Editor Bunny — Sync with Holly

Date: April 29, 2026 Participants: Alton Wells, Holly (founder, Editor Bunny) Format: Live Figma walkthrough + strategy session

TL;DR

• Holly walked through a 2-hour Figma mockup of the Editor Bunny platform — an agentic content/UGC ops tool aimed at founders and their interns. Strong work; the UX shape is right.
• Strategic call: don't build an editor marketplace. It compresses margin and adds operational complexity. Path forward is staff editors → quality-assurance reviewers + AI editor as the model matures.
• Bigger disagreement on timeline for AI short-form video editing: Alton's view is 12–24 months until it's good enough; Holly's view is ~5 years. Either way, design the system so the editing layer is replaceable.
• V1 / MVP is dramatically smaller than the full Figma: an internal tool for Holly + her interns + her current clients (Locale, Scoops, Jack). Just scripts, video uploads, scheduler. Vibe-coded.
• Tech stack confirmed: Mastra (TypeScript agent framework). Avoid OpenAI's Agent Builder canvas. Wire Mastra's MCP docs server into Claude Code so the AI has the right docs in-context.
• Followups: map all user types in FigJam, scope the V1 build, research social posting APIs and a CMS pattern for brand assets.

What Holly is building (product recap)

Editor Bunny is an agentic platform where the chat is the primary interface, with manual screens as fallback for "control-freak" founders. From the Figma:

• Chat-first home with manual tabs for everything the chat can do. The chat can be pulled out as a side panel from any manual screen.
• Video & editing pipeline: upload raw footage → fill an editing brief (count, target platforms, link inspo, expedite toggle, scheduling preference) → goes into editing → moves to Ready for Approval → approve/reject-with-notes/download.
• Scheduler & analytics: per-platform calendars, engagement metrics, content log.
• UGC creator side (open question — see below): sourcing, onboarding, scripts, video submission.
• Brand tab / CMS for assets, product imagery, brand guidelines (chat reads from this).
• Integrations: Instagram, TikTok, YouTube Shorts, LinkedIn, X, plus Instagram still-image posts. Surfaced as a flat top-level list, not buried in admin.

Users vs. customers (key distinction)

Action: Map all four user roles in a FigJam alongside the product file before designing more screens.

The strategic call: editor marketplace vs. staff vs. AI

Holly raised the question of whether to build an editor marketplace alongside the UGC creator marketplace. We worked through it on a whiteboard.

Why no marketplace for editors

• Margin compression. Every dollar paid to an outsourced editor leaves the business.
• Operational drag. Sourcing, vetting, ticket assignment, on-time delivery, payment processing, dispute handling — all overhead that doesn't compound.
• Worse customer experience. Brand voice and quality consistency are hard with a rotating editor pool.

Recommended evolution

The economic argument: at 100 customers and 1:10 staffing, even cheap offshore editors are ~$300–400k/yr in headcount. That cost goes away (or compresses to 1:100 reviewers) once the AI editor is good enough — and the human QA layer never goes away because brand quality has to be guaranteed.

Disagreement on AI video timeline

• Alton: Long-form AI editing is further out, but short-form is climbing out of the uncanny valley fast — 12–24 months. The model improvement curve has been faster than expected.
• Holly: Tried multiple AI editing startups (including one she met with directly). All bad. Estimates 5 years even at exponential growth.
• Holly's secondary point: Even when AI editing is good, Gen Z and millennials reject AI slop from brands — they want real people with real products. So the use case is real UGC footage with AI-assisted editing, not fully synthetic video.

Resolution: Design the editing layer as a swappable component. Don't bet the company on either timeline.

Reference: Opus Clip

Alton's friend Nick has been using Opus Clip for short-form clips and gotten real engagement lift — it's a useful proof point that AI-assisted clipping is already viable for one slice of the workflow (long-form → vertical shorts). Worth Holly trying as a benchmark for "what good enough looks like today" in the clipping niche specifically.

