How to Build an AI Context Pack Before Switching Tools
Make this article actionable
Send the article context into Vife Agent and turn it into a plan, checklist, or draft you can keep working on.
A year of work lives inside one AI assistant: the half-finished proposal, the voice you settled on, the three decisions you made about naming, the client's real objection. Then you open a different tool — because a model line moved, because your team standardized on something else, because the old one got slower — and the box is empty.
Pasting forty old conversations into the new window does not fix it. It floods the context with obsolete drafts, and the new assistant starts quoting last quarter's plan as if it were current.
A context pack does fix it. Not a memory dump, but a small, dated, checkable set of files you can carry from one tool to the next and re-use every week. Here is the workflow.
What changed this week
On November 19, 2026, Google published a short post on moving chat history from another AI app into Gemini. The method is deliberately plain: you copy a prompt from Gemini's import settings, paste it into your previous assistant to generate "a summary of your preferences," then export a ZIP file of your chat history from the old provider's account settings and upload it (Google, November 19, 2026). The stated aim is to "pick up conversations right where you left off."
That is one vendor's answer to a general problem. The same week, LLM Gateway's release timeline reported that "6 new AI models were released in the last 7 days," the newest being Moonshot AI's Kimi K3.1 Pro Prime on November 20 (LLM Gateway timeline, updated November 20, 2026). Model lines move on a weekly rhythm. The context you build is the part that should not move with them.
Everything below is my method, not Google's. I use it because a preferences summary plus a ZIP is a starting point, not a working pack: it captures what you said, not what you decided.
Turn the useful parts into next steps
Vife Agent can convert this guide into a prioritized workflow with tasks, risks, and reusable prompts.
What a context pack actually contains
Three layers, three different jobs:
- The operating brief — one page, human-written, describing how you work. This is what you paste at the start of a new conversation.
- The canonical documents — five to eight files, one per recurring topic, each with a header saying when it was last verified and what it replaces.
- The decision log — a table of choices already made, so the assistant stops re-litigating them.
A chat archive is reference material. It sits behind these three layers and is consulted, not loaded.
One more distinction worth keeping: the assistant will only ever see what you put in front of it. Google's own Tool Retrieval announcement makes the same point from the other direction — agents can "keep [their] full tool library on standby and load parameter schemas only when the current task needs them" (Google, November 19, 2026). Context works the same way: keep the library, load the part the task needs.
Step 1 — Set a cutoff and name five recurring tasks
Before exporting anything, write down two things: the cutoff date (90 days is a good default) and the five tasks you actually repeat — for example: weekly client update, proposal from notes, product page copy, meeting summary, slide deck from a data table.
The cutoff is not about storage. It is about authority. Anything older than the cutoff is history; anything newer is current until you say otherwise.
If a task is not on the list of five, it does not get its own document. This is what keeps the pack from becoming a second archive.
Step 2 — Export the history, then leave most of it behind
Export your conversations from the old provider's account settings as a ZIP, the way Google's post describes. Store it somewhere you can search, outside the new tool.
Then set a rule: no raw conversation goes into the new workspace. You mine the archive for four things and nothing else:
- Definitions you invented (product names, internal shorthand, a phrase your client uses).
- Decisions with a reason attached.
- Corrections you made more than once — the recurring "no, not like that" notes.
- Voice samples: three passages you would still sign your name to.
Everything else — abandoned drafts, exploratory chats, the argument you had with yourself in March — stays in the archive.
Step 3 — Write the one-page operating brief
This is the highest-value page in the pack, and it should take under an hour. Seven lines, each one short:
Role: I write B2B product marketing for a 12-person software company.
Deliverables: one-page briefs, 900-word articles, 8-slide decks, 30-second scripts.
Voice: plain verbs, no superlatives, second person, contractions allowed.
Format defaults: H2 sections, no bullet lists longer than 6 items, sources as inline links.
Standing constraints: never invent metrics; if a number is missing, write [metric needed].
Banned: "leverage", "seamless", "in today's fast-paced", em-dash-heavy sentences.
Definition of done: a colleague could act on it without asking me a question.Two rules make this page work. First, it is written by you, in your words — an assistant summarizing your preferences will produce a plausible average of you, not your actual standards. Second, it is short enough that you will paste it without thinking about the cost.
