recording — Claude Code skill: redacts PII during demos
Triggered by /recording, it replaces names, dates, money and other personal data with obvious placeholders in real time while you record or share your screen.
MIT 184 installs ★ 414 updated 2026-10-23
npx skills add glebis/claude-skills --skill recording -a claude-codeWhat recording does
The skill swaps every piece of personal information in agent output for clear dummies: Alex Doe instead of real names, [email protected] instead of email, Acme Co instead of company names, rounded figures instead of actual amounts, 2025-01-01 instead of real dates. Redaction covers people's names, cities, addresses, phones, financial and medical data, business details, tokens and paths that contain usernames. Code, libraries and public APIs stay exact.
Placeholders remain consistent within a session: the same real name maps to the same fake one every time, structure and length are preserved to keep the demo readable. /recording is a toggle — calling it again switches the mode off. Files on disk are untouched; only output visible to the audience gets redacted.
What it helps with
- mask personal data during screen recording
- show a demo without real client names and deal amounts
- hide meeting dates and addresses during a stream
- remove tokens and username paths from agent output
Example prompts for the agent
- Turn on
/recording, about to record the screen - Stop redacting, recording finished
- Starting a demo, switch to recording mode
How to install recording
1Skills CLI (Vercel)
npx skills add glebis/claude-skills --skill recording -a claude-code2Manually: put the files in place
~/.claude/skills/recording/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/recording/
1Skills CLI (Vercel)
npx skills add glebis/claude-skills --skill recording -a codex2Manually: put the files in place
.agents/skills/recording/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add glebis/claude-skills --skill recording -a cursor2Manually: put the files in place
.cursor/skills/recording/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install glebis/claude-skills recording2Manually: put the files in place
.github/skills/recording/SKILL.md1Skills CLI (Vercel)
npx skills add glebis/claude-skills --skill recording -a gemini-cli2Platform CLI
gemini skills install https://github.com/glebis/claude-skills --path recordingManually: put the files in place
.devin/skills/recording/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add glebis/claude-skills --skill recording -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about recording
How to install recording in Claude Code?
Run npx skills add glebis/claude-skills --skill recording -a claude-code, the skill lands in ~/.claude/skills/recording/SKILL.md.
What does recording do?
Activated by /recording, it replaces names, addresses, dates, amounts, tokens and other personal data in agent output with obvious placeholders so recordings or demos do not expose your data. Code and public APIs remain unchanged.
Does recording touch files on disk?
No, it only redacts agent responses visible to the audience. To write redacted content to a file, you need to ask the agent explicitly.
Source text
SKILL.md · MIT
--- name: recording description: Demo/recording mode that redacts personally identifiable and sensitive information from Claude Code's outputs. Use when the user invokes /recording or says they are about to record, screen-share, or demo their Claude Code session and want PII scrubbed in real time. --- # 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`, `[email protected]`, `+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.
© glebis · MIT · glebis/claude-skills