Back to search
ghostwriter
MarketplaceWorks with Claude CodeGitHub
Ghostwriter — a Claude Code plugin that keeps evidence-backed negotiation profiles of the people you meet, built from meeting transcripts and stored in your Notion. Never sends, every claim backed by a quote.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/rlfszn--ghostwriter)HTML
<a href="/listings/rlfszn--ghostwriter"><img src="/api/listings/rlfszn--ghostwriter/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
As a plugin:
/plugin marketplace add rlfszn/ghostwriter /plugin install ghostwriter@ghostwriter
Or copy the skills by hand (keep all seven folders together: they reference each other as siblings):
git clone https://github.com/rlfszn/ghostwriter cp -r ghostwriter/skills/* ~/.claude/skills/
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 23 lines
- {
- "name": "ghostwriter",
- "owner": {
- "name": "Elia Rolfi"
- },
- "metadata": {
- "description": "Ghostwriter: evidence-backed counterpart profiles from meeting transcripts, on Notion.",
- "version": "1.0.0"
- },
- "plugins": [
- {
- "name": "ghostwriter",
- "description": "Evidence-backed negotiation profiles of the people you meet, kept on Notion and built from meeting transcripts. Prepares meetings and drafts messages. Text only, never sends, every claim backed by a quote.",
- "version": "1.0.0",
- "author": {
- "name": "Elia Rolfi"
- },
- "license": "MIT",
- "category": "productivity",
- "source": "./"
- }
- ]
- }
README.md · README · 239 lines
- # Ghostwriter
- *[Leggi in italiano](README.it.md)*
- A Claude Code plugin that keeps an evidence-backed profile of the people you
- negotiate with, built from the transcripts of your meetings and stored in your own
- Notion workspace.
- ## The problem
- Between one meeting and the next you forget what people actually care about. Who
- needs to justify a decision to whom. What the client said about delivery dates three
- months ago. Which concession unblocked things last time. Meeting notes don't help:
- they record what was decided, not what each person wants, fears, and will accept.
- Ghostwriter reads the transcript of a meeting and keeps, for each counterpart, a
- profile that grows over time: drivers, decision style, effective levers, things to
- avoid, fears, firm and negotiable positions. **Every line is backed by a verbatim quote
- with its timestamp.** When there is no quote, the field stays empty. Before the next
- meeting, it tells you what each person risks if they say yes, what to ask, and what
- not to say. It drafts the follow-up. It never sends anything.
- ## How it works
- ```
- recording ──► transcript ──► process-meeting ──► Notion: Meeting page + updated profiles
- (PLAUD, or (text only, + briefing, recommended move,
- any recorder) never audio) follow-up draft, predictions
- prepare-meeting · compose-message · consult-profile read the profiles, write nothing
- ```
- 1. You record a meeting you took part in (a PLAUD device, a phone, anything).
- 2. You get a transcript: from PLAUD, from any transcription service, or locally with
- `tools/transcribe.py`.
- 3. `process-meeting` normalises it into numbered turns, identifies who spoke, extracts
- facts, positions and interests, closes the open predictions, and updates each
- person's profile on Notion — one person at a time, with a quote for every claim.
- 4. `prepare-meeting`, `compose-message` and `consult-profile` only read.
- Two Notion databases hold everything: **People** (one page per counterpart) and
- **Meetings** (one page per recording, with the transcript in a subpage). The state
- of every processing run lives on Notion, so an interrupted run resumes where it
- stopped instead of duplicating.
- Every briefing closes the open **predictions** and opens new ones, in falsifiable form
- (`IF I do X, THEN they will do Y; REFUTED IF Z`). A profile that cannot be proved wrong
- cannot be corrected either. A hit rate under 40% means the profile is wrong; above 90%
- means the predictions are too obvious.
