poltergeist

👻 poltergeist

Code reviews from contributors who aren't in the room.

Poltergeist is a Claude Code plugin and CLI that lets you invoke a pull request review through the lens of any contributor — using their voice, values, heuristics, and communication style — even when they're not available.

Build a ghost from a contributor's git history, GitLab review comments, Slack messages, and design docs. Then summon them on any diff.

git diff main | claude "review this as @alice"

---

Why

Code review quality is uneven. Senior contributors have hard-won instincts — patterns they watch for, questions they always ask, things that reliably bite the team later. When they're on leave, moved on, or just at capacity, that institutional knowledge doesn't transfer.

Poltergeist captures it.

It's not a linter. It's not a generic AI review. It's a reviewer with a specific point of view, a specific voice, and a specific history with your codebase.

---

How it works

[Build]   data sources → npx @poltergeist-ai/cli extract → ghosts/<slug>.md
[Invoke]  diff + ghost file → review in contributor's voice

1. Run the extractor against a contributor's git history, MR comments, Slack export, and docs
2. A ghost file is generated — a structured profile of their values, heuristics, communication style, and example comments
3. Manually validate the ghost with teammates who know the contributor
4. Invoke reviews through Claude Code at any time

The ghost file (ghosts/<slug>.md) lives in your repo alongside your code. It's version-controlled, team-editable, and human-readable.

---

Quick start

1. Install the plugin

Add poltergeist as a Claude Code plugin from GitHub:

claude plugin marketplace add gkweb/poltergeist
claude plugin install poltergeist

This gives you the /extract command and the review skill.

Note: We plan to submit poltergeist to the official Claude Code plugin marketplace once it's stable. After that, installation will be a single claude plugin install poltergeist command.

2. Build a ghost

No install needed for the CLI — run directly via npx:

npx @poltergeist-ai/cli extract \
  --contributor "Alice Smith" \
  --email [email protected] \
  --git-repo . \
  --gitlab-export ./exports/gitlab-notes.json \
  --slack-export ./exports/slack/ \
  --output ghosts/alice-smith.md

Then open ghosts/alice-smith.md and fill in the [fill in manually] sections. The manual pass is the most important step — especially the ranked values and example comments.

Or use the /extract command inside Claude Code which guides you through the options.

3. Invoke a review

# Review a diff
git diff main | claude "review this as @alice-smith"

# Review a specific file
claude "what would alice think of this?" < src/components/MyForm.vue

# Review an open MR (requires glab CLI)
claude "review MR !123 as @alice-smith"

---

Ghost file

Each contributor has a single ghost/<slug>.md file. It contains:

• Identity — role, domains, when the ghost was last updated
• Review philosophy — what they care about, ranked; what they ignore; dealbreakers
• Communication style — tone, severity prefixes they use, vocabulary, comment density
• Code patterns — what they introduce, what they push back on, refactors they suggest
• Known blind spots — areas they historically under-review (Claude flags these)
• Example review comments — verbatim excerpts that ground Claude's voice output

See packages/plugin/skills/poltergeist/references/ghost-schema.md for the full schema.

---

Data sources

Source	What it contributes	How to export
Git history	Coding patterns, naming conventions, commit style, primary areas	Available locally
GitHub PR reviews	Review voice, severity prefixes, recurring phrases	Auto-extracted from GitHub URL
GitLab MR comments	Review voice, severity prefixes, recurring phrases	GitLab API (see docs)
Slack export	Informal technical voice, how they discuss code in conversation	Slack admin export
Design docs / ADRs	How they reason about architecture, what they value in writing	Local directory

All sources are optional. More sources = more accurate ghost. GitLab comments are the single most valuable input for voice fidelity.

---

Repo layout