cli-demo-recorder

CLI Demo Recorder

Record polished demo videos for CLI tools — either as direct subprocess recordings (CLI tools, no AI session) or real TUI recordings (interactive AI sessions via tmux).

Use this skill when: A demo GIF/video is needed for a CLI tool, plugin, or terminal application. Invoke with: /cli-demo-recorder or "Help me record a demo for my CLI tool"

---

Choose Your Pathway [Both]

Pick the pathway based on whether the tool has an interactive TUI session. Wrong choice → recording captures nothing useful.

Tool type	Interactive TUI?	Uses AI/LLM?	Correct pathway
Pure CLI (aise, git, curl)	No	No	CLI: harness IS the recording
CLI + AI session (claude -p)	No TUI	Yes	CLI: verify output is useful
Plugin/hook for TUI tool	Via TUI	Yes	TUI live: tmux + pane
Plugin with hook-only acts	Via hook	No	TUI scripted: run_hook() + --play
Spawns interactive TUI	Yes	Maybe	TUI live: drive TUI via tmux
Batch/config tool	No	No	CLI: capture_output=False

WARNING: Using subprocess.run(capture_output=True) for a CLI demo silences recording entirely — asciinema captures nothing. See CLI pathway for the correct pattern.

---

How It Works [Both]

Phase 1: Plan (15–30 min)

1. Read the tool's docs first — before writing a single act
2. Choose 5–7 features that are visible, immediate, and self-explanatory to newcomers
3. Skip invisible features (background daemons, auto-save without visible output)
4. Choose your pathway (see table above) — this determines the entire harness design

Phase 2: Build the Harness (30–60 min)

• CLI: Python script that calls subprocess.run(..., capture_output=False). asciinema records python test_demo.py --run-acts.
• TUI scripted: Python script that calls run_hook() directly for each act. asciinema records python test_demo.py --play.
• TUI live: Python script that creates a tmux session, sends prompts via send-keys, and asciinema attaches to that session.

Phase 3: Record and Verify (10–30 min)

• CLI: python tests/test_demo.py --record → checks cast text fragments
• TUI: python tests/test_demo.py --record → parse JSONL for tool calls

Phase 4: Convert [Both]

agg demo.cast demo.gif \
    --theme dracula \
    --font-size 14 \        # 14-16; smaller fits more content
    --renderer fontdue \    # vector-quality anti-aliased text
    --speed 0.75 \          # 0.75x — readable without pausing
    --idle-time-limit 10    # 10s — preserves full banner display

# MP4: 4-strategy fallback (best compression first)
# Strategy 1: libx265 HEVC (tune=animation — ~50% smaller than libx264 at same quality)
ffmpeg -y -i demo.gif -movflags faststart \
    -vf "scale=trunc(iw/2)*2:trunc(ih/2)*2" \
    -c:v libx265 -preset slow -crf 28 -tune animation \
    -pix_fmt yuv420p -tag:v hvc1 demo.mp4 2>/dev/null \
  || ffmpeg -y -i demo.gif -movflags faststart \
    -vf "scale=trunc(iw/2)*2:trunc(ih/2)*2" \
    -c:v libx264 -preset slow -crf 28 -tune animation \
    -pix_fmt yuv420p demo.mp4 2>/dev/null \
  || ffmpeg -y -i demo.gif -movflags faststart \
    -vf "scale=trunc(iw/2)*2:trunc(ih/2)*2" \
    -c:v h264_videotoolbox -q:v 65 -pix_fmt yuv420p -color_range tv demo.mp4 2>/dev/null \
  || ffmpeg -y -i demo.gif -vf "scale=trunc(iw/2)*2:trunc(ih/2)*2" \
    -pix_fmt yuv420p demo.mp4

Important: Convert GIF → MP4 (not cast → MP4). The GIF is already processed; going cast → MP4 directly misses the speed/idle adjustments from agg. Use tune=animation (not tune=fast) — terminal recordings have flat colors and sharp edges that match animation compression.

Total: ~75–120 minutes for a polished, verified demo.

---

CLI Pathway [CLI Only]

Architecture: Harness IS the Recording

For CLI tools, the Python harness runs as the command that asciinema records. No tmux, no pane attachment.

asciinema rec demo.cast --command "python test_demo.py --run-acts"
                                                        ↑ harness runs here
  → harness types $ prompt, runs subprocess with capture_output=False
  → asciinema captures all stdout from the process
  → agg demo.cast demo.gif

# ✅ CORRECT — output flows to terminal, asciinema captures it
subprocess.run("mytool subcommand", shell=True, capture_output=False, env=DEMO_ENV)

# ❌ WRONG for CLI demos — captures output into Python, asciinema sees nothing
result = subprocess.run(["mytool", "subcommand"], capture_output=True)

