ve-intake

ve-intake — Stages 0–3

What this skill does

Takes a topic from the user and produces episodes/<slug>/packages.md — three meaningfully different angles scored for a shorts A/B test. Ends with a user checkpoint. Downstream: ve-produce.

Stage 0: Preflight

Before any paid work, run:

ve-doctor --live

This validates all API keys (ZapCap, ElevenLabs, OpenAI, Gemini, Upload-Post), binaries (ffmpeg, ffprobe, Chrome), and vendor connectivity. Do not proceed past Stage 0 until this exits 0. If any check fails, surface the output to the user — ve-doctor prints exact remediation per missing item.

If the plugin is freshly installed, keys come from the userConfig → keychain bridge (SessionStart hook). If the user updated a key, they may need to restart Claude Code for the hook to re-export.

See references/env_setup.md for the full per-key reference.

Stage 1: Intake

Confirm with the user:

• Topic (one sentence — broad area)
• Constraints (topical news peg, specific source required, evergreen vs trending)

Do not ask for a "thesis" or "title" yet — those emerge from the angles pass.

Set the episode slug (NNN-kebab-case-topic). Create $EPISODES_DIR/<NNN-slug>/ if needed.

Stage 2: Angles research

Always start with a broad research pass. The angle pool is the single most important creative decision in the whole pipeline.

Run the deep-research angles-mode flow per references/deep_research/SKILL.md. The vendored skill defines four facet agents (surprising facts, counterintuitive claims / overreach, unresolved debates, real-world anchors). Read its Phase 2 breadth-search prompt template and dispatch the four agents in parallel via the Agent tool. Synthesize per its angles-mode template in Phase 5.

(Optional shortcut: if the standalone deep-research skill is loaded in-session, invoke it via the Skill tool — identical output. The vendored copy is the durable contract.)

Save synthesis to $EPISODES_DIR/<slug>/research_angles.md. This file is permanent — never overwrite it later. Stage 4's per-angle hook research writes to separate files (research_hook_{A,B,C}.md).

Stage 3: Pick 3 angles for shorts

From the candidate-angles list, pick three that are meaningfully different from each other — different framings, hook concepts, visual anchors. Three variants of the same thesis is not three angles.

Use the schema in references/shorts_angles_template.md. Each angle has these fields:

• Core idea (1 sentence)
• Hook concept (1.7-second promise)
• First-frame concept (visually anomalous opening)
• Length strategy (loop-maximized 11–25s OR watch-maximized 40–60s — never 25–40s)
• Loop technique (visual-match-cut | narrative-callback | audio-continuity | cliffhanger-reversal | open-question-close)
• Payoff concept (what the last frame shows)
• Cover concept (separate from first-frame — wins on profile grids)
• Why this angle has its own short (variance vs B and C)
• Research anchors (2–4 sources from research_angles.md)

Write $EPISODES_DIR/<slug>/packages.md using this schema. Set flow_version: v4.0 in frontmatter. Set winner: null and judged_by: null (filled later by ve-analyze-shorts).

Checkpoint: present the 3 angles to the user as a compact summary (core idea + hook + length per angle, no commentary). User approves all 3 OR redirects ("drop B, swap in something on X"). Once approved:

Handoff to ve-produce

After Stage 3 checkpoint passes, the state is: packages.md written, research_angles.md saved, no per-angle hook research yet. Hand off to ve-produce — it runs Stage 4 (per-angle hook research) then Stage 5 (produce 3 shorts).