Working with assistants

The shape of every agent run

Every productive run, no matter the agent, follows the same three-part shape:

1. You hand the agent a brief and the right sources.

   Outcome, audience, constraints, context. The brief tells the agent what good looks like.

2. The agent produces a structured first draft.

   Format, sections, tone — already roughly right. The first draft is a starting point, not the answer.

3. You refine in place, one change at a time.

   Small, specific revisions. Each revision keeps the context the agent has already built up.

Master this loop and 80% of your agent runs land in three turns or fewer.

Briefing the agent well

A great brief is short but specific. The order doesn’t matter — completeness does.

The four pieces

Piece	One-sentence rule	Quick example
Outcome	Describe what you want to ship, not what you want to discuss.	”A two-paragraph internal update for our exec team.”
Audience	Name the audience so tone, length, and assumptions line up.	”Engineering leads and the VP of Product.”
Constraints	State the hard rules up front: tone, length, format, exclusions.	”Plain tone. Under 250 words. Skip launch dates that aren’t confirmed.”
Sources	Attach the material the agent should read.	”Standup notes from last week. Thursday’s customer call transcript.”

What to leave out

• Background you can’t verify. If you’re not sure last quarter’s number was X, don’t put it in the brief — let the agent ask or pull it from a source.
• Filler context. Long preambles (“here’s some background on our team and how we work…”) dilute the brief without changing the output.
• Multiple unrelated asks. One brief = one outcome. Run a second agent for the second thing.

Attaching the right sources

Agents work better when they read your material rather than guess at it. Three ways to give them sources:

1. Upload a file directly

PDFs, text files, transcripts, spreadsheets — drag them in and reference them by name in your prompt.

“Use the attached customer call transcript. Skip anything not mentioned in the call.”

2. Connect an app

When you’ve connected Google, Jira, Confluence, GitHub, Slack, or another platform, you can point the agent at a folder, a Jira project, a Confluence space, or a Slack channel. The agent reads what it needs in the moment.

“Reference the ‘Q3 launches’ Confluence space and the open Jira tickets tagged releases-q3.”

3. Paste it in line

Sometimes the fastest move is to paste the source — a meeting summary, three backlog items, a competitor’s product page — into your prompt.

Iterating without losing context

The biggest mistake people make with agents is starting over. Don’t.

Refine in place

Tell the agent what to change. Keep the conversation going. The agent already has the brief, the sources, and the previous draft in working memory — restart and you throw all of that away.

Good refinement prompts:

• “Shorten paragraph two to one sentence.”
• “Move the risks section to the top.”
• “Replace the second bullet with a quote from the transcript.”
• “Make the tone slightly more direct.”
• “Add a one-line summary at the top for the exec audience.”

One change per turn

If you tell the agent to “make it shorter, more confident, and replace the second bullet” — and you don’t love the result — you can’t tell which change to undo. Change one thing per turn.

When to restart

Restart when the brief itself was wrong:

• You picked the wrong agent for the outcome.
• The audience changed mid-task.
• The sources you gave were the wrong ones.

In those cases, a fresh start with a corrected brief is faster than refining around a flawed foundation.

Switching agents mid-task

Sometimes a draft is great but needs a different specialist to finish.

Common handoffs:

• Drafting agent → Editing agent — one agent writes a clean draft, another tightens it for tone or length.
• Research agent → Drafting agent — research surfaces the facts, drafting turns them into a deliverable.
• Refining agent → Comparison agent — refine a backlog, then compare to last sprint’s commitments.

To hand off:

1. Copy the output from agent one (or save it to your workspace).
2. Open the second agent.
3. Paste the output in as a source. Add a one-line brief for what the second agent should do.

For more complex handoffs — three or more specialists in sequence — that’s a sign you should be using VDF AI Networks instead of chaining agents by hand.

Reading and trusting the output

Agents produce confident-looking output even when they’re wrong. Always read the draft.

A quick read-back checklist

• Did the agent stay inside the sources you gave it?
• Are the numbers, dates, and names verifiable?
• Does the tone match the audience you specified?
• Are the constraints respected (length, format, exclusions)?
• Would you sign your name to this?

If yes to all five — ship it.

Saving and sharing winning briefs

The same brief, run twice, gives you (roughly) the same shape of output. That’s the point.

When a brief produces a result you love:

1. Save the brief as a template in your workspace.
2. Note the agent it worked with — different agents respond to the same brief differently.
3. Share it with the team if the task is one others repeat. Your win becomes their starting point.

Over a quarter, your team builds a library of known-good prompts. New hires inherit your patterns instead of figuring them out from scratch.

Where to go from here

• Use cases — six worked scenarios with brief + source + outcome.
• Tips & best practices — the patterns top teams use.
• FAQ — common issues and when to choose Chat vs Agents vs Networks.