- ## The profile
- | Field | What goes in it |
- |---|---|
- | Key drivers | Status, Security, Control, Autonomy, Recognition, Belonging, Financial result — only once a pattern is seen in 2+ meetings |
- | Decision style | Analytical, Intuitive, Consensual, Directive, Avoidant — only after an actual decision was observed |
- | Communication style | how they talk, with quotes |
- | Goals and interests | what a position lets them do or avoid |
- | Fears | risks they perceive, written as topics that need preparation, not weak spots |
- | Sensitive topics | business topics to prepare for |
- | Effective levers | interest → move → why it works for them → cost to you |
- | Things to avoid | what doesn't work or backfires |
- | Firm / Negotiable positions | negotiable only if they stated a condition ("if…, then…") |
- | Alignment with me | Ally, Favorable, Neutral, Skeptical, Opponent |
- | Profile reliability | Low / Medium / High, from the number of meetings with real evidence |
- | User notes | yours: the agent reads it, never writes it, and it always wins |
- Each claim carries its confidence level (`hypothesis`, `observed`, `pattern`) and the
- number of meetings it rests on. Hypotheses never enter the profile fields: they go
- into a "to verify" list as open questions. A profile with three fields filled and
- eight empty is a healthy profile.
- ## What it will never do
- - **Analyse audio.** No tone, pitch, hesitation, or emotion inferred from the voice.
- Text only. The EU AI Act (art. 5(1)(f)) prohibits inferring emotions from biometric
- data such as voice in the workplace; working only on the text is what keeps the
- system outside that ban.
- - **Send anything.** It writes drafts. An automatic send based on a psychological
- profile is exactly what must not be able to happen by mistake.
- - **Evaluate the people who report to you.** No performance judgements, no use for
- hiring, promotion, duties or discipline. Using it that way would make it a
- high-risk system under the AI Act (Annex III) and a workplace-monitoring tool under
- most labour laws. People listed as your subordinates get a profile that is mostly
- empty by design, and their words are not archived in full.
- - **Write anything without a quote.** An empty field beats plausible psychology.
- - **Guess who a speaker is.** If it is not certain, it stops and asks: a wrong
- attribution poisons two profiles for good.
- - **Record special-category data** (health, politics, religion, union membership,
- sexual orientation, ethnicity — GDPR art. 9). If it comes up, it is redacted and
- never used.
- - **Profile people who were not in the room**, or process recordings of conversations
- you were not part of.
- - **Keep data after a deletion request.** `remove-person` cleans the person out of
- every page and tells you the manual steps (trash, version history, local files) that
- a real GDPR erasure needs.
- The test behind all of it: every line must be one you could show to the person it is
- about. The framing is "what does this person risk if they say yes", not "where do I
- push".
- ## Requirements
- - [Claude Code](https://claude.com/claude-code)
- - The **Notion MCP connector**, connected to a **personal** Notion workspace (not your
- company's: a workspace admin can read everything in it)
- - Optional: a **PLAUD** recorder with its MCP connector, to fetch transcripts directly
- - Optional: **mlx-whisper** on an Apple Silicon Mac, to transcribe audio locally with
- `tools/transcribe.py`
- ## Install
- As a plugin:
- ```
- /plugin marketplace add rlfszn/ghostwriter
- /plugin install ghostwriter@ghostwriter
- ```
- Or copy the skills by hand (keep all seven folders together: they reference each
- other as siblings):
- ```bash
- git clone https://github.com/rlfszn/ghostwriter
- cp -r ghostwriter/skills/* ~/.claude/skills/
- ```
- ## First run
- Say **"set up Ghostwriter"**. The `setup-notion-memory` skill:
- 1. creates your **Ghostwriter home**, `~/.ghostwriter/`, and copies the templates into
- it: `configuration.md`, `style.md`, `deletions.md`, plus `transcripts/incoming/` and
- `transcripts/archive/`;
- 2. asks you to fill in `configuration.md` — at minimum `me` (your name, every alias a
- transcript may use for you, a contact email) and `subordinates` (everyone who
- reports to you). `me` keeps you from profiling yourself; `subordinates` keeps a
- team member from being treated as a counterparty;
- 3. checks that the Notion workspace has no other members, then creates the two
- databases, their relations, rollups and views, and writes the IDs back into
- `configuration.md`.
- Your private files live **outside the plugin**, so updates never touch them. For a
- separate instance (say, one for work and one personal), run the setup from a project
- folder and ask for a per-project instance: it creates `./.ghostwriter/` there, and the
- skills use it whenever you work from that folder.
