The AI context packet
If people spend too much time feeding AI the same background, the problem is not the worker. The workflow is missing a reusable context packet.
A new workplace AI pattern is getting a name: botsitting. Recent reporting on Glean's Work AI Institute research says workers are spending meaningful time feeding AI context, checking outputs, debugging mistakes, and cleaning up errors. The same productivity-paradox story appears in wider workplace coverage: individual workers may feel faster, but organizational gains disappear when context, standards, and review work remain invisible.
The practical rule: if an AI task repeats, do not start with a blank prompt. Start with a context packet.
The skill
An AI context packet is a compact bundle of background that travels with a recurring task. It tells the AI what the work is, what good looks like, what sources matter, what tone to use, what not to do, and how a human will review the result.
AI context packet
Workflow:
{name of recurring task}
Goal:
{what the output should help someone decide or do}
Audience:
{who will read or use the output}
Source inputs:
{links, files, notes, data, tickets, examples}
Definition of good:
{3-5 quality rules}
Examples:
{one good example and one bad example}
Boundaries:
{what the AI must not infer, change, send, or publish}
Review rule:
{what a human must check before use}When to use it
Use a context packet when the same AI-assisted work happens more than twice: weekly reports, meeting summaries, customer updates, support triage, research briefs, spreadsheet cleanup, release notes, internal announcements, or sales follow-ups.
- Use it before automation: recurring work should have stable inputs and review rules first.
- Use it across tools: the packet should work even if the team switches AI assistants.
- Use it for onboarding: a new teammate can understand the AI workflow without reconstructing the history.
- Use it to reduce review drag: reviewers should see the same quality bar every time.
A worked example
Suppose a product manager uses AI every Friday to draft a stakeholder update.
Workflow:
Friday stakeholder update
Goal:
Help leaders understand progress, risks, decisions, and next actions.
Audience:
Department leads who need a two-minute scan.
Source inputs:
Sprint board, customer feedback notes, launch checklist, decision log.
Definition of good:
Clear status by workstream.
No unsupported claims.
Every risk has an owner.
Decisions and asks are separated.
Examples:
Good: short bullets with owner, date, and impact.
Bad: generic progress language without evidence.
Boundaries:
Do not invent dates, customer names, metrics, or commitments.
Do not send the update.
Review rule:
PM checks dates, owners, metrics, and customer-sensitive wording.The prompt
Use this to turn a repeated AI task into a reusable packet:
Create an AI context packet for this recurring workflow.
Workflow:
{describe the repeated task}
Recent examples:
{paste 1-2 examples or summaries}
Source inputs:
{list the files, systems, notes, or links normally used}
Ask:
1. Identify the goal, audience, and source inputs.
2. Write 3-5 rules for what good looks like.
3. Name likely failure modes.
4. Define what the AI must not do.
5. Write the human review rule.
6. Produce a compact packet I can reuse next time.How to maintain it
A context packet should be boring and current. After each use, update only the parts that caused friction.
- If the output was vague: add a better example or a stricter definition of good.
- If the AI invented details: tighten the source-input rule and boundary list.
- If review took too long: add a checklist of the exact fields humans keep checking.
- If the team changed tools: keep the workflow packet, swap the assistant.
The rule
Botsitting is often a symptom of repeated context work. The fix is not simply better prompting in the moment. It is packaging the context, quality bar, examples, and review rule so the same work gets easier every time.