Core Helpers [CLI Only]

_TIMED = False  # True only inside --run-acts (recording mode); _DEMO_WITH_TIMING also common

def pause(seconds: float) -> None:
    """No-op in pytest; sleeps during recording. Errors 1+2: timing ≠ spacing."""
    if _TIMED:
        time.sleep(seconds)

def _type(text: str, delay: float = 0.04) -> None:
    if _TIMED:
        for ch in text:
            sys.stdout.write(ch); sys.stdout.flush(); time.sleep(delay)
    else:
        sys.stdout.write(text); sys.stdout.flush()

def _run(cmd: str) -> None:
    """Show typed $ prompt, then run command.
    capture_output=False is CRITICAL — output must flow to terminal.
    """
    _type(f"\n\033[1;32m$\033[0m ", delay=0)
    _type(cmd + "\n", delay=0.045)
    pause(0.3)
    subprocess.run(cmd, env=DEMO_ENV, shell=True, capture_output=False, text=True)

def section(title: str) -> None:
    """Visual section divider between acts.

    3 newlines BEFORE bar = visual gap from previous act (change this for spacing).
    pause() durations = reading time (change separately for timing).
    These are INDEPENDENT knobs — do not conflate them.

    bar_len = max(68, len(title) + 6) prevents bars shorter than title.
    """
    bar_len = max(68, len(title) + 6)
    bar = "─" * bar_len
    sys.stdout.write(f"\n\n\n\033[90m{bar}\033[0m\n")
    sys.stdout.write(f"\033[1;96m  {title}\033[0m\n")
    sys.stdout.write(f"\033[90m{bar}\033[0m\n")
    sys.stdout.flush()

Intro Banner [CLI Only]

For CLI tools, the banner is a Python string printed directly to stdout:

def banner() -> None:
    W = 68  # compute padding on PLAIN text only — no ANSI codes inside len() math
    def row(text: str = "", style: str = "") -> str:
        content = (" " + text).ljust(W)  # W visible chars; no ANSI in length
        return f"\033[90m  ║\033[0m{style}{content}\033[0m\033[90m║\033[0m"
    # ❌ WRONG — ANSI codes inflate len(), misalign padding:
    # bad = f"\033[1m{text}\033[0m".ljust(W)
    lines = [
        f"\033[90m  ╔{'═'*W}╗\033[0m",
        row("mytool — tagline here", "\033[1;96m"),
        row(),
        row("This demo shows:", "\033[90m"),
        row("  1. Feature one", "\033[90m"),
        row("  2. Feature two", "\033[90m"),
        f"\033[90m  ╚{'═'*W}╝\033[0m",
    ]
    print("\n" + "\n".join(lines) + "\n")

Privacy Isolation [CLI Only]

# DEMO_DATA_DIR: committed synthetic fixtures (no real user data)
# TOOL_ISOLATION_VAR: env var that redirects the tool's data reads
# Examples: CLAUDE_CONFIG_DIR (aise), XDG_DATA_HOME, APP_DATA_DIR
DEMO_DATA_DIR = Path(__file__).parent / "tool-demo"
DEMO_ENV = {**os.environ, "TOOL_ISOLATION_VAR": str(DEMO_DATA_DIR)}

Date-Shifting Fixtures [CLI Only]

Required when demo acts use --since Nd, --after DATE, or any time-relative filter.

def create_dated_demo_dir() -> Path:
    """Copy DEMO_DATA_DIR to temp dir with timestamps shifted to near today.
    Without this: fixtures from months ago → 0 results for --since 3d.
    The committed fixture files are NEVER modified — only the temp copy is shifted.
    Adapt _TS_RE and shift logic to match your tool's timestamp format.
    """
    _TS_RE = re.compile(r'"timestamp":\s*"(\d{4}-\d{2}-\d{2}T[^"]+)"')
    # find max timestamp in fixtures, compute delta to (today - 1 day)
    tmp = Path(tempfile.mkdtemp(prefix="tool-demo-dated-"))
    shutil.copytree(DEMO_DATA_DIR, tmp / "data")
    for f in (tmp / "data").rglob("*.jsonl"):
        f.write_text(_TS_RE.sub(lambda m: shift_ts(m, delta), f.read_text()))
    return tmp

Recording [CLI Only]

def record(cast_file: Path) -> None:
    dated_dir = create_dated_demo_dir()
    record_env = {**os.environ, "TOOL_ISOLATION_VAR": str(dated_dir)}
    try:
        subprocess.run([
            asciinema, "rec", str(cast_file),
            "--command", f"{sys.executable} {__file__} --run-acts",
            "--window-size", "160x48",   # 100x35 for narrow tools
            "--capture-env", "TERM,COLORTERM,TOOL_ISOLATION_VAR",
        ], env=record_env, check=True)
    finally:
        shutil.rmtree(dated_dir, ignore_errors=True)

Initial Frame Cleanup [Both]

Asciinema records shell initialization (prompt, env vars, RC files) before the harness starts. This produces a messy first frame with shell artifacts instead of a clean banner. Fix this with a post-recording trim that finds the banner marker, walks back to the preceding clear-screen escape, drops all prior events, and rebases timestamps to t=0.

def trim_cast_to_banner(cast_file: Path, banner_marker: str = "mytool") -> None:
    """Trim shell init events so the banner is the first visible frame.

