do-debrief

name	do-debrief
description	Use when sending a spoken debrief to a Telegram chat. Collects context, drafts a 30-second executive brief, synthesizes it via TTS, and delivers it as a native voice message. Triggered by 'send a voice debrief', 'speak this update', 'do-debrief', or any request to deliver an audio summary.
argument-hint	<scope-or-notes> --chat <chat>
allowed-tools	Bash, Read, Grep
user-invocable	true

/do-debrief — Spoken Executive Brief to Telegram

Construct a 30-second executive brief, speak it, deliver it as a Telegram voice message. The skill does the construction work — it does not just synthesize whatever text you hand it. The output is shaped for decisions, not information.

Why this exists

Executive bandwidth is the scarce resource. A brief earns its 30 seconds only when every item is something the recipient needs to decide, act on, or know for a meeting they'll walk into that day. Anything you're already handling should never appear. Calendars get echoed back — surface only the anomaly.

Inputs

• scope (required, positional) — Either a short framing of what to brief on ("morning standup", "deploy debrief", "post-merge update for psyoptimal") or raw notes you want shaped into a brief.
• --chat (required) — Target chat name (e.g. "Dev: Valor") or numeric chat ID.
• --voice (optional) — Voice name; defaults to am_michael (Kokoro). bf_alice is the female alternative. See tools/tts/README.md.
• --reply-to (optional) — Telegram message ID to reply to (required for forum-group topics).
• --no-preface (optional) — Suppress the one-line text preface that normally precedes the voice note. Default is to send the preface; pass this flag for pure-audio delivery.

The brief shape (target: ~30s, ~70 words spoken)

1. Top decision + your recommendation (~10s, ~25 words). Lead with the ask, not the context. Use default-and-confirm, not open questions: "I'm pushing the vendor call to Thursday unless you want it sooner" beats "What should we do about the vendor?" Their job is to veto, not deliberate.
2. Second decision OR critical heads-up (~8s, ~20 words). One thing only. Drop this slot if there isn't one — silence beats padding.
3. Batched FYIs (~8s, ~20 words). Open with "Three quick FYIs:" then one clause each. No laddering ("Also… and another thing…").
4. Close: "I've got the rest." (~2s).

Construction phases

Do these in order. Skipping phases produces flabby briefs.

1. Collect (parallel)

Pull raw material in parallel — single message with multiple Bash calls:

• git log --oneline -20 origin/main — recent commits
• gh pr list --state all --limit 10 — open + recently merged PRs
• python -m tools.valor_session list — session activity
• valor-telegram read --chat "<scope-relevant>" --since "24 hours ago" — outstanding chat threads (only if scope names a chat)
• Calendar anomalies only for daily/morning briefs: gws calendar events list --params '{...}' — surface only items that moved, conflict, or are net-new since yesterday. Never read the agenda back.

If the user passed raw notes as the scope, skip the pulls and treat those notes as the corpus.

2. Categorize

For each candidate item, classify it into exactly one bucket:

• Decision — needs the recipient's yes/no.
• Critical heads-up — they'll get blindsided in a meeting today if they don't know.
• FYI — material but non-urgent; fits in one clause.
• Already-handled — drop. You handling it ≠ them needing to know.
• Noise — drop.

If Already-handled + Noise together account for >70% of raw material, the brief probably isn't worth sending today. Tell the user that and exit, instead of synthesizing filler.

3. Identify gaps

For every Decision item, confirm you have:

• the current default course of action,
• the deadline or next checkpoint,
• the cost of being wrong.

If any of those are missing, consolidate the unknowns into a single clarifying question to the user before drafting. Do not ask piecemeal. If a Decision item still can't be resolved, demote it to Critical heads-up or drop it.

4. Draft pass A

Write the brief in the shape above. Apply these rules verbatim:

• Default-and-confirm phrasing on every decision: state your intended action + the unless-clause.
• Contractions ("I'm", "don't", "we're") — written prose reads stiff aloud.
• Never recite issue/PR numbers (or any multi-digit identifier). TTS reads "1195" as either "one thousand one hundred ninety-five" (cumbersome) or "one-one-nine-five" (meaningless) — both waste the listener's attention, and the number isn't actionable in audio anyway. Refer to the work by substance: "the continuation-PM crash," not "issue 1195." If the listener needs traceability, follow the voice message with a written brief that includes the numbers.
• Proper-noun respelling for TTS prosody. Product names benefit from hyphenation when they should sound like one word with multiple syllables (e.g., spell "Yudame" as You-duh-may). Dictionary-style hints only — never IPA in slashes; the phonemizer reads /.../ literally and doubles the clip duration.

