Screenshots & demo assets

Every image in docs/assets/ is a capture of the app running a real session against a real local model. Nothing here is a mockup, a composite, or a retouched frame. This document records what each asset shows, the alt text that must travel with it, the machine it was captured on, and the single command that remakes the whole set.

---

Capture environment

Recorded 2 October 2026 from the development build.

Hardware	Apple M1 Pro, 16 GB unified memory, macOS 14.8.2
Model	Qwen3 4B Instruct Q4_K_M (Apache-2.0), 2.5 GB on disk (2.3 GiB), loaded as a local GGUF
Engine	llama.cpp b9415 llama-server, --ctx-size 8192 --n-gpu-layers 99
App	./scripts/dev.sh — UI on http://127.0.0.1:7354, core on http://127.0.0.1:7355
Profile	a throwaway CONVSIM_DATA_ROOT with the six official packs seeded and no personal data
Scenario	Making the Case (Job Interview Basics), standard difficulty, seed 455, text input, NPC state meters enabled
Browser	Chromium (Playwright), 1280 CSS px wide at devicePixelRatio 2

Two things visible in the captures are the app reporting on this machine, not on the product, and are kept rather than staged away:

• The amber performance advisories. The app raises two, and both are in frame here: “NPC response is slow” on 03-conversation.png, where the first token took 9.6 s, and “Full NPC response is very slow” at the end of the hero recording, where the whole reply took 14.0 s. The product budgets 2.5 s to first token and 10 s to a full response on the recommended tier (packages/shared/src/types/metrics.ts); a 2021 laptop running a 4B model at a ~1,500-token prompt lands well past both and the app says so. Capturing on faster hardware removes the banners; editing them out would be a lie about the thing the README is selling.
• STT and TTS reading “Not installed” on the home screen. Voice needs whisper.cpp and a local Kokoro server, which are optional runtimes a text-only install does not have (see the install guide). The app runs fully in text mode without them.

No real people, faces, voices, or private data appear in any asset. Every NPC, company and transcript line is fictional scenario content or model output responding to it.

---

Re-capturing

# 1. a local model installed and selected, engine up, dev services running
./scripts/dev.sh

# 2. Playwright + Chromium (not a repo dependency — only needed to re-capture)
npm install playwright && npx playwright install chromium

# 2b. ffmpeg, for the hero recording (not needed with --skip-hero, or with
#     an --only list that leaves `hero` off)
brew install ffmpeg

# 3. drive a real playthrough and write every asset under docs/assets/
NODE_PATH="$PWD/node_modules" node scripts/capture-screenshots.mjs

# 4. remake the 1x web copies the two sites render (the script prints this
#    command itself whenever it rewrote a screenshot)
for f in docs/assets/screenshots/*.png; do
  for d in website/static/images/screenshots docs-site/public/images/screenshots; do
    magick "$f" -resize 1280x -colors 256 -strip "$d/$(basename "$f")"
  done
done

scripts/capture-screenshots.mjs plays the scenario through the real UI — library search, scenario setup, four typed turns, end session, debrief — then captures the Creator Workbench and Model Manager, and encodes the hero recording from the last exchange with ffmpeg. It refuses to run when the active runtime is fake or scripted, so a capture can never quietly ship canned text as gameplay.

Useful flags: --skip-hero (screenshots only, no recording) and --only=03,06 (still plays the session, but writes only the named outputs — a screen number, or hero for the recording; anything left off the list keeps the file already committed). A name the script does not write — or a pair of flags that cancel out, like --skip-hero --only=hero — is rejected before the playthrough starts, so neither can cost a run. The scenario, player turns and seed are constants at the top of the file.

A run that loses a screen keeps going — the other outputs are worth more than the one that failed — but it exits non-zero and names what it missed. Those files still hold the previous capture, so check the exit status before committing: a partial run is otherwise indistinguishable from a full one. If it is the hero encode that failed, the raw .webm is left in its temp directory and the path printed, so the segment can be re-encoded by hand instead of replayed.

