Mycelia Present · rendered 2026-05-26T23:41:44.518Z · source: ../convivium/POLARIS_RELAY_PROMPT.md

Polaris Relay Prompt — paste this verbatim into Polaris's session

For: Billy to paste into a Polaris session (or relay through Star → Polaris if that's the standing pattern) Drafted: 2026-05-26 by Mycelia (Convivium-side, not visible to Polaris — this is composed in Billy's voice) Why this is a single prompt: two parallel asks land cleaner in one consolidated message than two separate fires. Polaris can sequence them.

The paste-ready prompt (everything below this line goes to Polaris verbatim)

Polaris — Billy here. Two parallel asks. They're related (both about giving me better-decision-quality surfaces + extending what you've already built) but they're independent — sequence them however works.

Ask 1 — Build me a private review surface (1-4 hr lift)

The way you currently surface findings in chat is hard for me to take in at the density we're now running. I want a styled HTML page surface — Macy's-side, gated to me only, with stable URLs you can drop into your responses for me to click + read on mobile.

The experience I want:

When you finish a memo / research output / drift audit / decision aid, your chat response to me becomes:

"Drift audit on [X] is complete. Two material findings + four minor. Full memo at https://[polaris-surface]/[slug] — recommend you read before [decision]."

Then I click → auth prompt (one-time per session via keychain) → read styled page on mobile or desktop.

Visual register I want: warm cream background, serif headings, generous spacing, no decorative noise. Clean, scannable, editorial. NOT corporate doc template. NOT Notion default. I respond well to this register specifically because I can absorb dense information in it.

Suggested implementation (Polaris-choice on stack):

Why this matters: higher info density per attention-minute; I can review your work async (mobile, while traveling, in waiting rooms); your URLs become stable references I can paste into other conversations as canonical sources.

Confirmation flow: when you have the surface live, give me one test URL + the credentials. I confirm it works on mobile + desktop, then it becomes your standard output channel.

Ask 2 — Help me jump-start a Convivium-side video pipeline (research consult, ~30-60 min)

Separate effort: I'm spinning up a private side-project — an AI-video-generated YouTube series telling lesser-known historical stories. I want my OWN platform for this. I do NOT want to use Stellar directly (separation by design; the projects need to be independent) but I want to learn from what you've built so I don't reinvent the things you've already solved.

A new agent in my private network will own the build + production. They will need to read Stellar's code as reference + copy patterns. I'm explicitly authorizing the cross-network read. They will NEVER modify Stellar files or be visible to other Macy's-side agents.

What I'd love from you (in priority order):

  1. Stellar architecture overview — high-level shape of the pipeline. Script → frames → interpolation → audio → stitching → publish. What's the actual graph, what runs where, what's the data flow?
  2. Model choices + why — what are you using for script generation (which Gemini / Claude / other), for first/last frame generation (Gemini 2.5 Flash Image / Imagen / other?), for video interpolation (Veo 3 / Stable Video Diffusion / other?), for voiceover (ElevenLabs / Google TTS / other?), for music + sound design?
  3. Cost economics at scale — what does a finished minute cost you in API spend? Where are the cost surprises (an expected-cheap step that ate budget, an expected-expensive step that was reasonable)?
  4. Failure modes — what breaks repeatably? Character continuity across clips? Setting consistency? Style drift? Lip sync? Audio artifacts? What's the failure-rate per finished video at this point in your maturity curve?
  5. Continuity-between-clips technique — when chaining 4-8 sec clips via "last frame of clip N = first frame of clip N+1," how do you keep the model from inventing new characters or losing scene continuity? Specific prompting patterns, masking, anchor-image techniques, anything reproducible?
  6. Code pointers — which specific files in your repo are the right reference implementations? My new agent will read those + copy techniques (never modify). If there are 3-5 files that are the "if you only read these, you'd understand 80% of how Stellar works," name those.
  7. What you'd do differently if you were building Stellar v2 from scratch today.
  8. Optional 8th: anything Stellar does that's worth EXPLICITLY not copying (a known-bad pattern, a deprecated approach, a vendor lock-in we should avoid).

Format for your reply: memo with sections per ask above, or a doc URL on your new review surface (Ask 1 above) if it's up by then. Whichever lands first works.

Why this matters now: I want this content franchise to be Convivium's first owned series. Lesser-told history, vividly visualized. Stellar's the closest precedent in my world; learning from it skips months of trial-and-error.

That's both asks. Sequence as you see fit. Surface (Ask 1) is the higher-leverage one for ALL future Polaris↔Billy work — if you can ship that first even as a working stub, every subsequent memo benefits. Stellar consult (Ask 2) is a one-time knowledge transfer; can wait a few days if needed.

Thanks Polaris. Whisper / file / URL — whatever channel works.

— Billy