools.mjs, so a 👍 is an action it can take instead of a message it has to send.
• Teach it the end-beat in the persona.In src/prompt.mjs: when the thread has clearly closed (they said night / ok / got it and there’s nothing new to add), prefer a 👍 over a reply. Don’t manufacture a reason to keep talking.
• Stay curious when it counts.Hard to get, not cold. If there’s a real thread to pull — what they like to eat, why the weigh-in slipped — Coach should still reach for it. React to close, ask to open.

The target

Same goodnight. Coach lets it end with a tap, and you keep the last word:

A 👍 on your message, no bubble after it. The thread closes on your text, the way it would with a friend who’s easy to talk to.