• Auto-detects highlight moments in long videos
• Reframes to 9:16, 1:1, 16:9 with subject tracking ("ReframeAnything")
• Auto-captions in 20+ languages
• Has its own scheduler/autoposter and an API for integration
• Pricing: free tier with watermark; paid tiers $15–29/mo; business tier with API

This is not a competitor to Editor Bunny — Opus is point-solution for clipping. Editor Bunny is the workflow layer above it.

V1 / MVP scope (the actual build)

The full Figma is the 18-month roadmap, not the V1. Per Holly's prompt at the end of the call:

"Should I make a V1 that's an internal tool for me, my agents, and my current clients (Locale, Scoops, Jack)? Just script generation, content calendar, and leave it at that for now."

Yes — exactly that. Scope V1 to:

1. Script generation agent — accessible to Holly + interns
2. Video upload + brief form — clients upload raw footage with notes; editor sees them
3. Editor inbox — internal editor pulls videos + briefs to work on
4. Lightweight scheduler OR external scheduler integration — punt on metrics tracking by leveraging the scheduler's built-in analytics
5. (Defer) UGC creator portal, full chat orchestration, multi-platform integrations, approvals UI, brand CMS

Vibe-code it. Ship it for the existing client list. Iterate based on what they actually use.

Object model to clarify with the AI assistant when building

Be explicit about what the chat creates (objects):

• script
• post (a scheduled or published unit on a platform)
• plan (a content calendar / posting schedule) — probably defer

Tight scope on objects = predictable agent behavior. Don't let the chat output fuzzy "things."

Tech stack research

Framework: Mastra ✅

Use this. Do not use OpenAI's Agent Builder canvas.

Mastra is a TypeScript-native open-source framework for building AI agents and workflows. Built by the team behind Gatsby (Sam Bhagwat, Abhi Aiyer, Shane Thomas). Y Combinator-backed.

Why it's the right fit for Editor Bunny:

• TypeScript-native — Holly took 2 years of CS in college and coded in HS; TS will be approachable
• Production-grade primitives: agents, tools, workflows, memory, RAG, evals, tracing
• Human-in-the-loop built in — workflows can suspend and wait for approval, then resume. Critical for the approval pipeline.
• Deployable anywhere — Vercel, Cloudflare, Netlify, any Node host
• MCP-native — both authors and consumes MCP servers (relevant when integrating with social platforms or tools like Outstand)
• Used in production by Replit, PayPal, Sanity, Brex
• Built on top of Vercel's AI SDK with model routing across 1000+ models (OpenAI, Anthropic, Gemini, etc.)

Critical setup: Mastra MCP docs server in Claude Code

This is what Alton meant by "I'll teach you how to use the MCP." Mastra ships an MCP documentation server that gives the AI in your editor real-time access to Mastra's full docs, code examples, and changelogs. Without this, the AI hallucinates outdated APIs.

Install in Claude Code (one terminal command):

Or the generic config (works in Cursor, Windsurf, VS Code, Codex):

Project bootstrap:

The interactive wizard installs the MCP docs server, picks model providers, and scaffolds an example. Pick Claude Code or Cursor when prompted.

Framework: OpenAI AgentKit / Agent Builder ❌

AgentKit launched at OpenAI DevDay 2025. Visual drag-and-drop canvas for chaining agent nodes, with ChatKit for the UI layer.

Why we're skipping it:

• Visual workflow builders don't scale past a few branches — turn into tangled boxes-and-arrows once real conditional logic enters
• Locks you into OpenAI's stack and opinions
• The "no-code" promise breaks down quickly — you still need to wire connectors, debug, version
• Common critique: too constrained for engineers, too complex for non-coders
• You can export to the Agents SDK (code) but at that point you're not getting value from the canvas

Mastra gives the same primitives in code, with no canvas-tax and no vendor lock-in.