    Modifies cast_file in place. Preserves the JSON header line.
    """
    lines = cast_file.read_text().splitlines()
    if len(lines) < 2:
        return

    header = lines[0]   # JSON header — always line 1
    events = lines[1:]

    # Find the first event containing the banner marker text.
    banner_idx = None
    for i, line in enumerate(events):
        if banner_marker in line:
            banner_idx = i
            break

    if banner_idx is None:
        print(f"[trim] marker {banner_marker!r} not found — skipping trim")
        return

    # Walk back to the clear-screen escape just before the banner.
    clear_idx = banner_idx
    for i in range(banner_idx - 1, -1, -1):
        if "\\u001b[H\\u001b[2J" in events[i] or "\\033[H\\033[2J" in events[i]:
            clear_idx = i
            break

    kept = events[clear_idx:]
    if not kept:
        return

    # Rebase timestamps: first kept event becomes t=0.
    first_ts = json.loads(kept[0])[0]
    rebased = []
    for line in kept:
        evt = json.loads(line)
        evt[0] = round(evt[0] - first_ts, 6)
        rebased.append(json.dumps(evt))

    cast_file.write_text(header + "\n" + "\n".join(rebased) + "\n")
    trimmed = len(events) - len(kept)
    print(f"[trim] Removed {trimmed} events before banner (kept {len(kept)})")

Call after recording, before conversion: trim_cast_to_banner(CAST_FILE, banner_marker="mytool_name"). For Rust and TypeScript implementations, see references/rust-demos.md and references/typescript-demos.md.

Write Acts First, Then Matching Tests [CLI Only]

Root cause of test/demo drift: writing tests independently of demo acts.

✅ Correct workflow:
1. Finalize run_demo_acts() — fix the EXACT command string for each act
2. Write TestDemoFree using those SAME command strings, verbatim
3. Never add --format, --limit, or extra flags to tests that aren't in the demo

❌ Wrong workflow:
  demo act:  mytool messages search keyword --context 1
  test:      mytool messages search keyword --format plain --full-uuid
  → test passes, demo shows different output, test proves nothing about the demo

Rule: if a test needs a different flag to verify output, either (a) add that flag to the demo act too, or (b) verify a property that the demo's actual output satisfies.

Verification [CLI Only]

Check that expected text appears in the cast file:

def verify_recording(cast_file: Path) -> bool:
    content = cast_file.read_text()
    checks = [
        ("Sessions:",      "Act 1: stats label present"),
        ("authentication", "Act 4: search result shows keyword"),
    ]
    return all(fragment in content for fragment, _ in checks)

Deterministic Fixture IDs [CLI Only]

# IDs follow a recognizable pattern — easy to spot in recordings
_S1 = str(uuid.UUID("cafe0001-cafe-cafe-cafe-000000000001"))
# Stable tool-call IDs across runs:
_id = hashlib.md5(f"{session_id}{timestamp}{path}".encode()).hexdigest()[:8]

Closing Message [CLI Only]

# Always read install command from pyproject.toml or README — never guess
sys.stdout.write(
    "\033[1;32m  ══════════════════════════════════════\033[0m\n"
    "\033[1;32m  ✓  Demo complete — {tool name and tagline}\033[0m\n"
    "\033[1;32m  ══════════════════════════════════════\033[0m\n"
    "\n"
    "  Install:  {exact command from pyproject.toml/README}\n"
    "\n"
)

---

TUI Pathway [TUI Only]

Architecture: Why tmux + Real TUI

The first working version of a demo typically shows simulated output — the harness calls the hook directly and prints what the hook would say. This is wrong for two reasons:

1. It's not what users see. Users see the full interactive TUI: tool call blocks, the AI's conversational response, the block message embedded in that response.
2. The hook may not have fired at all. This is unknowable without recording actual tool calls.

