Daily Drop & Character Roster

social/daily-drop is the multi-character fan-out workflow. Point it at a roster YAML, and it produces one or more talking-head videos per character in parallel — with a hard, pre-flight cost cap so misconfiguration can’t quietly empty your FAL credit pool.

Caution

social/daily-drop does not post the produced videos. Posting belongs to a consumer that subscribes to workflow.run.completed — see ADR 0001. The workflow output is a list of artifact paths + script text + formula labels; pick them up from your posting layer.

Roster schema

Rosters are portable, declarative documents in YAML, TOML, or JSON — format selected by extension. The Fabric team’s example roster is checked in at examples/character-rosters/fabric-team.yaml; your own roster lives wherever you point roster_path at, with default auto-lookup at ~/.fabric/characters.{yaml,toml,json}.

Note

Repo example rosters under examples/character-rosters/ are NOT auto-discovered — pass an explicit roster_path to load one. This avoids accidentally running a stale committed roster instead of the one you’ve been editing locally.

roster_name: my-team
daily_budget_usd: 10.00      # hard preflight cap (USD/day)

shared:
  language: en
  duration_secs: 30
  cost_path: cheap-fal       # local | cheap-fal | premium-fal

characters:
  - name: builder            # alnum / -/_ only — used as artifact prefix
    voice_style: "warm, technical"
    gender: male
    presenter_look: "engineer with glasses, soft window light"
    persona: "indie engineer demystifying systems"
    daily_video_count: 1
    default_script_formula: youre_doing_it_wrong
    allowed_formulas: [youre_doing_it_wrong, reframe, listicle]
    topics_pool:
      - "Why your AI workflow is leaking dollars"
      - "Stop hand-prompting in chat — commit your prompts to Git"

The full schema is enforced via pydantic (workflows/social/_roster.py):

Field	Required	Description
roster_name	yes	Stable identifier — used as artifact filename prefix.
daily_budget_usd	no (default 10)	Hard cap. Preflight refuses to start if estimated spend exceeds this.
shared.cost_path	no (default local)	local / cheap-fal / premium-fal — picks the model preset.
shared.duration_secs	no (default 30)	Per-drop target duration.
characters[*].name	yes	Slug-safe character name.
characters[*].topics_pool	yes (for unattended runs)	Topic strings rotated by slot_offset.
characters[*].daily_video_count	no (default 1, max 10)	Drops per character per call.
characters[*].default_script_formula	no	Default narrative formula — see Script Formulas.
characters[*].allowed_formulas	no	Rotation pool overrides default_script_formula.

Cost paths

Path	Text	TTS	Avatar/Lipsync	Per-drop estimate
local	gemini-2.5-flash	Kokoro (local)	SadTalker (local)	~$0.025
cheap-fal	gemini-2.5-flash	FAL Kokoro	Kling Avatar v2	~$0.45
premium-fal	gemini-2.5-flash	ElevenLabs Turbo v2.5	Kling Avatar v2	~$0.55

Override individual models per-character (avatar_model, tts_model, text_model) or roster-wide (shared.avatar_model, etc.) when you need something off-preset. Resolution order: character > shared > preset.

Running it

Recommended: sharable run-spec file

The most ergonomic path is a self-contained run-spec file (plan 075):

fab-workflow --from-file examples/run-inputs/daily-drop-fabric-team.yaml

The file names both the workflow and the input — it’s a single artifact you can commit, hand to a teammate, or modify per-run. CLI --input k=v flags still layer on top:

fab-workflow --from-file examples/run-inputs/daily-drop-fabric-team.yaml \
    --input slot_offset=$(date +%j)

See the examples/run-inputs/ directory for starter files.

Default-lookup roster

fab-workflow social/daily-drop

Looks for ~/.fabric/characters.{yaml,yml,toml,json} only. Repo examples are intentionally not auto-discovered.

Explicit roster path

fab-workflow social/daily-drop --input roster_path=./my-roster.yaml
fab-workflow social/daily-drop --input roster_path=./my-roster.toml
fab-workflow social/daily-drop --input roster_path=./my-roster.json

Slot rotation across the day

fabric run social/daily-drop --input slot_offset=$(date +%j)

slot_offset rotates each character’s topics_pool and allowed_formulas so consecutive runs produce different topics. A common pattern: schedule the workflow once per day with slot_offset set to the day-of-year, so a 7-topic pool yields 7 unique days before repeating.

Pre-flight cost cap

Before any heavy stage runs, daily-drop sums the estimated cost across every (character × slot × stage) against cost.PROVIDER_COSTS. If the total exceeds daily_budget_usd, it raises:

RuntimeError: daily-drop preflight: estimated $24.50 exceeds
daily_budget_usd=$10.00.
Per-stage breakdown:
  builder:avatar: $9.0000
  builder:tts:    $1.5000
  ops:avatar:     $9.0000
  ...
Either raise daily_budget_usd, lower daily_video_count, or switch
shared.cost_path to 'local'.

This is a hard fail — the workflow refuses to begin. Cost overrun is the fastest way to ruin a dogfood run, so the preflight is intentionally strict.

Output shape

{
  "roster_name": "my-team",
  "drops": [
    {
      "character": "builder",
      "slot": 0,
      "topic": "Why your AI workflow is leaking dollars",
      "script_formula": "youre_doing_it_wrong",
      "video_path": "/.../my-team-builder-slot0.mp4",
      "asset_id": "asset_abc123",
      "script_text": "You're paying for tokens you don't actually need...",
      "error": ""
    }
  ],
  "total_estimated_usd": 1.35,
  "success_count": 3,
  "failure_count": 0
}

Failed drops carry error text instead of aborting the whole run — one flaky FAL call shouldn’t kill the rest of the day’s content. Check failure_count and the per-drop error fields after each run.

Dogfood playbook

If you’re running the official Fabric character roster, the operating playbook lives at specs/playbooks/01-fabric-character-roster.md. It documents the daily / weekly / monthly review cadence, what success looks like, and the discipline around revenue claims (we don’t make them without 6 months of audited data).