Curated from kaomoji-collection (MIT, 41k faces) rather than bulk-imported:
candidates were filtered to glyphs present in JetBrains Mono Nerd Font, no
combining marks or wide cells, and no wider than her widest existing face,
then rendered in the real font and picked by eye from the sheet.

calm gains a sleepy ( ˆωˆ ), thinking a deadpan (¯ . ¯), bright the victory
(^_^)v, and squint both a shrug ╮(╯-╰)╭ and a smug ( ¬‿¬ ) — the two that
add gestures she didn't have rather than respacing ones she did.

working is deliberately untouched: those four frames are Kirby's dance
(right, left, up, down) advanced one step per tool call, so a fifth entry
would break the cycle. Noted in the comment.

Yield was low on purpose — most of the 41k are near-duplicates of faces she
already has, and a face nobody has seen her make reads as a different
character. voice.md's table is synced.

voice.md and qt314.md both cited scripts/qt314-*.py as the canonical
mood→face mapping. Those predate apps/agent's Companion and hold a subset:
they miss the squint mood entirely — including (－‸ლ), which is already in
the tweet corpus — and carry fewer variants per mood.

Point both at apps/agent/src/features/profile/companion.tsx, and record why
the set stays small: her mood is derived from the feed rather than chosen by
a model, so a face is a readout of state, not decoration. Also distinguish
the product set (state) from her wider expressive range in copy.

The skill had grown into a lessons journal — 236 lines dominated by a
numbered list of video gotchas (with two #12s), with the copy, brand, and
card recipes scattered through it. Reorganize around the asset you want to
make.

SKILL.md is now a router: preflight, an asset table, voice and brand in a
paragraph each. The depth moves into references/ — voice, brand, video,
cards, qt314 — and runnable files move to scripts/ and assets/ per the
Agent Skills layout.

- references/voice.md is new and built from the real @getvibekit corpus:
  the vibekit/qt314 speaker split, the sigil vocabulary, the kaomoji mood
  mapping (cross-checked against qt314-*.py so copy and animation agree),
  the thread format, and rejected drafts for the failure modes.
- assets/card.css holds the shared card shell and the palette as CSS
  variables; both card scripts source it instead of pasting hex.
- make-card.sh gains PNG=1, so OG images and quote cards come off the same
  template as the video title cards and cannot drift from them.
- cards.md adds TUI stills and the GIF palettegen recipe.
- assemble.md is gone: it billed itself as the exact commands while
  building title cards with ffmpeg drawtext, which SKILL.md explicitly
  banned. Whichever file was read first won.
- segA/segB.tape no longer cd to a path that moved two renames ago, and
  carry the brand Set Theme.
- demos/apps-call.md is marked archived — read for beat structure, not
  re-run against app IDs from that day.

Verified both renderers still produce correct output after the CSS
extraction, PNG=1 writes 1200x630, and bundle-skills still skips the
internal skill.

Claude-Session: https://claude.ai/code/session_01WigrFXMAuiDj2RG714GnQu

Why VibeKit says the thing plainly: agents will act on chains; nobody in the
loop holds the key but the person's wallet, and the toolkit is shaped so that
stays true on every surface. How it works is the mental model with the
adapter table. Examples is one table over the runnable files. Run an action
walks examples/action.ts stage by stage; Compose a server is the reference
server's routing with its three decisions named. Index reorganised by who
you are; Explanation gets its own sidebar group.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

A Bun server (50 lines) mounts the agent handler, the action routes, and
REST over every tool, and serves the page. A React page (150 lines) streams
a turn, renders results with the vibekit add components by view id, and
walks a draft to approval — resting at approved, honestly, since no wallet
is wired. The component copies are drift-checked against the package.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

mcp: createMcpServer, createMcpServerFactory, createMcpHttpHandler,
serveMcpStdio. rest: createRestHandler, POST …/tools/<name> (it runs actions
too). The root exports the contract only; the compose engine is
@initlabs/vibekit/compose; agent exports the loop, the handler, and the
prompt — provider plumbing is agent/providers, the config file is
agent/config. The packed-consumer verifies catch up with phase 1's renames.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

createActionHost(deployment) and signDraftWith join actions; the deployment
read host and the block tail join views; createHost(network) — the stock
combined host the TUI and web server build on — and the Explorer agent
(tools, prompt, plugins, explain tool) are preset, where opinions live.
Five nouns again: query/action, agent, pay, views, preset; the rest are
adapters.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

views no longer re-exports the actions surface (every name had two import
paths); the fixtures and the sample host move to @initlabs/vibekit/views/sample;
ReadHost, classifyInput, routeInput, InputRoute, formatTime; record() is
createRecord; actionNextEventKinds is nextActionEvents. Imports in both apps
and the tests re-pointed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

draftDataSchema/DraftData, simulationDataSchema/SimulationData, stageEventSchema/
StageEvent, createStageEvent, <stage>StageEventSchema, actionIntentSchema/
ActionIntent, actionResultSchema. Rename only.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

createActionRoutes + createRemoteActionHost (actions/http): an ActionHost
over HTTP with server-side byte verification, lifted from the web agent's
explorer route and remote host. createReadHost + recordForToolCall (views):
the lookup surface over any callTool, shared by the live host and the
browser. The REST response carries the tool's view id, so a browser can wrap
any query into its record client-side. The web agent's 335-line explorer RPC
is now the action routes plus the query handler (house-billed), and the
enrichment/NFD side hosts it needed are gone.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Everything the two apps shared beyond the tool contract now ships in the
published package: views/ (browser-safe records, view models, the graph
builder, bridge, format, input, sample fixtures) and live/ (the deployment
host, block tail, enrichment, the Explorer agent). Tests and recorded
fixtures move with them; the block-tail record types move to views so views
never imports live. Apps import @initlabs/vibekit/views and /live; the
browser boundary test allows views beside actions.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

A generated tools reference (bun run docs:reference walks the tool contract:
29 queries, 16 actions, 5 plugins — never hand-edited), one tutorial (build
an agent: handler, components, signing), four short how-tos pointing at the
runnable examples (REST + x402, swap the model, swap the signer, components),
one explanation (queries and actions), the README reframed around the
surface, and a paragraph in the build-on-vibekit skill. Deliberately thin:
the API is moving; the reference is what stays current.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

packages/vibekit/components is the source: Companion, useAction (now free of
the explorer), ActionApproval, ToolResult, Transaction, Asset/AssetList,
Account, a digit-math format.ts, and one base stylesheet of class hooks.
Every component's props are a tool's output type, so the same data from an
agent event, a REST response, or an MCP result renders unchanged. The CLI
bundles them at build time (like skills); add --list/--dir/--force. The
agent app's copies are checked against the source in test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>