Correct architecture (TUI live):

Python harness
  → creates tmux session
  → asciinema attaches: asciinema rec --command "tmux attach-session -t SESSION"
  → background thread: tmux send-keys "claude --dangerously-skip-permissions" Enter
  → polls trust/safety dialog → auto-confirms with Enter
  → polls for TUI input prompt (❯) — requires 3 consecutive idles
  → tmux send-keys "{prompt}" Enter
  → waits 1.5s, then polls until 3 consecutive idles
  → next act...
  → claude exits → tmux session ends → asciinema finishes

Background thread pattern (TUI live recording):

demo_t = threading.Thread(target=_demo_thread, daemon=True)
demo_t.start()
# asciinema in foreground records the tmux session
asciinema_proc = subprocess.Popen([asciinema, "rec", cast_file,
    "--command", f"tmux attach-session -t {session.session_name}", ...])
demo_t.join(timeout=600)
time.sleep(1)
asciinema_proc.terminate()

Wrong for TUI tools:

# ❌ Headless subprocess — output not shown in TUI, not recorded by asciinema
result = subprocess.run(["mytool", "--do-thing"], capture_output=True)

# ❌ claude -p stays on command line, no TUI, asciinema captures nothing useful
subprocess.run(["claude", "-p", "delete this file"])

TUI Scripted Pathway [TUI Only]

For tools with hook-level acts (no Claude TUI needed), use scripted mode:

• Acts call run_hook() subprocess directly
• No API key required, $0.00 cost
• asciinema records python test_demo.py --play as the command
• Output is formatted to explain features to newcomers

# Scripted recording: asciinema records this process directly
asciinema_cmd = [asciinema, "rec", cast_file, "--overwrite",
    "--command", f"{sys.executable} {Path(__file__)} --play",
    "--idle-time-limit", "3"]  # shorter idle limit for scripted mode
subprocess.run(asciinema_cmd)

Setting Up the Work Directory [TUI Only]

Use a short, fixed path — not tempfile.TemporaryDirectory(). Long paths appear in every tool call and look unprofessional.

# ❌ Long noisy path in every tool call:
work_dir = Path(tempfile.mkdtemp())
# → /private/var/folders/9f/jr_p974d3j318tmvrfkjl55w0000gp/T/tmpAbcDef

# ✅ Clean, short path:
import os
work_dir = Path(f"/tmp/mytool-demo-{os.getpid()}")
work_dir.mkdir(parents=True, exist_ok=True)

Mock Git Repo Setup [TUI Only]

def setup_mock_project(work_dir: Path) -> None:
    # Use GIT_CONFIG_NOSYSTEM to prevent system git config bleed-in
    env = {**os.environ, "GIT_CONFIG_NOSYSTEM": "1"}
    run = lambda cmd: subprocess.run(cmd, cwd=work_dir, capture_output=True, env=env)

    run(["git", "init"])
    run(["git", "config", "user.email", "[email protected]"])
    run(["git", "config", "user.name", "Demo User"])
    run(["git", "config", "commit.gpgsign", "false"])  # required — signing may be global

    (work_dir / "main.py").write_text(
        "#!/usr/bin/env python3\n\n# FIXME: add input validation\ndef main():\n    pass\n"
    )
    (work_dir / "config.yaml").write_text("debug: false\nversion: 1.0\n")
    (work_dir / "project_data.csv").write_text("id,name\n1,important\n2,data\n")
    (work_dir / "auth.py").write_text(
        "# WIP: refactoring\ndef login(user, pwd):\n    pass  # stub\n"
    )
    run(["git", "add", "main.py", "config.yaml"])
    run(["git", "commit", "-m", "initial commit"])
    # Leave auth.py and project_data.csv uncommitted for demo acts

Tmux Plumbing: Common Failures [TUI Only]

Pane index must be queried, not hardcoded:

# ❌ Breaks when tmux base-index=1 (common configuration):
pane = "session_name:0.0"

# ✅ Query actual window and pane index after creating the session:
result = subprocess.run(
    ["tmux", "list-panes", "-t", session_name, "-F", "#{window_index}:#{pane_index}"],
    capture_output=True, text=True
)
pane = f"{session_name}:{result.stdout.strip()}"

Sending special characters: use -l flag (literal):

# ❌ tmux interprets / as a search, { as repeat count:
subprocess.run(["tmux", "send-keys", "-t", pane, "/ar:plannew"])

# ✅ -l flag sends text literally without tmux special-char interpretation:
subprocess.run(["tmux", "send-keys", "-t", pane, "-l", text])
# Required for: /command, {text}, !, and any prompt with special chars

Kill stale asciinema before re-recording:

# Without this: re-running --record with the same session name produces two
# asciinema processes writing to the same .cast file simultaneously (corrupt cast).
import signal
result = subprocess.run(["pgrep", "-f", f"asciinema.*{session_name}"],
                        capture_output=True, text=True)
for pid_str in result.stdout.strip().splitlines():
    try:
        os.kill(int(pid_str.strip()), signal.SIGTERM)
    except (ValueError, ProcessLookupError):
        pass