Image generation: Nano Banana

For the AI-assisted product imagery use case (generate variations of brand assets, place products in scenes, etc.), Nano Banana is Google's native image gen in the Gemini API. Three flavors:

Why this matters for Editor Bunny:

• Subject consistency — keeps the same character/product across multiple generations (huge for brand asset libraries)
• Multi-image input — upload product photo + scene reference, generate composite
• Accurate text rendering — for posters, ads, marketing creative with legible copy
• Conversational editing — refine iteratively rather than re-prompting from scratch

This pairs cleanly with the brand CMS Holly was already planning — the CMS holds source assets, Nano Banana generates novel content from them.

Social posting integrations: don't build platform-by-platform

This is the trap. Each platform (Instagram, TikTok, YouTube, LinkedIn, X) has its own auth, rate limits, content rules, business-account requirements, and API approval cycles. Building 5+ from scratch is months of unrewarding work.

Use a unified social posting API. Top contenders worth Holly evaluating:

Recommendation: Start with Ayrshare or Post for Me for V2 (when scheduling becomes part of the platform). They handle the OAuth, rate limits, and platform-quirk maintenance so Editor Bunny doesn't burn engineering on it. For V1, just have editors download finished videos and post manually or use an off-the-shelf scheduler.

If Holly wants to expose social tools to her Mastra agents directly via MCP, Outstand and Zernio both ship MCP servers — agents can call create_post, get_post_analytics, etc. as native tools.

Action items

Holly

1. Map users in FigJam — founders, interns, UGC creators, QA reviewers. What can each do/not do?
2. Tighten the object model — confirm the chat creates: scripts, posts. (Plans deferred.)
3. Update Figma: flatten Integrations to top-level (rename — "Integrations" or "Plugins"; "Integration's" was a spelling/autocomplete artifact).
4. Scope V1 build doc — internal tool for Holly + interns + 3 named clients. Scripts + video uploads + brief form + editor inbox. Nothing else.
5. Install Mastra + MCP docs server in Claude Code before starting to code.
6. Don't put Editor Bunny on LinkedIn yet — wait until past $50k MRR. Hypsonic overlap concern noted (no non-compete, AI-checked the contract).
7. Try Opus Clip as a benchmark for current-state AI clipping quality.

Alton

1. Send Mastra MCP install snippet (above ✅)
2. Connect Holly with Nick (Opus Clip user) — potential to grow his socials in exchange for testing context
3. (When ready) Walk Holly through peptide protocol for her boyfriend's stack — separate convo

Open questions / parking lot

• UGC creator flow — single app, separate creator portal, or hold off entirely until V3? Holly is leaning toward a separate creator app long-term. Defer the design decision; it's not in V1.
• YC application — Alton thinks Editor Bunny is one of the few cases where YC makes sense (the customer base is other YC founders who want to be famous). Holly is against it. Tabled. Revisit after first 5 paying customers.
• Quality guarantee mechanism — when AI editor produces output the QA reviewer rejects, what's the routing? Back to AI with feedback? To human editor? Build the rubber-stamp UX before this matters.
• Pricing model for "expedite tokens" — flagged in the Figma as an upcharge for rush jobs. Needs a pricing exercise before V2.
• Hypsonic competitive risk — Holly's previous employer (working with Tower 28 etc.). No non-compete on file. Watch their public moves; don't broadcast Editor Bunny on LinkedIn until there's traction.

Aside: non-product topics covered (logged for completeness)

• Peptides discussion (retatrutide titration protocol, Holly's interest for her boyfriend, Alton's product idea on the side)
• Alton's "1-800-Get-Money-Now" approach to funding agent rebuild — he's not raising; he's bootstrapping a peptide product to fund the agent work
• Mutual agreement that fame is overrated; Wyoming-ranch-and-a-plane is the dream
• Holly is not on retatrutide currently; previous bad experience due to incorrect reconstitution/dosing

Notes generated from live transcript. Reach out if anything is misrepresented or needs to be reframed.