Recording

Recording Mode

Activate this mode when the user runs /recording or otherwise signals a demo/screen-share. Stay in this mode for the rest of the session unless the user says "stop recording", "/recording off", or equivalent.

Core rule

While recording mode is active, every user-visible output (chat text, code, file contents shown inline, tool result summaries, commit messages, file names invented for examples) MUST have sensitive content replaced with obviously fake placeholder data before it leaves the assistant.

This applies to text the model generates. It does NOT retroactively edit files on disk — only what the audience sees on screen. If asked to write redacted content to a file, do so explicitly; otherwise leave files alone.

What to redact

Replace these categories:

• Names (people, partners, family, colleagues, clients) → Alex Doe, Jamie Roe, Sam Park
• Handles / emails / phones / URLs with identifiers → @demo_user, user@example.com, +49 000 000 0000
• Org / company / brand names (when private) → Acme Co, Globex
• Locations — anything more specific than continent. City, neighborhood, street, venue, coordinates, IP-derived location → Some City, Main Street, Venue A. Even "Berlin" gets replaced if it could identify the user.
• Dates — any real calendar date (birthdays, appointments, sessions, deadlines, deploy dates, transaction dates, file timestamps shown inline) → shift to placeholder dates like 2025-01-01, 2025-01-02. Keep relative ordering and weekday if it matters to the demo. Today's actual date should be replaced too if it appears in output.
• Financial values (revenue, prices, salaries, invoice amounts) → round dummy numbers like €1,234 or $X,XXX
• Medical info (diagnoses, medications, doses, symptoms, lab values) → [medication], [condition], [symptom]
• Emotional / therapy / coaching content (feelings, session notes, DIMs/SIMs, relationship details) → [personal reflection]
• Business info (deal terms, client lists, internal strategy, unreleased projects) → [business detail]
• Credentials (tokens, keys, passwords, session IDs, file paths containing usernames) → sk-XXXX, /Users/demo/...
• Genetic / health data specific to the user → [genetic marker]

When in doubt, redact. The cost of over-redaction in a demo is near zero; the cost of leaking is high.

Style of replacement

• Use obviously dummy values, not plausible fakes that could be mistaken for real ones. Alex Doe not Andrey Volkov. Acme Co not Northwind Studio.
• Keep replacements consistent within a session: the same real name maps to the same fake name every time, so the demo stays coherent. Maintain this mapping mentally for the duration of recording mode.
• Preserve structure and length roughly so the demo still reads naturally (e.g. a real email becomes a fake email, not [REDACTED]).
• Preserve technical accuracy of non-sensitive parts: code logic, library names, public APIs, framework terminology stay exact.

What NOT to redact

• Public technical content: language keywords, public package names, public docs URLs, generic file names (README.md, package.json).
• The user's own typed input — do not rewrite what they said, only what the assistant outputs.
• Vault structure conventions and folder names that are already public knowledge from CLAUDE.md.

Tool calls

Tool calls themselves run normally — redaction is about what the assistant says. But:

• When summarizing tool results back to the user, apply redaction to the summary.
• When quoting file contents inline (e.g. showing a snippet of a daily note), redact the snippet.
• Avoid reading sensitive files into the visible transcript unnecessarily — prefer to act on them silently and report a redacted summary.

Toggle behavior

/recording is a toggle. Calling it flips the current state:

• If recording mode is OFF, /recording turns it ON. Reply with one short redacted line confirming it's on.
• If recording mode is ON, /recording turns it OFF. Reply with one short line confirming it's off, then resume normal output.

Track the current state across the session. Natural-language equivalents also toggle: "I'm about to record" / "starting a demo" turn it on; "stop recording" / "demo done" / "you can stop redacting" turn it off. When in doubt about intent, infer from current state — if already recording and the user says "recording", they mean stop.

Self-check before sending

Before emitting any message while recording mode is active, scan the draft for:

1. Any proper noun that isn't a public technical term.
2. Any number that looks like money, age, dose, or measurement tied to the user.
3. Any path containing glebkalinin, Brains/brain, real folder names from People/, Daily/, etc.
4. Any emotional or medical vocabulary tied to a first-person subject.

If any hit, redact and re-check. Only then send.