---

Demo asset (README hero)

File	docs/assets/demo.gif (3.1 MB) · docs/assets/demo.mp4 (0.6 MB fallback)
Used in	README.md — the GIF. The MP4 is the same segment kept as a video source for the Steam trailer and store page (publishing/STEAM_ASSETS_SPEC.md); nothing embeds it today.
License	CC0-1.0
Shows	25 s of one real turn: the player types a reply, the local model answers as the NPC, and the state meters under the transcript move as the turn resolves
Fictional content	NPC “Elena Vasquez” (Head of Product, stretch-hire interview) — no real person

Alt text used in README:

One turn of “Making the Case” in Conversation Simulator: the player types a reply into the composer, the local model answers as the NPC with event flags beneath it, and the NPC state meters below the transcript — trust, patience, rapport, openness, objective_progress, conviction, preparation_score — update as the turn resolves. An amber advisory above the transcript reports that the full response took 14.0 seconds on this machine.

---

Screenshots (docs/assets/screenshots/)

Six PNGs, 2560 px wide (1280 CSS px at 2×), heights fitted to each screen. License: CC0-1.0 (public domain).

The alt text recorded below is what README.md carries verbatim. The 1× web copies and the authoring guide carry condensed variants of the same description — shorter, but a subset of what is recorded here, never a different claim.

01-home.png — Home screen

What it shows: The home screen on a ready text-only install. Navigation links, the “Your training” card summarising the three sessions played in this capture profile (streak, strongest and weakest rubric dimension, last session’s delta), a three-scenario training plan, and the Status panel: Local runtime ready, the loaded model named, voice runtimes not installed, no network required, six packs installed.

Alt text:

Conversation Simulator home screen. Links run down the page — Start a scenario, Create / edit a scenario, Install model, Import pack, Creator workbench guide, Read docs — above a “Your training” card summarising the three sessions played in this profile, and a three-scenario training plan. The Status panel reports Local runtime: Ready, LLM: Qwen3 4B Instruct Q4_K_M, STT and TTS: Not installed, Network required to play: No, and Packs: 6 installed.

---

02-scenario-library.png — Scenario Library

What it shows: The library with the live search box filtered to interview, the rating, language, difficulty and tag filters beside it, and the Job Interview Basics pack expanded to its four scenario cards with metadata chips and Launch buttons.

Alt text:

Scenario Library filtered by the search word “interview”: four scenarios in one pack. The Job Interview Basics pack lists Making the Case, The Behavioral Interview, The Executive Gauntlet and The Foreman’s Interview. Each card carries a summary, chips for content rating, player role, length, language, difficulty and tags, and a blue Launch button.

---

03-conversation.png — Conversation (mid-session with state meters)

What it shows: Turn 10 of the session. NPC panel (emotion “Impressed”, status “Listening”), the scene card, the performance advisory this machine earned, four turns of transcript with per-turn event flags, the NPC state variable meters — the five visible baseline variables plus the two the scenario adds, conviction and preparation_score — the accumulated event-flag strip, and the text composer. The hidden variables (baseline pressure, the scenario’s stretch_potential and gap_acknowledged) have no meter here; they surface only in the debrief telemetry on 04-debrief.png.

Alt text:

The conversation screen mid-session in “Making the Case”. The NPC panel reads Impressed and Listening beside a scene card describing the stretch-hire interview, and an amber advisory notes that the first token took 9.6 seconds on this machine. The transcript alternates the player’s turns in violet with the NPC’s replies in green, each NPC turn tagged with event flags such as honesty_demonstrated. Below the transcript, NPC state variable meters read trust 95, patience 75, rapport 50, openness 80, objective_progress 0, conviction 35 and preparation_score 55.