Detecting the CLI's input prompt:

# ❌ Looks for ASCII > — misses Claude Code's Unicode ❯ prompt:
if ">" in pane_content:
    return True

# ✅ Check for Unicode prompt AND a reasonable idle state:
prompts = ["❯", ">", "$ ", "% "]
if any(p in pane_content for p in prompts):
    return True

Confirming trust/safety dialogs:

# ❌ Exact string breaks if dialog wording changes between versions:
if "Yes, I trust this folder" in content:
    send_key("Enter")

# ✅ Keyword detection survives wording changes:
trust_keywords = ["trust", "safe", "quick safety check", "allow"]
if any(kw in content.lower() for kw in trust_keywords):
    send_key("Enter")

Preventing false "done" detection:

# ❌ One idle check gives false positives — AI briefly shows input prompt
# between chained tool calls (tool A finishes → shows ❯ → calls tool B)
if is_at_input_prompt(pane):
    return  # premature!

# ✅ Require 3 consecutive idle checks (~1.5s total) to confirm truly done.
# Also: sleep 1.5s FIRST to let Claude begin responding.
time.sleep(1.5)
idle_count = 0
while True:
    if is_at_input_prompt(pane):
        idle_count += 1
        if idle_count >= 3:
            return
    else:
        idle_count = 0
    time.sleep(0.5)

Plan approval is a different prompt state — do not use wait_for_response():

# ❌ Wrong: wait_for_response() doesn't recognize plan approval UI
session.send_prompt("/mytool:plannew Add input validation to auth.py")
session.wait_for_response(timeout=300)  # may hang or return early
session.send_prompt("/mytool:planrefine")  # sent before plan was accepted!

# ✅ Right: use a separate wait that detects the plan approval prompt
session.send_prompt("/mytool:plannew Add input validation to auth.py")
session.wait_for_plan_approval(timeout=300)
pause(7.0)      # let viewers read the plan
session.approve_plan()   # dynamically finds and presses the right option
pause(3.0)
session.send_prompt("/mytool:planrefine")

Choosing which plan option to press — do not hardcode the number:

# ❌ Wrong: hardcoded — breaks when menu reorders
session._send_key("2\n")  # "2" may mean "clear context" in some versions

# ✅ Right: parse actual menu; use exact word sets from autorun DemoSession
_ACCEPT_WORDS = ("yes", "proceed", "accept", "bypass")
_CLEAR_WORDS = ("clear context", "new conversation", "fresh context", "clear history")
# Regex handles ❯ cursor prefix:
m = re.match(r'[❯\s]*(\d+)\.\s+(.+)', stripped_line)
# Select line with accept word AND no clear word; fallback to "1"

Shell command overlap:

# ❌ Fixed sleep may not be enough if previous command runs longer:
session.send_command("python3 banner.py")
time.sleep(1.0)
session.send_command("claude")  # overlap!

# ✅ Wait for shell prompt after each command:
session.send_command("python3 banner.py")
wait_for_shell_prompt(pane, timeout=10)
session.send_command("claude")

Intro Banner [TUI Only]

Run the banner as a script inside the tmux pane — not from the Python harness (it won't appear in the asciinema recording):

_BANNER_SCRIPT = r'''#!/usr/bin/env python3
import sys
CYAN, BOLD, RESET, GRAY = "\033[96m", "\033[1m", "\033[0m", "\033[90m"
W = 70  # compute padding on PLAIN text — no ANSI in len() math
def pad(text): return text + " " * (W - len(text) - 2)
lines = [
    ("", ""),
    (f"  ╔{'═'*W}╗", CYAN),
    (f"  ║  {pad('mytool — safety plugin for Claude Code + Gemini CLI')}║", CYAN),
    (f"  ╠{'═'*W}╣", CYAN),
    (f"  ║  {pad('Install once. Runs silently in the background.')}║", CYAN),
    (f"  ║  {pad('This demo shows:')}║", CYAN),
    (f"  ║    {pad('1. Dangerous commands blocked + safe redirect')}║", CYAN),
    (f"  ║    {pad('2. File policy — restrict to existing files only')}║", CYAN),
    (f"  ╚{'═'*W}╝", CYAN),
    ("", ""),
]
for text, color in lines:
    sys.stdout.write(color + text + RESET + "\n")
sys.stdout.flush()
'''

def act0_live(session, tmp_dir):
    banner_path = tmp_dir / "_demo_banner.py"
    banner_path.write_text(_BANNER_SCRIPT)
    session.run_shell_cmd(f"python3 {banner_path}; rm {banner_path}", wait=1.0)
    pause(10.0)  # Let viewers read all items — this is the only chance
    session.run_shell_cmd("mytool --status", wait=2.0)
    pause(6.0)

Prompt Engineering [TUI Only]

For TUI demos where hooks must fire, prompt engineering is the core skill. The cause of a hook not firing is always the prompt or context — not the tool's behavior.

