Finale pitch

The live round performs something already built, and it is won before the slot starts: inputs rehearsed until boring, numbers that match the repo, limits said before the jury finds them, and a rehearsed plan for every way the demo can fail. This skill writes four files (RUNSHEET.md, PITCH-SCRIPT.md, QA-BANK.md, PREFLIGHT.md), drills them, runs the day and closes the event out. It also covers rounds where mentors hand over a surprise challenge to build into the project within fixed hours.

Arguments

Inputs

• BULLSEYE.md (template: BULLSEYE.template.md): ## Contract for the round format (slot length, Q&A length, onsite or online, the jury, required deck), ## Thesis, the golden path in ## Plan, ## Claims, ## Risks.
• facts.json, CLAIMS.md and LIMITS.md from the sibling skill prove-it (install the full set with npx skills add SaudSatopay/bullseye if missing). Every number in the pitch and the answers comes from them.
• The deck (sibling skill pitch-deck), the backup video (sibling skill demo-video) and the latest findings from the sibling skill judge-panel.
• The live URL, a local production build that runs without the network, seeded demo data with a reset, and every device that appears in the demo.

Procedure

1. Fit the structure to the slot and the jury

Write the format at the top of PITCH-SCRIPT.md: slot length, Q&A length, onsite or online, jury type, language, deck required or not, demo live or recorded. Then pick the structure:

Weight the jury: founders and VCs get the product and the money first (ROI with its basis, the buyer, the wedge); engineers get the mechanism and the failure modes; bankers, regulators and government get governance (maker-checker, audit, a human at the risky step); sponsor judges get what their platform did and honest friction. Jury types, what each rewards and what loses them: judge-panel personas. Openers, closers, word budgets and the ask: references/pitch-structures.md.

Gate: every section has a start time and a length, and they sum to the slot minus at least 15 seconds.

2. Assign roles and freeze

• One pitch owner speaks and owns PITCH-SCRIPT.md. One demo driver drives and owns RUNSHEET.md. The speaker never waits for the screen; narration continues over any lag.
• Solo: you are both. Pre-open every tab, use reset and autorun deep links so no beat needs typing beyond its input, and keep the run-sheet in a second window or on paper.
• The presenter does not pull the all-nighter. After the feature freeze the presenter rehearses and sleeps.
• After the freeze, only golden-path and honesty fixes land; anything else waits.

Gate: roles in the header of both files; the freeze time in ## Decisions.

Field note (Dhaal · MUJ HackX 4.0): The team declared a feature freeze at hour 12, yet hardening ran until 05:35 with judging at 11:00. A freeze holds only if it is written down with the one exception: golden-path and honesty fixes.

3. Write RUNSHEET.md

Copy assets/RUNSHEET.template.md. Every beat has a paste-ready input in a code block, the expected result with the number the screen shows, and what to say. Never improvise inputs on stage. Rehearse until boring.

• The judge five-beat, a 3-minute skeptic's tour for the judge who says "show me it's real": J1 an unfamiliar case no rehearsed input used; J2 a legitimate case (it does not cry wolf); J3 an ambiguous input (it asks instead of guessing); J4 the mechanic-level insight (the case a keyword system misses); J5 one full loop whose effect shows on a second device.
• A contrast beat in the main path too: the case where it stays quiet.
• Latency lines: for every step that takes more than 3 seconds, the line said while it runs. Never wait silently. The pause reads as gravitas if you own it.
• The closer: the judge's own phone, inbox or data when the product allows it, with a printed card holding beat 1's input as the fallback. Ask first, and never keep or screenshot what a judge shares.
• The header holds the devices, the hotspot, the live and local URLs, the backup video's two locations and a pre-paired backup link for any fragile setup step.

Gate: every input was pasted and run in rehearsal on the production build, and every expected result matches today's screen.

4. Write PITCH-SCRIPT.md

Copy assets/PITCH-SCRIPT.template.md.

