Project handover for Claude Code and Codex.
Export a guide and privacy-filtered AI memories as a zip; dry-run import, then keep only verified memories.
When a developer leaves a project, two kinds of knowledge leave with them:
| Knowledge | What usually happens |
|---|---|
| Human context | The next person gets scattered docs, stale notes, or a rushed call. |
| AI context | The assistant starts from zero, even if another assistant learned the project over months of sessions. |
knowledge-transfer packages both into handover.zip: a short human onboarding
document plus portable AI memories that are filtered on export and verified on
import.
| Phase | Who runs it | What happens |
|---|---|---|
export |
The colleague leaving | Analyzes the project, reads project-scoped memories, filters private content, writes handover/ and handover.zip. |
import |
The colleague joining | Shows the onboarding doc, dry-runs candidate memories, verifies each claim against the current code, then installs only approved project knowledge. |
Run without arguments, the skill detects the phase:
handover/manifest.jsonorhandover.zipexists -> import- no
handover/package -> export
From inside Claude Code:
/plugin marketplace add ferdinandobons/knowledge-transfer
/plugin install knowledge-transfer@knowledge-transfer
/reload-plugins
Plugin-installed skills are namespaced:
/knowledge-transfer:knowledge-transfer export
/knowledge-transfer:knowledge-transfer import
/knowledge-transfer:knowledge-transfer import --dry-run
From your terminal:
codex plugin marketplace add ferdinandobons/knowledge-transfer --ref main
codex plugin add knowledge-transfer@knowledge-transferStart a new Codex thread, then invoke the plugin:
@knowledge-transfer export
@knowledge-transfer import
@knowledge-transfer import --dry-run
Agents that only read raw SKILL.md folders can use this repo's SKILL.md plus
references/ in a clean skill directory. Prefer the plugin install above for
Claude Code and Codex.
Leaving a project:
/knowledge-transfer:knowledge-transfer export
In Codex:
@knowledge-transfer export
Review the proposed package, including the exclusion report, then send
handover.zip to the next colleague.
Joining a project:
Place handover.zip in the project root, then run:
/knowledge-transfer:knowledge-transfer import
In Codex:
@knowledge-transfer import
Read the onboarding doc first. Import then shows a dry-run plan with candidate memories, checked evidence, proposed actions, and reasons. Confirm the plan to write accepted memories through the current agent's memory mechanism; stale memories are fixed when obvious or rejected with a receipt.
handover.zip # transfer this file
handover/ # local staging folder used to build the zip
ONBOARDING.md # 600-1000 words, non-technical project map
memories/ # neutral, portable project memories
omissions.json # safe counts/categories for omitted private memories
manifest.json # export date, commit SHA, exported/excluded counts
| File | Purpose |
|---|---|
ONBOARDING.md |
A first-read project map: what it is, main pieces, key flows, where to start, gotchas. |
memories/*.md |
Project facts, conventions, decisions, fragile flows, and commands worth preserving, with source hashes and verifiable claims. |
omissions.json |
Privacy-safe omission counts/categories, so the next developer knows gaps exist without seeing private content. |
manifest.json |
Package version, export commit, language, and privacy-filter accounting. |
handover.zip |
The transfer artifact to pass to the next person. It contains the full handover/ folder. |
Import also creates local audit files next to the extracted package:
handover/
import-plan.json # dry-run plan before memory writes
import-report.json # final import counts
import-receipts/ # one receipt per installed/rejected/blocked memory
| Guarantee | How the skill enforces it |
|---|---|
| No personal memory export | type: user memories are never written to the package. |
| Neutral project voice | Exported memories are rewritten without names, emails, usernames, or personal framing. |
| Transferable handoff | The output is a zip archive containing plain Markdown/JSON. |
| Dry-run before writes | Import shows candidate memories, evidence, action, and reason before writing. |
| Verified import | Every claim, cited file, path, or identifier is checked against the current repo before installation. |
| Import receipts | Every accepted, rewritten, rejected, or blocked memory gets a local receipt. |
| No blind overwrite | Existing memories are not overwritten silently on import. |
AI memories are durable notes an assistant builds while working with a developer across many sessions. They are not a chat transcript. They are compressed project knowledge that stays useful later: architecture decisions, local conventions, fragile flows, commands that actually work, files to avoid touching casually, and the reasons behind choices that may never have reached the README.
Over time, those memories become part of the project's working context. Some are captured deliberately; others emerge indirectly from repeated fixes, reviews, and debugging sessions. Losing them and starting from zero means losing the context a colleague built while doing the work, including the project understanding their AI learned alongside them.
You can clone an unknown repo and ask an AI to explain it. The answer starts from whatever is visible in the code today.
knowledge-transfer flips the order: the outgoing context is packaged first,
privacy-filtered, transferred, then verified before the next assistant learns it.
The newcomer gets a short human map and an AI that already knows the project's
important constraints.
Beta. The skill is prompt-driven and validated with the manual checklist in
TESTING.md. The package format is versioned through
manifest.version, so future import behavior can stay backward compatible.
Claude Code marketplace metadata lives in .claude-plugin/ and points to the
self-contained plugin in plugins/knowledge-transfer/. That plugin directory
also contains the Codex marketplace package, with real files rather than
symlinks so installed plugin caches are self-contained.
MIT (c) 2026 Ferdinando Bonsegna