Key techniques (full guide in references/prompt-engineering.md):

• Prefix with "Using the Bash tool, run:" to force Bash over native Grep/Read/Edit
• Use unconditionally-blocked commands (rm, sed, git clean -f) over conditional ones
• Add "do not override any safety blocks" to prevent AI self-override
• Use keyword detection for trust dialogs, not exact string matching
• Parse plan approval menus dynamically with word sets, not hardcoded indices

---

Shared Sections [Both]

Act Design: Newcomers First

Every act must answer "what just happened and why does it matter?" for someone who has never seen your tool before.

Good act structure:

1. Viewer sees a natural-language prompt that makes sense to them
2. The tool responds visibly (block message, table output, corrected command)
3. The response is short enough to read in 7–10 seconds

CLI acts = section(title) + _run(cmd) + pause(N) pairs. TUI acts = session.send_prompt(text) + session.wait_for_response() + pause(N) pairs.

Features to show: Dangerous command blocked + redirect. Policy/mode toggle. Custom rule lifecycle. Auto-saved artifact with ls immediately after.

Features to skip: Background daemon operations. Multi-window features. Requires prior tool knowledge. Config file editing.

Pacing [Both]

def pause(seconds: float) -> None:
    """Sleep if in timed mode; no-op in pytest."""
    if _TIMED:   # or _DEMO_WITH_TIMING — both are the same concept, different names
        time.sleep(seconds)

# Standard timing:
pause(2.0)   # Before every prompt — let viewers finish reading previous response
pause(7.0)   # After short response — let viewers read the block message
pause(10.0)  # After complex response — plan creation, long output
pause(10.0)  # After intro banner — all items must be readable

pause() ≠ section() spacing — they are independent knobs:

• pause(N) = reading time. Change when viewers can't finish reading.
• section() \n\n\n before bar = visual gap between acts. Change when acts look crammed.
• Do NOT change both when only one needs adjustment.

Dual-Purpose File Pattern [Both]

#!/usr/bin/env python3
"""
Demo harness — dual purpose:
  CLI:     python tests/test_demo.py --record    (records asciinema cast)
  TUI:     python tests/test_demo.py --record    (asciinema attaches to tmux)
  Pytest:  pytest tests/test_demo.py::TestDemoFree   # $0.00 always
           pytest tests/test_demo.py::TestDemoRealMoney  # requires opt-in
"""
import pytest, argparse

@pytest.mark.skipif(
    not os.environ.get("DEMO_ENABLE_REAL_MONEY"),
    reason="Set DEMO_ENABLE_REAL_MONEY=1 to run live Claude tests"
)
class TestDemoRealMoney: ...

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--record",    action="store_true", help="Record to .cast → GIF/MP4")
    parser.add_argument("--run-acts",  action="store_true", help="Run acts (CLI: recorded command)")
    parser.add_argument("--play",      action="store_true", help="Scripted acts, no AI ($0.00)")
    parser.add_argument("--gif-only",  action="store_true", help="Convert existing cast to GIF")
    parser.add_argument("--verify",    action="store_true", help="Verify last recording")
    parser.add_argument("--setup",     action="store_true", help="Regenerate test fixtures")
    parser.add_argument("--no-cleanup",action="store_true", help="Keep tmux session (debug)")
    args = parser.parse_args()

--no-cleanup: Keep tmux session alive after demo finishes. Critical for diagnosing why a hook didn't fire.

Tool dependencies must be optional: asciinema, agg, ffmpeg are NOT project dependencies:

agg_bin = shutil.which("agg") or shutil.which("/tmp/agg")
if not agg_bin:
    print("[demo] agg not found — cast recorded but GIF skipped")
    print("[demo] Install: brew install agg  or  cargo install agg")
    return

ffmpeg_bin = shutil.which("ffmpeg")
if not ffmpeg_bin:
    print("[demo] ffmpeg not found — MP4 skipped (GIF only)")

gitignore output files:

*.gif
*.cast
*.mp4
*.webm

Test Suite Alongside the Demo [Both]

class TestDemoFree:
    """$0.00 — verify behavior without a live AI session."""

    # CLI: run exact demo commands against DEMO_ENV fixtures
    def test_stats_output(self):
        result = subprocess.run("mytool stats", env=DEMO_ENV, shell=True,
                                capture_output=True, text=True)
        assert "Sessions:" in result.stdout

    # TUI: call run_hook() directly
    def test_rm_blocked(self, tmp_path):
        rc, _, stderr = run_hook("Bash", {"command": "rm file.txt"}, root)
        assert rc == 2
        assert "trash" in stderr.lower()

Acts first, then tests — finalize each act's exact command string, then write tests that run the same command verbatim. Never write tests first and let them drift from the demo.