• Timed sections with start times. [SQUARE BRACKETS] are actions: do them while speaking, never silently.
• Bold every number with your voice. They are the pitch. Every number comes from facts.json; an external statistic carries its date, said out loud when pressed.
• Open on a dated number or the judge's own experience. Close on the ask, or on "Round 1, we promised this. Round 2, it runs." when the event had rounds.
• The demo section names RUNSHEET beat numbers; the slide notes carry the same beats.
• Write an EMERGENCY 90-SECOND CUT and a NUMBERS CHEAT CARD (every number on one card, glanced at before going on).
• Time it aloud. Cut words; never speed up.

Gate: three timed run-throughs after the freeze, each inside the slot, and the 90-second cut timed once.

Field note (FLOODLIGHT · MUSA CodeX 2026): The Round 2 script was planned section by section to about 4:30, made the quiet day with zero alerts its own act ("THE PROOF"), opened the free host two minutes early because it sleeps, kept a local run as the fallback, and ended with a 90-second cut and a numbers cheat card. The project had qualified from Round 1.

5. Build the Q&A bank and drill it

Copy assets/QA-BANK.template.md and keep it gitignored.

• Honest limits first: say these before judges find them, each with its roadmap line.
• Every answer in three parts. LIVE: 2 to 5 speakable sentences, the first one answering the question. SHOW: one thing on screen. IF PUSHED: a one-line deeper defense.
• Questions by persona: technical, business, domain expert, sponsor, skeptic, team and process.
• Never quote a number without the caveats printed beside it in CLAIMS.md. Every answer points at something committed: a file, a test, a command.
• What did the AI build? Say it plainly. Judges check that you understand it, so know every row of the build table.
• One dense line of numbers at a glance, re-verified on the day.
• Drill it with the sibling skill judge-panel (/bullseye:judge-panel qa, or /judge-panel qa when installed with npx skills add): a fresh session asks the bank's questions one at a time and grades each answer against the project's facts. Log every miss in the bank and fix the answer.

Gate: every persona's questions answered aloud in under 30 seconds each, no number without its caveat, every limit said at least once. Answer craft and hostile patterns: references/qa-craft.md.

Field note (CrackWatch · Gappy AI Hackathon 2026): In the top 25, the team prepared for a 20-minute judging call with a brief verified against the code: honest limits to say first, a T-15 pre-flight, a decision rule, one line of numbers and fourteen speakable answers.

6. Pre-flight, decision rule and failure drill

Copy assets/PREFLIGHT.template.md and run it at T-15:

• Warm everything 2 minutes before: free hosts sleep, and the first request can take a minute.
• Re-verify the rehearsed numbers on the screen; demo stores keyed to the hour or the date drift.
• Devices on a phone hotspot, not venue Wi-Fi. The local fallback running: a production build, demo mode on (a DEMO_MODE=1 style flag that serves vetted cached AI output), and no --reload (a reloading server restarts on every edit and can die mid-restart).
• Backup video in two places (the presenter's phone and the laptop desktop) and a pre-paired backup link.
• A fresh browser profile or a cleared cache, with index.html served no-store, so no stale bundle reaches the stage.
• Decision rule: newest run COMPLETED → demo live; FAILED → walk the graph on the recording; never bluff a broken step.
• Failure drill, rehearsed once end to end: slow → narrate; API dead → switch to the local run; everything dead → the backup video, zero apology, keep narrating.

Gate: the drill was rehearsed once, including the switch to the local run and to the video.

Field note (ORCA · SIH 2026): A stale browser tab once showed the previous day's bundle on stage. The fix was serving index.html with no-store, so every load fetches the current build.

7. Online judging

• Share the Chrome tab, not the window, with "Also share tab audio" on; a shared window sends no sound.
• A second monitor holds notes only: the run-sheet and your own printed bank. Nothing listens to the call.
• Test the setup the day before on the same platform with a teammate: screen-share permission (macOS asks for Screen Recording and needs a browser restart), camera, mic, who shares, the meeting link.
• Zoom the browser to about 125% so text survives compression; close other tabs; notifications off.
• Paste the live URL into the chat so judges can click it, and keep the deck as a PDF in a tab as the fallback.

