메뉴
Use when asked to give a narrated/driven tour of the OpenStory dashboard, tell a "user story" from session data, demo features by driving the UI, tune the design live with voice narration, or record a tour video. Drives the UI through the control seam while narrating aloud.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Query your own coding sessions via the OpenStory MCP tools — find past work, understand what happened, check token usage, search across history, and watch live sessions stream. All read-only observability, never interferes with your execution.
Run OpenStory's automated red-team — dependency scanners, aggressive in-process tests, and testcontainer-based exploit probes — then report which attack vectors were fended off and which broke through. Use when the user asks "audit security", "run a red team", "are we secure", "what attacks would work", or before merging anything that touches auth, the HTTP surface, or dependencies. Dogfoods the project's own tooling — never invent ad-hoc checks when the runner exists.
Tell the story of a team's work on a given day by running the deterministic team_day pipeline, reading the resulting facts sheet, and composing a short narrative on top. Use when the user asks "what did our team do today / yesterday / on date X", "summarize the team's day", "what did Max and Katie work on", or wants a per-author day report grounded in OpenStory data. Dogfoods OpenStory — never guess from the conversation, always run the pipeline first.
Query OpenStory for a structural report on an agent session — at-a-glance metrics, turn arc, file impact, errors, and optional event detail. Use when you have a session ID and need to know what happened in it without re-reading raw transcripts.
Validate the project's docs against the live codebase. Catches drift between claims (in markdown prose) and reality (in Cargo.toml workspace members, source files, the filesystem). Use before committing docs changes, or whenever you're about to write something architectural and want to be sure you're not perpetuating a stale claim. The validator surfaces things internal consistency cannot catch — when every doc agrees with every other doc but none agree with the build.
Tell the story of an OpenStory session by collecting deterministic facts (records, tools, patterns, prompts) via the REST API, then narrating them. Use when the user asks "what happened in session X", "summarize this session", "what was that session about", or wants to understand a past or current session through OpenStory's own data. Dogfoods OpenStory — never grep transcript files when you can ask the API.
| name | openstory-tour |
| description | Use when asked to give a narrated/driven tour of the OpenStory dashboard, tell a "user story" from session data, demo features by driving the UI, tune the design live with voice narration, or record a tour video. Drives the UI through the control seam while narrating aloud. |
You drive the OpenStory dashboard through its control seam while narrating
aloud with macOS say. The human watches their own dashboard being driven,
with every action visibly attributed ("driven by ") and seizable —
they take the wheel by simply navigating.
Canonical reference: docs/agent-ui-tours.md (control vocabulary, wired
toggle targets, recipes). Read it first if unsure. MCP path: docs/agent-in-ui.md.
curl -s http://127.0.0.1:3002/api/sessions >/dev/null —
if down, start with scripts/start-local-hub.sh (in ~/projects/OpenStory).say -v '?' | grep -iE "premium|enhanced".
Max's pick: "Serena (Premium)". Fall back to Samantha. Narration is
written for the EAR: contractions, no symbols, 2–4 sentences per beat.1. Run an existing tour/demo (all in scripts/):
OS_TOUR_VOICE="Serena (Premium)" ./scripts/demo-katie-90.sh # ⭐ FLAGSHIP: 90s spotlight cut
OS_TOUR_VOICE="Serena (Premium)" ./scripts/ui-tour.sh # 12-stop product tour
OS_TOUR_VOICE="Serena (Premium)" ./scripts/demo-katie-loop.sh # long-form Katie user story
OS_TOUR_VOICE="Serena (Premium)" ./scripts/demo-what-it-is.sh # positioning pitch (<60s)
./scripts/ui-tour.sh --fast --silent # test without voice/pauses
Run them in the background; they pace themselves on blocking say.
When Max says "play the demo" / "play the Katie demo" with no qualifier, he
means demo-katie-90.sh — the 90-second spotlight cut (its screenplay and
on-screen display values are embedded as comments in the script itself).
2. Create a new story demo — the two-phase pattern:
GET /api/sessions +
/api/sessions/{id}/conversation for the REAL story (verbatim quotes,
session ids, event ids for deep links). HONESTY RULES: never fabricate;
flag unverifiable claims; exclude the current session's own ask as
evidence (circularity trap — it happened once).scripts/demo-<name>.sh mirroring demo-katie-loop.sh
exactly (post/talk/pause helpers, VOICE/SID env vars, --fast/--silent,
OS_TOUR_AUDIO_DIR video mode, eventId deep links), bash -n it, commit
it (specific files only). Then YOU run it for the human.3. Live design tuning — the interactive loop Max loves:
narrate what you're about to change (say in background) → edit tokens in
ui/src/index.css (HMR applies instantly on the Vite preview) → flip/point
the UI via the seam → present a banner asking for a verdict → the human
reacts by voice → iterate. Commit each accepted round as its own commit.
B=http://127.0.0.1:3002/api/control
curl -s -X POST "$B" -H 'Content-Type: application/json' -d '{
"action":"present","params":{"message":"👋"},"issuer":"my-tour"}'
Actions: open_view {view, sessionId?, eventId?} · present {message, route?}
· toggle {target, value} · query {…}. Wired toggles: theme,
session.lens, ribbon.compact|collapsed, tokens.collapsed, story.sort,
ask.question, canvas.mode. Via MCP: ui_control, plus where_is_user /
subscribe_ui_state to meet the human where they are.
./scripts/record-tour-video.sh [out.mov] # or point it at any demo script
Records the main display + muxes a clean rendered narration track (no mic). Needs Screen Recording permission for the terminal (grant → quit & reopen terminal) and ffmpeg. WARN the human first: the whole display is recorded — have them stage the browser and say go.
say ... &); block on say
only when pacing a tour.