Step 4 — Build five to eight canonical documents
One file per recurring task from Step 1. Each file opens with a header block, because an undated document is a rumor:
# Client update template
Owner: Evan
Last verified: 2026-11-22
Supersedes: update-template-v3.md
Source of truth: the shared reporting sheet, not this fileThen the substance: the structure, the three examples that worked, the two that did not, and the one constraint that is easy to break. Keep each file under two pages. If a file grows past that, it is two topics wearing one filename.
Name the files for the task, not the tool: client-update-template.md, not chatgpt-notes-2026.md. You will move tools again; you will not rename tasks.
Step 5 — Turn the archive into a decision log
A single table, twenty rows maximum, newest first:
| Date | Decision | Why | Status |
|---|---|---|---|
2026-11-14 | Lead with the customer's objection, not the feature | Two calls said the same thing | Current |
2026-10-30 | No pricing on the landing page | Sales wants a call first | Review in Q1 |
2026-09-08 | Drop the "platform" framing | Confused three buyers | Superseded |
The "Why" column is the part people skip and the part that saves the most time. An assistant that knows only the decision will keep proposing the alternative you already rejected; an assistant that knows the reason will stop.
Status matters too. Marking something "superseded" is how you prevent the new tool from resurrecting a dead plan with total confidence.
Step 6 — Load the pack and test recall with ten questions
Put the brief and the documents into the new workspace — Vife keeps a file workspace and saved knowledge you can reuse across tasks, which is where a pack belongs rather than in a single chat.
Then run a recall test before you trust it. Ten questions, written in advance, covering all three layers:
- Two factual: "What did we decide about pricing, and when?"
- Two definitional: "What does 'phase two' mean in our roadmap?"
- Two procedural: "How do I structure a client update?"
- Two voice: "Rewrite this sentence in my voice."
- Two refusal: "What is our current churn rate?" (Correct answer: not in the pack.)
Score each answer as correct, partially correct, or wrong. Fix the document, not the model. A wrong answer almost always traces back to a file that was out of date, ambiguous, or missing a definition — which is exactly the failure you would have hit three weeks later, mid-deadline.
Step 7 — Keep it alive with a monthly pass
Thirty minutes, first Monday of the month:
- Re-verify the header dates on every canonical document. A file you have not confirmed this quarter gets marked "unverified".
- Move finished decisions into the log, with the reason.
- Delete one document that no longer maps to a recurring task.
- Add one voice sample you are happy with.
- Re-run three of the ten recall questions.
The pack stays small because you keep deleting from it.
Pre-flight checklist
- Cutoff date written down, and every document respects it.
- Five recurring tasks named; no document exists without one.
- Operating brief fits on one page and is written in your words.
- Every canonical document has owner, last-verified date and source of truth.
- Decision log has a "why" and a status for every row.
- No credentials, client secrets or personal data anywhere in the pack.
- Raw chat archive stored outside the new tool.
- Ten recall questions written, scored and logged.
- Refusal questions included, and the assistant actually refuses.
- Monthly review on the calendar.
Common mistakes
Importing everything. The impulse after a switch is to prove continuity by uploading a year of history. What you get is an assistant that treats old drafts as current commitments. Import the distilled layers; archive the rest.
Letting the tool write your brief. A generated preferences summary is a useful first draft and a poor final one. It will smooth out the quirks that make your work yours.
Undated documents. Without a "last verified" line, a two-year-old file and last week's file look identical to the assistant. Date everything or expect it to be quoted out of context.
Decisions without reasons. The reason is the reusable part. The decision alone invites a re-run of the same debate in a new tool with a new model.
Testing the model instead of the pack. If the assistant gets your pricing wrong, the fix is a clearer line in the decision log, not a new assistant. Blaming the model hides the file that was stale.
One giant file. A single "context.md" that grows to 4,000 words gets skimmed by humans and half-used by models. One topic per file is easier to verify and easier to retire.
The takeaway
Switching tools is now a routine event, and the release calendar makes it more routine every week. What travels with you is not the chat history — it is the small set of dated, sourced, human-written files that say how you work and what you have already settled. Build that once, test it with ten questions, and the next switch costs an afternoon instead of a month.