- Fill in `style.md` with 3-5 real messages you sent before using `compose-message`,
- otherwise the drafts will sound like the agent, not like you.
- By default the skills run in **autonomous** mode: they never stop to ask, they leave
- anything doubtful marked "to confirm" on Notion. Start a request with
- `MODE: interactive` (or put that line in your `CLAUDE.md`) if you want to be asked.
- While `confirm_before_writing_profiles` is `true`, interactive runs show you every
- profile change before writing it.
- ## Usage
- | You say | Skill | Result |
- |---|---|---|
- | "process this morning's meeting" | `process-meeting` | Meeting page + updated profiles + briefing |
- | "tomorrow I'm seeing Mark about the budget, how do I approach him?" | `prepare-meeting` | goal, what he risks if he says yes, questions, zero-cost concessions |
- | "write me the message for Barlow" | `compose-message` | a draft in your register, to reread |
- | "what do we know about Farnsworth?" | `consult-profile` | the profile, and how many meetings it rests on |
- | "delete X's profile" | `remove-person` | full clean-up for GDPR, after your confirmation |
- | "set up Ghostwriter" / "create the databases on Notion" | `setup-notion-memory` | first-run setup, once |
- The skills reply, and write on Notion, in your language. Notion property names stay in
- English: they are the contract the queries rely on.
- ### Transcribing locally
- `tools/transcribe.py` transcribes audio on your Mac without sending it anywhere:
- ```bash
- uv venv --python 3.11 .venv && uv pip install --python .venv/bin/python mlx-whisper numpy
- .venv/bin/python tools/transcribe.py --language en meeting.m4a
- ```
- The output goes to `transcripts/incoming/` in your Ghostwriter home, without speaker
- labels: the agent attributes speakers from the content and says so. With PLAUD and the
- speakers renamed in its app, this step is not needed.
- ## Using it with other agents
- The skills are plain `SKILL.md` markdown files with standard frontmatter. Any agent
- that loads skills in that format can use them; there is no code to integrate, only
- text to load. Point the agent at `AGENTS.md` for the always-on rules, give it a Notion
- connector, and keep the same Ghostwriter home. The external contract is the Notion
- schema (`skills/profiling-references/references/notion-schema.md`): anything that
- reads or writes the two databases follows it.
- The skills were written to run unattended as well (a webhook from your recorder, a
- scheduled job). In that mode they never ask, never overwrite a filled field, never
- delete, and `remove-person` refuses to run. See
- `skills/profiling-references/references/execution-modes.md`.
- ## Layout
- ```
- .claude-plugin/ plugin and marketplace manifests
- AGENTS.md always-on rules, for any agent
- skills/
- profiling-references/ LIBRARY: core rules, method, schema, limits, templates
- process-meeting/ the main flow
- prepare-meeting/ read-only
- compose-message/ read-only + drafts
- consult-profile/ read-only
- remove-person/ destructive, asks for confirmation
- setup-notion-memory/ first-run setup
- tools/transcribe.py local transcription with mlx-whisper
- tests/ 4 synthetic cases + 8 negative tests
- ```
- One copy of the logic: the analysis method, the Notion schema and the legal limits
- live in the library, and the skills point to them.
- ## Tests
- `tests/` holds four synthetic meetings with expected results and eight negative tests
- where the right behaviour is to stop (insufficient material, noisy transcript,
- unknown speakers, sensitive data, deceptive message, single meeting, HR use, voice
- analysis). Run them against sandbox databases before your first real recording: say
- "run the test suite in the sandbox". Instructions in `tests/README.md`.
- ## Legal
- This is not legal advice. You are responsible for complying with the laws that apply
- to you: in many countries recording a conversation requires telling the other people,
- or their consent; workplaces often have policies on recording devices; and data
- protection law (the GDPR in the EU) applies to the profiles you keep. Check your
- company's policies before recording any meeting — if they forbid it, don't. Tell
- people you record, stop when they ask, and if someone asks what you have written about
- them, show them the page. `skills/profiling-references/references/ethical-legal-limits.md`
- explains the rules the agent applies and the ones that are up to you.
- ## License and author
- MIT. Created by Elia Rolfi.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.