5. Pass B — cut and re-shape

• Count words. If >80, cut. If <55, you're padding — drop the weakest FYI.
• Verify the first sentence is the ask, not setup.
• Verify nothing in the brief is something you're handling yourself.
• Read it through once mentally — first pass at filler-word removal almost always frees ~10s for what matters.

6. Preface line

Build a one-line text preface that lands in the chat right before the voice bubble. It exists so the recipient can glance at the chat and decide whether to tap play, and so the message is searchable later by something other than waveform.

Format: Brief update as of <H:MM AM/PM> · <N> items · <Q> questions

• items — total things mentioned in the brief (Decisions + Critical heads-up + FYIs). Already-handled and Noise are dropped during categorization and don't count.
• questions — number of Decision items (each one is a default-and-confirm the recipient can veto).
• timestamp — the recipient's local hour:minute with am/pm. Use date +"%-I:%M %p" (machine localtime is fine for a same-day brief).

If the user passed --no-preface, skip this entirely.

7. Review gate

Show the final transcript and preface to the user with one line: Final transcript (~Xs, Y words). Preface: "<preface>". Synthesize and send? Only proceed on explicit confirmation. This is the cheap moment to catch a bad default-and-confirm before it lands as audio in someone's chat.

Delivery (only after confirmation)

OUT=$(mktemp -t debrief).ogg

valor-tts --text "$TRANSCRIPT" --output "$OUT" || {
    echo "Synthesis failed"
    rm -f "$OUT"
    exit 1
}

# Preface (skippable). Send before the voice note so it lands first in the chat.
if [ -z "$NO_PREFACE" ]; then
    valor-telegram send --chat "$CHAT" "$PREFACE"
fi

valor-telegram send \
    --chat "$CHAT" \
    --voice-note \
    --cleanup-after-send \
    --audio "$OUT"

The relay owns the audio file from the moment the payload is pushed — it deletes on successful send OR after dead-letter placement on retry exhaustion. Synchronous deletion races the relay's retry loop.

Anti-patterns

• Reading the calendar back. They have it. Surface only the anomaly — "Your 2pm moved to 3" — not the full agenda.
• Open-ended questions. "What about the vendor?" wastes 30 seconds of attention. State your default; let them veto.
• Laddered FYIs. "Also… and another thing… oh and…" — pre-batch into "Three quick FYIs:" with one clause each.
• Status-of-status. "I'm working on the migration" isn't brief-worthy. Either it shipped (FYI) or it's blocked on a decision (Decision).
• Synthesizing raw notes verbatim. That's a dictated memo, not a brief. The construction phases above are the value the skill adds.
• Reciting numbers in audio. Issue numbers, PR numbers, port numbers, and dollar figures with many digits all read badly through TTS and don't help a listener anyway. Substance over identifier; numbers go in the written follow-up.
• Skipping the review gate. TTS + Telegram is one-way; the wrong "I'm pushing the vendor call to Thursday unless you want it sooner" is permanent once it lands.

Error handling

• valor-tts exits non-zero → STDERR carries Error: <message>. Surface it verbatim. Delete the partial file. Do not push to the outbox.
• valor-telegram send exits non-zero → the payload was not enqueued. The temp file is still on disk; remove it manually so it doesn't leak.
• Bridge relay not running → the payload sits in Redis until the relay starts. If you need synchronous confirmation, run ./scripts/valor-service.sh status first.

Why no /tts skill

tools/transcribe/ has no /transcribe skill — its CLI + README is the stable agent-facing surface. tools/tts/ mirrors that. Agents invoke valor-tts via Bash directly for raw synthesis. /do-debrief is the one user-invocable composite, and the construction phases above are why it earns its own skill instead of "pipe text to valor-tts."

Related references

• tools/tts/README.md — full TTS API, voice catalog, troubleshooting
• bridge/telegram_relay.py — _send_queued_message voice-note branch + cleanup_file honoring
• docs/features/tts.md — feature design + dual-backend rationale