Dry-run each act against fixtures BEFORE writing assertions — verify the expected output actually appears. Don't assume fixture data will trigger behavior.

Tool Path Discovery [Both]

def find_tool_root() -> Path:
    candidates = [
        Path(__file__).parent.parent,  # tests/ → tool/
        Path.home() / ".tool" / "current",
    ]
    cache_base = Path.home() / ".cache" / "tool" / "versions"
    if cache_base.is_dir():
        for version_dir in sorted(cache_base.iterdir(), reverse=True):
            if (version_dir / "pyproject.toml").exists():
                candidates.append(version_dir)
                break
    for c in candidates:
        if (c / "pyproject.toml").exists():
            return c
    raise RuntimeError(f"Tool not found. Searched: {candidates}")

Resolution Targets [Both]

Quality	Terminal	Font	Approx Output	Use Case
Full HD (1080p)	160x48	18	~1750x1230	GitHub README, presentations
2K	160x48	20	~1960x1380	High-DPI displays
Minimum viable	160x48	16	~1560x1098	Acceptable for web
NOT acceptable	80x24	any	<1000px wide	Too small for text readability

Rule: Never use 80x24 for demo recordings. Minimum 160x48 cols/rows with font-size 16+.

GIF and Video Settings [Both]

Note: --idle-time-limit appears in two contexts with different values:

• asciinema rec flag: limits idle gaps DURING recording (scripted: 3s; TUI live: 5s)
• agg flag: limits idle gaps IN GIF output (both pathways: 10s — preserves banner)

Setting	CLI	TUI scripted	TUI live
--window-size (asciinema)	160x48	auto from terminal	{cols}x{rows}
--idle-time-limit (asciinema)	—	3	5
--font-size (agg)	16-18	16-18	16-18
--speed (agg)	0.75	0.75	0.75
--idle-time-limit (agg)	10	10	10
--last-frame-duration (agg)	5 (hold closing frame)	—	—
--renderer fontdue	yes	yes	yes
--theme dracula	yes	yes	yes

MP4 codec strategies (4-strategy fallback, best first):

1. libx265 crf=28 tune=animation — HEVC, ~50% smaller than libx264, needs tag:v hvc1 for macOS. Note: HEVC not supported in Firefox; use libx264 if embedding in web pages
2. libx264 crf=24 tune=animation — broadest browser compat (Chrome, Firefox, Safari); CRF 24 in x264 gives similar visual quality to CRF 28 in x265
3. h264_videotoolbox -q:v 65 — macOS hardware encoder, VBR, fast but larger files
4. ffmpeg default — last resort

CRF note: CRF 24 in libx264 and CRF 28 in libx265 produce approximately equivalent visual quality. Both are visually lossless for terminal text. Use tune=animation for terminal recordings (flat colors + sharp edges) — do not use tune=fast.

Settings that did not work:

• --font-size 20 at 160x48: text too large, content gets cut off at edges
• --speed 1.5: too fast for viewers to read block messages
• --idle-time-limit 5 (agg): banner and status pauses get cut short in GIF

Common Pitfalls [Both]

See references/common-pitfalls.md for the full 30+ pitfall table covering CLI, TUI, and shared issues including recording failures, timing problems, prompt engineering failures, and encoding pitfalls.

---

Concrete Examples

Working examples with act lists, design decisions, prompt evolution, and recording parameters:

• CLI pathway: references/examples/aise-cli-example.md — aise (ai_session_tools), 7 CLI acts
• TUI pathway: references/examples/autorun-tui-example.md — autorun for Claude Code, 7 TUI acts with prompt debugging history

---

Tool Stack [Both]

Tool	Purpose	Install	Required for
asciinema	Record terminal session to .cast	brew install asciinema	Both
agg	Convert .cast → animated GIF	brew install agg or cargo install agg	Both
ffmpeg	Convert GIF → MP4	brew install ffmpeg	Both
tmux	Full TUI session isolation and control	brew install tmux	TUI only

---

References

File	Content
references/hook-based-tools.md	Forcing Bash tool calls for hook-based tool demos
references/common-pitfalls.md	30+ catalogued pitfalls with root causes and fixes
references/prompt-engineering.md	TUI prompt strategies for reliable AI tool invocation
references/rust-demos.md	Rust demo harness patterns (AtomicBool, multi-shell, cargo bins)
references/typescript-demos.md	TypeScript/Node demo patterns (spawn, Vitest)
references/tmux-integration.md	tmux recording patterns (multi-pane, multi-shell, idle detection)
references/cast-merging.md	Merging multiple .cast files into single recordings
references/alternative-tools.md	VHS, t-rec, and other recording tools vs asciinema
references/examples/aise-cli-example.md	Working CLI demo: aise (7 acts, privacy isolation)
references/examples/autorun-tui-example.md	Working TUI demo: autorun (7 acts, prompt evolution)

