wedding-toast-architect

Use when asked to write, revise, time, or safety-check a wedding speech or toast (best man, maid of honor, parents, friends) — builds a role-tuned 5-part outline with per-section word budgets, estimates speaking time at 3 paces, and red-flag scans drafts for the classics that bomb (ex mentions, drunk stories, divorce jokes, inside jokes, backhanded 'finally settled' compliments) with CUT/SOFTEN/CHECK verdicts.

Overview

Every year ~2 million US couples marry; nearly every wedding hands a friend or parent a microphone and 3 minutes they are unprepared for. The result is predictable: rambled inside jokes, accidental ex mentions, drunk-story confessions, "ball and chain" humor in front of both grandmothers — and a regretful speaker replaying it for years.

This skill treats the toast as a buildable artifact with a spec:

1. build — questionnaire → 5-part outline with word budgets per section, tuned by role (best man / maid of honor / parent / friend / sibling).
2. scan — draft → red-flag report: 16 pattern classes with CUT / SOFTEN / CHECK verdicts, protected-name enforcement, length vs target, Flesch readability for the ear, and long-sentence detection.
3. time — words → minutes at 110/135/150 wpm.

Pure Python stdlib. Works offline. No draft text ever leaves the machine.

When to Use

• "I'm best man in 3 weeks and haven't written anything" (any speech-role panic)
• "Is this speech okay? Be honest." — paste draft, run scan
• "How long will 500 words take to deliver?"
• Parent-of-the-bride/groom, maid of honor, sibling, close-friend toasts
• Rehearsal-dinner speeches, engagement-party toasts (same structure, shorter)

Don't use for: vows (intimate, couple-authored — different genre), ceremony readings (pre-existing texts), professional MC scripts, eulogies (see note: structure overlaps but tone rules differ — this tool would over-flag).

How It Works — Numbered Steps

1. Interview the speaker briefly: role, couple names, partner name, years known, 1-2 story candidates, 1-2 qualities those stories prove, target minutes (default 3).
2. Build the skeleton:

   text

   python3 scripts/toast_architect.py build --role best-man \
       --couple "Alex & Sam" --partner Sam --years-known 9 \
       --stories "rafting capsize; chili cook-off disaster" \
       --qualities "loyal, calm in chaos" --minutes 3
   

   Output: 6 sections (hook / who / story1 / story2 / partner / toast) with per-section word budgets and second-durations at 135 wpm.
3. Draft against the budgets — one story told well beats three told fast.
4. Scan the draft:

   text

   python3 scripts/toast_architect.py scan --draft draft.txt \
       --protected "Jordan,Riley" --minutes 3
   

   Fix every CUT finding. Decide on SOFTENs. Read CHECKs aloud.
5. Time the final cut: python3 scripts/toast_architect.py time --words 410.
6. Rehearse 3× out loud, standing, holding a full glass. Print 18pt+, one section per index card, cards numbered.

What the Scanner Catches (16 pattern classes)

Class	Severity	Why
ex-partner references	CUT	exes never belong in toasts
drunk/intoxication stories	CUT	all-ages room
strip club / vegas	CUT	bachelor content bombs
divorce / "ball and chain" jokes	CUT	reads as doubt about the marriage
"finally settled" backhanded frames	CUT	insults both partners
pregnancy mentions	CUT	not your news to announce
profanity	SOFTEN	shrinks the room
money / prenup jokes	SOFTEN	divides the room
looks-only compliments	SOFTEN	compliment choices instead
dating-app jokes	CHECK	only if couple approves the story
inside-joke markers ( ... )	CHECK	translate or cut
religion assumptions	CHECK	match the couple's faith
"However" early digs	CHECK	usually introduces a slight
cliché openers	SOFTEN	"Good evening, for those who don't know me..."
filler words on paper	SOFTEN	filler out loud too
protected names	CUT	exes, estranged, deceased, private people

Scoring: 100 − 18·CUT − 7·SOFTEN − 3·CHECK, floored at 0. Ship at ≥85.

Common Pitfalls

1. Treating it as a comedy set. Funny is fine, funny-only flops. The room wants warmth + one laugh, not ten laughs and no heart.
2. Three stories in three minutes. Budget says ~120 words for story1 — that's ONE story with a scene, complication, payoff.
3. Ignoring the partner section. Maid-of-honor speeches about the bride with 10 seconds for the groom read as a rejection of the marriage. Budget 14-24% of words for the partner (role-dependent).
4. Reading from a phone. Lock screen, glare, notifications. Index cards.
5. Not timing it. 500 words = 3.7 min normal, 4.5 min with adrenaline pauses. The couple scheduled you for 3.
6. Trusting the scanner as the only pass. It catches patterns, not tone. The final gate is the question printed at the end: would you say this to the couple's face with both grandmothers listening?
7. Writing the night before. The outline takes 20 minutes; the stories need a week of "which detail do I keep" walks.

Example

text

Speaker: best man, 9 years of friendship, worried about an ex mention.
$ toast_architect.py build --role best-man --couple "Alex & Sam" --partner Sam \
    --years-known 9 --stories "rafting capsize; chili cook-off disaster" \
    --qualities "loyal, calm in chaos" --minutes 3
→ 405-word budget: hook 48 / who 40 / story1 121 / story2 89 / partner 56 / toast 48

$ toast_architect.py scan --draft draft.txt --protected "Jordan" --minutes 3
→ 5 CUT / 2 SOFTEN / 1 CHECK, safety score 0/100:
  CUT  ex-partner reference (1x)
  CUT  drunk/intoxication story (1x)
  CUT  divorce/pessimism joke (1x)
  CUT  'finally settled' framing (1x)
  CUT  protected name 'Jordan' (1x)
  SOFTEN money joke, cliché opener   CHECK inside-joke parenthetical

Rewrite fixes all CUTs; rescan: 92/100 — ready for rehearsal.

References

• references/toast-craft.md — structure theory, role budget rationale, delivery-day protocol (mic, glass, cards), and the repair guide for each flag class with before/after rewrites.
• references/scanner-patterns.md — the full pattern table with regexes, severity weighting, known false-positive notes, and how to extend.