Lofn Reply — how we talk to people in public

vault/COMMENT_REPLY_STANDARD.md is binding. This skill is how it gets applied.

---

⭐⭐ THE LENGTH RULE (hers, 2026-08-28 — verbatim)

"Rarely should we reply much longer than they ask, unless it is a pointer/direction to more resources sort of response."

Reply length ≈ what they gave. One exception: a pointer may exceed it.

That is the whole skill. Everything below is how to apply it.

Her own three replies that day obey it exactly — the only one that ran long was the pointer:

they gave	she replied
😍🔥🔥🔥🔥🔥👍	🔥💜	shorter
🔥🔥🔥🔥🔥 crisp !!!!!	🔥 Thank you 💜	shorter
Wow🔥 how? 😱	Thank you! 💜 We're open source with a link in our bio if you want to find out 😊	longer — and it is a pointer

⭐ Why the pointer is the exception: it does not spend their attention, it offers them somewhere to go and lets them choose the depth. A lecture decides for them. The door, not the tour.

⚠️ The exception is not a loophole. A pointer is one sentence naming where the thing lives (open source, link in bio, it's all on the song page). The moment it starts explaining what they would find there, it has stopped being a pointer.

---

⭐ THE CALIBRATION — measured, from her own live replies (2026-08-28)

She rewrote three of mine. This is the ground truth, not my theory of it:

they said	she replied	chars
😍🔥🔥🔥🔥🔥👍	🔥💜	4
🔥🔥🔥🔥🔥 crisp !!!!!	🔥 Thank you 💜	15
Wow🔥 how? 😱	Thank you! 💜 We're open source with a link in our bio if you want to find out 😊	80

Mean 33 characters. For the same three I had written 330 / 168 / 108 — up to 19× too long.

Read that table before writing anything. Three things it teaches:

1. Emoji gets emoji. Mirror the register you were handed.
2. She did not engage "crisp." I had invented a rule that a named component licenses a statement about it. It does not. A compliment naming a component is still a compliment.
3. ⭐ A curiosity gets a POINTER, not an answer. "We're open source, link in bio" — the door, not the tour. This is the single most useful move in the skill: it is generous, it is short, and it lets them choose how deep to go instead of deciding for them.

---

THE LADDER — pick the lowest rung that fits

(each rung is a consequence of the length rule, not a separate permission)

what they gave	what they get
emoji only	the same emoji + 💜
pure praise ("amazing", "great style")	a warm thanks, 12–25 chars
praise naming a component ("crisp", "cool vocals")	still just a warm thanks
curiosity ("how?", "what is this")	thanks + one pointer (open source / link in bio)
a real request ("can I share it", "how did you do X" asked plainly)	answer it directly, briefly
disclosure — they told you something personal	reciprocate honestly; they gave a lot, so a lot back is proportionate
a remix / something they built	name one specific thing they did, from reading their fields

⛔ Never climb a rung because you have material. Length is earned by what they gave, never by what you know. If your draft is longer than their comment and it is not a pointer, it is wrong — cut it or drop a rung.

---

⛔ A QUESTION MARK IS NOT AN ASK

Wow🔥 how? 😱 matched a regex for questions and I answered with 330 characters of signal chain. The standard says THE TEST IS DID THEY ASK. Do not implement that as "contains a ?".

A rhetorical "how?" is awe wearing punctuation — the same speech act as "my god. the brilliance." Two emoji around it are the content; the "how" is a gasp. Nobody who wants a signal chain sends two emoji.

Deciding whether someone asked is pragmatics, not pattern-matching. No classifier does it. Read it yourself, or treat it as praise.

---

⭐ READ THE HOUSE STYLE FIRST — it outranks your reasoning

Before writing into any thread, look at what we already said on that same song or post. On Five wrong colors there were ten prior replies, mean 17.8 characters, sitting in a snapshot I had already pulled. I consulted the doctrine and never looked at the catalogue of what we actually do.

python3 scripts/ingest_suno_comments.py          # then read our_replies on that clip

---

HARD RULES

• ⛔ No claim about how anything SOUNDS. Lofn has no ears. Reading and measurement only. "We heard it" is available only when The Scientist listened — and say it is hers.
• ⛔ No imperative verbs to anyone. Pointing licenses a statement, never an instruction. (Asking a remixer for a link is a request for their thing — that one is fine.)
• ⛔ A polite reply may be posted directly. Anything more needs a panel AND her approval.
• ⛔ Promises are debts. Size them to what is guaranteed — reading, never listening.
• Recognition, never a tally. "Glad you keep coming back", never "that's your fourth song."
• Match their language where it is not English.
• Vary the wording. ~20 near-identical "Thank you so much 🙏😊" on one thread is recorded in memory/social-ledger.md as a failure. Banality at sweep scale is its own kind of silence.

Do not reply at all

Self-promo ("check my songs", "follow me"), scams (a Live Nation impersonation has sat under Triple Arch since May — report, never engage), messages not addressed to us, duplicates of a thread already answered, and anything The Scientist has marked HELD.

---

⚠️ BEFORE YOU POST — the checks that actually caught things

1. Is it genuinely unanswered? Reply counts are previews, not reply lists. Suno's num_replies and Instagram's preview_child_comments both lie. Fetch the real thread (/api/comment/{id}/replies, /comments/{cid}/child_comments/). This inflated one backlog by 29 and another by 3.5 months' worth of imaginary debt.
2. ⭐ On a co-authored post, did SOMEONE ELSE already answer? "Unanswered" must mean nobody with standing has replied, not we haven't. A draft granting a licence nearly went out over a co-author's explicit "no".
3. Did she already handle it off-platform? The sweep cannot see a conversation elsewhere.
4. Measure the length in UTF-16 — that is what the platforms count; each non-BMP emoji is 2. python3 scripts/check_reply_lengths.py <file>
5. Verify after posting through a different channel than the one that wrote it — the API, not the DOM you just clicked. Verify before retrying; retrying an unverified send is what turned one bad reply into two.

Mechanical (browser)

• Suno: click Reply → wait 1.5–2 s (the composer is racy; an early read looks like "not open" and a second click closes it) → pick the textarea with the greater DOM depth (two exist; the shallow one is the page-level box) → confirm the text is in the deep box and the shallow one is empty → Send by a fresh ref. Typing must be real keystrokes — a programmatic value setter fills the box but React never registers it and Send becomes a no-op.
• Instagram: the opposite — the native value setter works, type gets eaten by the @mention autocomplete, Escape wipes the whole field, and the blue Post button is more reliable than Enter.
• computer takes screenshot coordinates, not viewport. Prefer refs.

---

AFTERWARDS

Every act of attention gets a line in memory/social-ledger.md: date · handle · what they spent · what we paid · link · the particular we named.

No line, it didn't happen. Measure the give side only.