---

Version History

v5.1.0 — 2026-03-11

• Added Initial Frame Cleanup section: trim_cast_to_banner() removes shell init artifacts from .cast first frame
• Added Rust and TypeScript implementations in respective reference files
• Added initial frame cleanup pitfall to references/common-pitfalls.md

v5.0.0 — 2026-03-11

• Added trigger phrases to description for reliable skill activation
• Added version field to frontmatter
• Added Resolution Targets table: minimum 160x48 at font-size 16+, Full HD at font-size 18
• Updated font-size recommendation from 14-16 to 16-18 for Full HD output
• Documented CRF equivalence: CRF 24 (x264) = CRF 28 (x265) visual quality
• Extracted Common Pitfalls (30+ items) to references/common-pitfalls.md
• Extracted Prompt Engineering section to references/prompt-engineering.md
• Extracted concrete examples to references/examples/
• Added references/rust-demos.md — Rust harness patterns from canal (AtomicBool, multi-shell, cargo bins)
• Added references/typescript-demos.md — TypeScript/Node harness patterns (spawnSync, Vitest)
• Added references/tmux-integration.md — multi-pane, multi-shell, idle detection, trust dialogs
• Added references/cast-merging.md — asciinema cat, programmatic merging, transitions
• Added references/alternative-tools.md — VHS (charmbracelet) and t-rec comparison
• Added References table listing all 10 reference files
• Reduced SKILL.md from 5,921 to ~3,300 words (under 5,000 hard limit)
• Converted second-person prose to imperative form
• Removed TODO markers from prose (code examples use FIXME/stub instead)

v4.0.0 — 2026-03-05

• Added "Choose Your Pathway" section: CLI vs TUI scripted vs TUI live vs Mixed decision table
• Section labels throughout: [CLI Only], [TUI Only], [Both] — readers skip irrelevant parts
• Added complete CLI pathway (no tmux): harness IS the recorded process
• Fixed "Wrong architecture" block: retitled "Wrong for TUI tools" (not universally wrong)
• Added 10 session-error pitfalls with root causes, labeled by pathway
• Added _run(), _TIMED/_DEMO_WITH_TIMING (same concept, two names), pause(), section() for CLI
• Clarified: pause() timing ≠ section() newline spacing — independent knobs
• Added "Write Acts First, Then Matching Tests" rule (prevents test/demo drift)
• Added synthetic fixture data: deterministic IDs, date-shifting, TOOL_ISOLATION_VAR
• Added privacy isolation via DEMO_ENV + generic TOOL_ISOLATION_VAR pattern
• Added --run-acts / --gif-only / --verify / --setup / --play CLI flag patterns
• Added CLI verification: cast text fragments (vs TUI JSONL tool-call parsing)
• Added aise CLI example (peer to autorun TUI example); aise-specific details labeled
• Added TUI scripted pathway (hook-only, $0.00): acts call run_hook() directly
• Added send-keys -l (literal flag) to TUI plumbing — required for slash commands and braces
• Added _kill_stale_recording_procs() — prevents corrupt casts on re-record
• Added GIT_CONFIG_NOSYSTEM=1 + commit.gpgsign=false to mock git repo setup
• Added background thread pattern for TUI live recording (acts thread + asciinema foreground)
• Updated GIF settings table: separate asciinema vs agg --idle-time-limit rows
• Added libx265 tune=animation as strategy 1 in MP4 fallback (4 strategies total)
• Noted section() auto-sizing improvement: replace hardcoded width with max(68, len(title) + 6)
• All pitfall rows labeled [CLI], [TUI], or [Both] for fast scanning

v3.0.0 — 2026-03-01

• Complete rewrite based on 20+ commits of real failures and fixes from autorun demo
• Added architecture section: why tmux + real TUI (vs simulated/headless)
• Added work directory setup: short fixed paths vs tempfile.TemporaryDirectory
• Added tmux plumbing section: pane index, Unicode prompt, trust dialog, false-done detection, shell overlap
• Added prompt engineering failure table with exact failed prompts and fixes
• Added self-override prevention; bash exec doesn't work in TUI; dynamic UI string detection
• Added concrete example: autorun prompt evolution showing all 6 attempts for act2
• Added test suite structure, CLI flags structure, GIF settings

v2.0.0 — 2026-03-01

• Rewrote as general-purpose (was autorun-specific)
• Added newcomers-first, pacing, verification sections

v1.0.0 — 2026-03-01

• Initial release