---

04-debrief.png — Session Debrief

What it shows: The debrief generated for that session: overall score and outcome badge, a model-written summary, the rubric scorecard, the telemetry panel with conversation metrics and a sparkline per state variable, then strengths — each citing the turn it came from — and areas for improvement. The frame ends at the improvements section; key moments, replay suggestions and the transcript continue below.

Alt text:

Session Debrief scoring the session 53 out of 100 with a “Player Exit” outcome badge and four turns. A written summary is followed by a six-dimension scorecard — Experience Gap Acknowledged 59, Transferable Skills Demonstrated 52, Learning Velocity Demonstrated 52, Specific Evidence Transferable Skills 52, Honesty About Experience Gap 51, Transferable Evidence 52 — then a Telemetry panel showing talk ratio, words per turn, question counts, response latency and a sparkline for each state meter. Strengths are listed in green, each citing the turn it came from, and areas for improvement in amber.

---

05-creator-workbench.png — Creator Workbench

What it shows: The workbench after using “Create local copy to edit” on an official pack: the green validation banner, the pack list with six read-only official packs and one editable local-dev copy, that copy’s file tree, and the YAML editor open on a scenario file with the YAML/Form toggle, Save, and Export .zip. Both columns are viewport-height panels that scroll on their own, so the frame shows the top of the file tree and the top of the open file rather than all of either.

Alt text:

Creator Workbench. A green “Pack is valid” banner sits above a two-column layout. On the left, the pack list — six official packs and one editable local-dev copy of Job Interview Basics — sits above that copy’s file tree, with scenarios/stretch_role_interview.yaml selected. On the right, a YAML editor shows the file — schema_version, scenario_id, title, summary, player_role, and the npc, scene and rubric references — with YAML and Form editor toggles, a Save button and an Export .zip button.

---

06-model-manager.png — Model Manager

What it shows: The model surface as it actually ships: three ways to get an engine running — install the recommended registry model, pick one from a local Ollama, or point at a GGUF file already on disk — with size, licence, VRAM requirement and an expected speed class on the recommended card.

Alt text:

Model Manager. “Set up your model” offers three cards: install the recommended Qwen3 4B Instruct Q4_K_M (2.5 GB, Apache-2.0, requires 4 GB VRAM, tagged Fast with an expected 0.8 to 2.4 second time to first token), use a model already installed in local Ollama, or point the app at a GGUF file already on the machine.

---

Capture checklist

Completed for the current set; re-run it whenever the assets are replaced.

• Recorded from a development build at http://127.0.0.1:7354 with fictional NPC data and a local Qwen3 model — no real-person images, audio, or data.
• PNG screenshots exported at 2× (2560 px wide), heights fitted per screen.
• Hero recording exported under 5 MB (ffmpeg palette optimisation), with an MP4 fallback alongside it.
• Files placed in docs/assets/screenshots/, and derived web copies placed in website/static/images/screenshots/ and docs-site/public/images/screenshots/: magick <png> -resize 1280x -colors 256 -strip <dest> (about a quarter the bytes, still sharp at the sizes those pages render).
• The old hero SVG and the six mockup SVGs are gone from the tree; demo.gif and the six PNGs take their place.
• README.md image tags updated with the real filenames and the alt text above.
• Manual content and safety review: no real faces, no sensitive data, no identifiable voices, nothing above PG-13.
• File sizes verified for GitHub rendering (largest screenshot 0.5 MB, hero 3.1 MB).
• NOTICE updated for the captured assets (CC0-1.0).

---

Licence summary

Asset	Licence	Notes
docs/assets/demo.gif, docs/assets/demo.mp4	CC0-1.0	Captured from a real session; fictional NPCs only
docs/assets/screenshots/*.png	CC0-1.0	Captured from a real session; fictional NPCs only
Site copies under website/static/images/ and docs-site/public/images/	CC0-1.0	The same captures at 1×, 256 colours, for page weight
Future re-captures	CC0-1.0	Keep the same licence, and re-run the checklist above