A context pack is the bundle of truth you give a model before you care about wording. It is the opposite of opening a blank chat and performing intelligence.
If Bye Prompt has a single practical invention to steal, it is this: treat context as an artifact, the way you already treat a brief, a ticket, or a design folder.
What goes in a pack
Keep it boring. Five folders (or five headings in one doc) cover almost every job:
| Slot | What it is | Example |
|---|---|---|
| Goal | One paragraph of done | “Draft a 1,200-word explainer for merchants, US English, no legal advice.” |
| Must / must-not | Constraints | Must cite public docs. Must not invent GA dates. Must not name private customers. |
| Sources | Files and links the model may trust | Spec PDF, last week’s notes, three URLs |
| Examples | What good looks like | A previous article, a table format, a tone sample you own |
| Non-goals | Work you are not asking for | “Do not redesign the brand. Do not write tweets.” |
Optional sixth slot: tools allowed (browser, repo, calculator, none).
That pack is more valuable than a 900-word “you are an expert…” preamble. The preamble is theater. The pack is input.
Why packs beat prompt folklore
Models are good at continuing a distribution. If you do not show the distribution you want, they continue the internet.
A pack:
- Survives a new chat, a new model, a new teammate
- Makes evaluation possible (did it finish?)
- Reduces secret leakage — you attach this folder, not your whole disk (secrets)
- Turns “super prompts” into something honest: a compiled spec generated from the pack
How to assemble one in ten minutes
- Name the job in a single sentence you could put on a ticket.
- Dump raw material into a folder: PDFs, screenshots, CSV, markdown, URLs in a
sources.txt. - Write non-goals while you are annoyed — that is when they are honest.
- Add two examples of output you already like (yours, licensed, or public domain).
- Strip secrets. No keys, no customer PII, no unpublished financials unless the tool is actually bound to that environment.
- Freeze a version.
pack-2026-08-31/beats “the stuff in my downloads.”
Then — and only then — open the model. Point it at the pack. Your typed instruction can be short: “Execute the brief in GOAL.md. Do not use knowledge outside SOURCES.”
Packs for different jobs
Research. Sources are the product. The model’s job is to read what you gave it and flag gaps, not to invent a bibliography. Pattern: research.
Making. References (stills, brand kit, shot list) belong in the pack; “cinematic, 8k, trending” does not. Create By Prompt is the vertical. Pattern: making.
Buying. Spec, budget, “must work with X,” return window. The pack is a decision matrix, not a shopping personality. Buy By Prompt.
Ops. Runbooks, last incident note, what “green” looks like. Agents that skip the pack will “fix” the wrong service.
Size: enough, not your company
Stuffing 400 files into a window is not a pack. It is a haystack. Prefer:
- A manifest (what each file is for)
- The smallest set that could change the answer
- Retrieval only if you already have it — do not build a vector database for a two-page job
If the model keeps asking questions you already answered in a file, the file is not in the pack (or the instruction did not tell it to read the pack). That is a plumbing bug, not a “prompt engineering” bug.
Living packs vs frozen packs
| Frozen | Living |
|---|---|
| One-off article, a purchase decision | A product, a weekly report, a codebase |
| Zip and archive with the output | Update the pack when reality changes |
| Evaluate once | Evaluate every run against the same rubric |
Living packs need owners. If nobody is responsible for deleting stale files, the agent will eventually obey a 2024 constraint you no longer mean.
What a pack is not
- Not a jailbreak
- Not a persona (“you are a Harvard professor…”)
- Not your entire Google Drive
- Not a substitute for human review on irreversible steps
Checklist
Build a context pack
Three fields. Copy a GOAL.md-style pack. Do not paste secrets.