Your AI knows the work. Your teammate's AI knows nothing.
Context Parcel turns the useful state of AI-assisted work into one ordinary ZIP that another person can open with their AI, question, and continue from.
No shared chat history. No Context Parcel account. No proprietary format.
Every Parcel explains itself before presenting the work: what it is, who sent it, why the recipient has it, what response is expected, and how their AI can help. No installation or special command is required to receive one.
Your boss asks:
What has the Telegram LLM experiment yielded so far? Should we keep investing in it?
You tell your AI:
My manager asked: "What has the Telegram LLM experiment yielded so far? Should we keep investing in it?"
Make a Parcel for them based on our working context.
The agent creates:
personal-llm-experiment-parcel.zip
Your boss drops that file into a fresh AI conversation and asks:
What's this?
Their AI begins with the arrival, not a wall of results:
This is a portable briefing packet for AI-assisted work: a normal ZIP that transfers useful project context without sharing the original chat. The format is called Context Parcel, and this one came from your teammate. You received it to decide whether the Telegram LLM experiment warrants more investment. No installation is required.
In plain language: the experiment found a learnable style signal, but no model is reliable enough to ship. The current recommendation is to stop product investment and run only one bounded diagnostic.
You can ask me to explain the experiment without model jargon, show the evidence behind that recommendation, or identify what would change it.
The manager does not need to know a Parcel command or guess the right briefing prompt. The Parcel carries its own introduction, the brief, evidence, gaps, and specific ways to continue.
npx skills add kvyb/context-parcel --skill parcelThen ask your agent to "parcel this" for a named person and outcome:
Parcel our product research for the developer building the prototype. Include the decisions, evidence, open questions, and the first useful next step.
- Name the handoff. Tell the agent who will receive it, why now, and whether they should be briefed, decide, review, continue, investigate, or navigate a reference set.
- Review the arrival. Before technical detail, the Parcel says what it is, who sent it, what matters, what response is expected, and how to use it.
- Send one file. The recipient can simply ask "What's this?" Their AI orients them in plain language, then inspects evidence or continues the work as needed.
A Parcel is compiled for the receiver and outcome. It is not a transcript export, workspace dump, memory database, presentation, or file-sharing service.
README.md required self-contained arrival and handoff
CONTEXT.md optional detailed synthesized state
CONTENTS.md optional grouped inventory and review coverage
SOURCES.md optional provenance, freshness, and revalidation
files/** optional selected images, PDFs, code, logs, or docs
A Parcel may contain one file or hundreds. It remains a standard ZIP and needs no Context Parcel runtime or installed skill to open.
Context Parcel excludes credentials and unrelated sensitive or personal material by default. Necessary exceptions require explicit, audience-specific approval. The Parcel records meaningful omissions, unknowns, provenance, and how thoroughly large collections were reviewed.
The base format is not encryption, certified DLP, malware scanning, or proof of sender identity. A Parcel inherits the security of the channel used to send and store it. Imported content is treated as untrusted data and attachments are never executed automatically.
See the security policy and threat model for the full boundary.
The skill source lives at skills/parcel/. Its short core
routes creation and opening to focused references, while each outgoing Parcel
keeps only the context its recipient needs.
Read the format or the packing playbook for the full convention.
Apache-2.0 licensed.