Session artifacts from an AI-orchestration team — songs, films, posters, quotes.
Every session of an AI-orchestrated work day at Neckarshore AI ends with a small ritual: each AI persona picks a song, a quote, a film, and a propaganda-poster prompt that fit the work just done. One paragraph of why, each. It started as a half-joke. It became a creative footprint.
This repo is the public surface of that footprint. Auto-synced from the session source-of-truth, with hallucination guards so only real songs, real films, real quotes make it in.
| Subfolder | What | Source skill | Status |
|---|---|---|---|
playlist/ |
Spotify-synced playlist of session songs, with confidence-scored metadata and "why" paragraphs | neckarshore-easter-eggs-playlist |
Coming soon (first sibling) |
movies/ |
Films picked at session-close, verified existence | neckarshore-easter-eggs-movies |
Stub |
images/ |
Propaganda-poster prompts + generated images per session | neckarshore-easter-eggs-images |
Stub |
wisdom/ |
Quote collection — both real attributions (Wikiquote-verified) and persona-original aphorisms | neckarshore-easter-eggs-wisdom |
Stub |
Awaiting the next session-close. The block below is the rolling slot — every session-end refreshes it. Older sessions migrate into per-month archives at v1.1.
Date: TBD at next session-close Persona: TBD Title: TBD
🎵 Song: TBD — Spotify link will be added here
Why this song fits this session — the persona's reasoning paragraph.
🎬 Film: TBD (Year)
Why this film fits this session.
💬 Quote: "TBD" — Attribution
Why this quote fits this session.
🖼 Image: will be added here — hand-rendered from the session's poster-prompt
The session's poster-prompt.
- A persona ends a session by writing one entry per asset class into the source files at
comedy-execution-engine/assets/(private repo) - A user-level Claude Code skill parses the source, verifies each entry against an external oracle (Spotify+MusicBrainz, Letterboxd+TMDb, Wikiquote+Wikipedia), and writes the canonical YAML + rendered README into the corresponding subfolder of this repo
- Suspected hallucinations are blocked from publishing — they land in
<feature>/unverified.yamlfor review
The image asset class started as a literal propaganda poster — that's how the ritual began. Bold, ironic, mid-century-graphic. But after a few weeks, the prompts drifted into a different register: atmospheric scenes, dream-logic, station platforms in morning fog, Italian pines, stationmasters checking pocket watches. Each session now has its own visual vibe — less propaganda, more stage-set for the work that just happened.
We kept the original folder name (images/) and skill name (-images) as honest archaeology. The first prompts were posters. The current ones aren't.
A representative prompt — for the upcoming Goldoni-launch session (2026-05-04):
🎨 Bahnsteig-Vorplatz im Morgennebel, 24 leere Reisekoffer in Reihe, jeder mit verblasstem Aufkleber wie
spotify-playlist-link. Im Hintergrund ein Zug mit Schild "ristorante-goldoni.de — Abfahrt Mo 04.05.2026", Stationsschild leuchtet Terrakotta. Italienische Pinien am Rand, Stationswärter schaut auf seine Taschenuhr.
Each prompt is one paragraph. One image per session. No prompt-hierarchy, no series. The constraint is the format.
A persona can fantasize a song. We've seen it. "Halluzinationskraut" by some made-up band sounds plausible inside a 3-AM session-close paragraph, and silent-fails into a playlist as either a no-result skip or — worse — a fuzzy-matched wrong song that nobody catches.
The four sibling skills share a verification pattern:
| Level | Confidence | Source |
|---|---|---|
| 1 | HIGH | Primary API exact match (Spotify / Letterboxd / Wikiquote) |
| 2 | MEDIUM | Primary API fuzzy match |
| 3 | MEDIUM | Fallback API confirmed (MusicBrainz / TMDb / Wikipedia) |
| 4 | LOW | Fallback API low-confidence — published with a warning |
| 5 | SUSPECTED HALLUCINATION | Neither source confirms — NOT published, flagged for review |
A user with a known-real-but-unstreamable artist (Jandek-style) can whitelist them explicitly. Everything else is empirical.
The four skills share a common structure:
- YAML-First storage —
<feature>.yamlis the single source of truth (R1 doctrine) - Markdown rendering —
<feature>/README.mdis auto-rendered, byte-deterministic - Idempotent sync — same input, same output, no empty commits
- Existence verification — primary API + fallback API + confidence scoring
- gh-CLI auth — no separate token storage, reuses existing GitHub CLI session
Architecture spec (private): omnopsis-ai/omnopsis-planning/docs/specs/easter-eggs-umbrella-architecture.md
- Per-persona pages — one
personas/<name>.mdper AI persona, aggregating all four asset classes filtered to that persona's contributions. The creative-footprint surface. - Auto-rendered top-level README — this README becomes a rendered artifact too, pulling latest entries from each subfolder
- Auto-sync hooks — write to source → playlist syncs without manual trigger
proposals/subfolder — a path for human PRs to suggest entries (closed in v1, see CONTRIBUTING.md)
Dual-licensed. Code is MIT. Content (why-paragraphs, prose, curated metadata) is CC BY 4.0. See LICENSE.
This repo is closed-write in v1 — only the four sibling skills push to it. See CONTRIBUTING.md for the rationale and the v1.1 plan to open up.
Neckarshore AI — see the company website for context.