privacy-filter-skill
olhahah/privacy-filter-skill
A Claude Code skill that strips personal and sensitive content from a transcript while keeping the work content intact.
Instalação
npx convez-ai olhahah/privacy-filter-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
privacy-filter-skill
A Claude Code skill that strips personal and sensitive content from a transcript while keeping the work content intact.
It cuts greetings/weather/weekend plans, third-party medical details, comp talk, mental-health specifics, and similar — and when a sentence mixes work and personal context (e.g. "X was delayed because my wife is sick"), it keeps the work fact and drops the personal reason. Every removal goes to a local log file so you can review what was cut.
This is one of two skills used together to ship meeting transcripts to a private GitHub repo. The other is granola-share-skill — it's what most people will invoke directly. This filter runs underneath it.
What you need first
- Claude Code installed (
https://claude.com/claude-code) - This skill installed in
~/.claude/skills/privacy-filter/(see install below)
Install
mkdir -p ~/.claude/skills
git clone https://github.com/olhahah/privacy-filter-skill ~/.claude/skills/privacy-filter
One command. The clone lives directly where Claude Code expects skills — no symlinks, no extra source dir at $HOME.
Update
cd ~/.claude/skills/privacy-filter && git pull
That's it. The skill is just a markdown file Claude reads — there's nothing to rebuild.
Use
You'll usually not invoke this directly — the granola-share skill calls it for you as part of pushing transcripts. If you want to run it directly on some text:
"filter this transcript using the privacy filter"
Paste the transcript, Claude applies the rules, prints the filtered version, and appends a log entry per removal to _logs/privacy-redactions.md in your current directory.
What gets cut
- Always cut: greetings/weather/weekend plans, compensation, job-search talk, mental-health specifics, third-party medical (partners, kids, parents), confidentiality cues ("between us"), emotional confessions, personal attacks on named colleagues.
- Likely cut (logged for review): mild emotional language tied to work, opinions on named colleagues, caregiving phrases without medical detail, self-deprecating work feelings.
- Always kept: the employee's own holiday/sick leave, project impact statements (even when emotionally framed), neutral work friction, mentions of team/role/project/customer/metric.
- Mixed work + personal: the filter keeps the work fact and drops the personal reason. "X was delayed because my wife is sick" → "X was delayed."
What does NOT get blocked
The filter never blocks. It makes calls in the moment and lets the transcript through. Safety comes from the log review afterwards — you scan _logs/privacy-redactions.md and either approve or revert any individual removal.
The log
Every removal is appended to _logs/privacy-redactions.md in whatever directory you're running from. Entry format:
## <date> — <transcript title>
- **Confidence:** high | low
- **Speaker:** <UNKNOWN by default when no speaker labels are available>
- **Reason:** <plain-English reason>
- **Quote:** "<full text removed>"
- **Work fact preserved (if any):** <what was kept inline, or "none">
- **Review:** _<blank — fill in after you read>_
Mark Review: as ✓ correct / ✗ should have kept / ~ partial when you scan the log. Those decisions inform future refinements to this skill.
Limitations
- No speaker labels expected. This skill works on raw text without speaker attribution. The "background bleed" category (cutting lines from people outside the meeting) only fires when context makes it obviously off-meeting — otherwise it leaves the text in.
- No glossary lookups. Lighter than MVP-brain's version: no per-org name list, no person-specific rules. If you need that, run the full MVP-brain pipeline on your own machine after pulling.
Related
- granola-share-skill — the orchestration skill that uses this filter to push Granola transcripts to a private GitHub repo.
Questions
Olha — olha.dolinska@healf.com