Gate: a full dry run on the same platform, with sound, the day before.

8. Surprise-challenge rounds

When mentors hand each team a challenge to build into its existing project within fixed hours:

• Before the round: in the earlier stage, pre-pitch the seams the challenge will land in (a new agent is one file, a rule is a config line, a language is a table) and the hour plan. "Hand us any challenge. The crew absorbs it."
• H0 to 1: read the brief, transcribe it verbatim into a plan file, pick the seam it lands in, post the plan with a brief-to-task table and global constraints: rehearsed numbers unchanged, healthy-path text byte-identical, no new dependencies.
• H1 to 6: build inside the seam with tests named after the brief's scenarios; the rest of the project stays untouched.
• H6 to 9: wire it into the UI and every language's strings.
• H9 to 11: smoke test green, plus a demo fixture for the new path.
• H11 to 12: rehearse twice. The old demo stays byte-identical.
• The deck pays off seam by seam: the brief with "what the mentor checks", before and after, the drills, edge cases, real test output with counts before and after, a timeline slide from the git log (brief at 12:05, shipped by 14:50, commits, lines, tests), and "What the brief asked us to show ✓ … page N".

Gate: every item in the brief maps to a task, a test and a page. Plan skeleton and deck structure: references/surprise-rounds.md.

Field note (ORCA · Vibeathon 2026): The Stage 1 deck promised the seams and a 12-hour playbook; in Stage 2 the mentor's safety-gate brief arrived at 12:05 and shipped by 14:50: 27 commits, 3,213 product lines and 219 new tests, with the healthy demo untouched.

9. After the round

• Do not edit the submission after the deadline. Later work goes after a tag, under an "After the deadline" note.
• Keep judge access alive until results: sandbox invites, logins, the deployed URL, keep-alive for free tiers.
• Thank sponsors and mentors with specifics: what you used, what broke, the one feature you would ask for. Credit any team whose finding helped you.
• Rotate every secret that passed through chat or appeared on screen.
• Record the outcome exactly as announced and one lesson in TRACKER.md with the sibling skill hackathon-tracker (archive <event>).

Gate: the tracker holds the outcome and the lesson, and the status line in BULLSEYE.md reads phase=done.

Outputs

• RUNSHEET.md, PITCH-SCRIPT.md, QA-BANK.md, PREFLIGHT.md in the project root; in a surprise round, a plan file such as .claude/plans/<date>-<challenge>.md. Print the run-sheet and the numbers cheat card.
• Git: QA-BANK.md is always gitignored. The others follow the process-file policy recorded in BULLSEYE.md (private by default).
• BULLSEYE.md: status phase=finale with the next action; ## Decisions (structure, roles, freeze time); ## Risks (each live-demo risk with its mitigation); a ▶ STATUS · rehearsal block in ## Log per run-through (time, misses, fixes); ## Handoff rewritten before the slot; after the round, phase=done and the outcome.

Rules

• Never improvise inputs on stage. Rehearse until boring.
• Never bluff a broken step: walk the recording and say what it shows.
• Never quote a number without its caveat, and never claim what has no committed file, test or command behind it.
• Say the honest limits before the judges find them.
• No AI answers judges live. The bank is rehearsed, not consulted by a listening assistant: a hidden assistant can break the event's rules and the judges' trust. If an assistive tool is used during judging (captions, translation, accessibility), disclose it and follow the event's rules.
• After the freeze, only golden-path and honesty fixes. The presenter sleeps.
• In a surprise round, the old demo stays byte-identical and every rehearsed number stays true.
• Ask before using a judge's phone or data, and keep nothing.
• Do not edit the submission after the deadline.

Reference map