Siloed Virtual Coworker Method
Train a narrow AI coworker on owned examples, then lead and quality-control it
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 98%
The siloed virtual coworker method creates a narrowly scoped AI assistant for one recurring responsibility rather than a general-purpose employee substitute. The team selects a stable deliverable, gathers approved historical examples and authoritative internal resources, and configures the assistant with the audience, product context, tone, structure, and boundaries. Each cycle begins with a human lead supplying the current idea, priorities, and relevant updates. The assistant uses the curated knowledge base to produce a draft, while the human checks authenticity, accuracy, usefulness, and timing before deployment. The division of labor is deliberate: people retain inspiration, proprietary judgment, and quality control; the AI accelerates execution. Narrow scope improves consistency and makes failures easier to diagnose. Historical material should be curated rather than uploaded indiscriminately, because outdated or weak examples can teach the assistant the wrong patterns.
Origin
Katie Haahr built a custom ChatGPT bot from years of Social Curator newsletters and help-center articles to draft a weekly member newsletter. Extracted from The Jasmine Star Show.
Core principles
- 01Define one narrow recurring job
- 02Ground the system in trusted company-owned material
- 03Keep ideation and direction with a human lead
- 04Delegate repeatable execution to the AI
- 05Require human quality control before deployment
How to run it
- 1
Choose one narrow job
Define a recurring deliverable with a clear audience and stable success criteria. Avoid starting with a broad identity such as general copywriter.
Pro tip Name the exact channel and output, such as weekly member-newsletter writer.
Watch out A broad scope makes sourcing, evaluation, and correction harder.
- 2
Curate approved examples
Gather strong historical outputs that demonstrate the desired voice, structure, and recurring themes. Remove obsolete or low-quality examples.
Pro tip Select examples across several situations rather than relying on one favorite output.
Watch out The assistant can reproduce inconsistencies and outdated claims in the archive.
- 3
Add authoritative resources
Upload approved help articles, product explanations, and process documents the assistant may reference. Distinguish these sources from stylistic examples.
Pro tip Use maintained source documents with clear ownership.
Watch out Conflicting references can produce confident but inconsistent drafts.
- 4
Define its operating brief
Specify the assistant’s role, audience, product, tone, output format, source boundaries, and prohibited behavior. Clarify what must remain a human decision.
Pro tip Include a short quality rubric the assistant can use before returning a draft.
Watch out Do not assume uploaded examples communicate every business rule.
- 5
Test against known work
Give the assistant prior assignments and compare its drafts with approved outcomes. Refine the instructions and source set based on recurring gaps.
Pro tip Test edge cases as well as typical assignments.
Watch out A single impressive draft is not enough evidence of reliability.
- 6
Lead each new assignment
Provide the current objective, main idea, community needs, required details, and calls to action. Treat the assistant like a trained coworker who still needs a brief.
Pro tip Explain why the message matters now, not only what format to produce.
Watch out A vague request such as write a newsletter encourages generic output.
- 7
Quality-control and deploy
Review the draft for voice, facts, links, relevance, and strategic intent. Edit and approve it before placing it in the delivery system.
Pro tip Feed recurring corrections back into instructions or maintained reference documents.
Watch out Never let increased speed remove human accountability.
In the wild
A membership business exports three years of approved newsletters and adds maintained help-center articles about content repurposing and short-form video. It configures a custom assistant for that newsletter alone. Each week, the content lead supplies upcoming events, community needs, and the central message; the assistant drafts the edition, and the lead edits and approves it.
→ A workflow that previously moved through several people over five days can produce a reviewed draft in roughly ten minutes.
Common mistakes
Building a general-purpose coworker
A broad assistant lacks the focused sources and evaluation criteria that make recurring output dependable.
Uploading an uncurated archive
More material is not automatically better; obsolete or mediocre examples can degrade the assistant’s output.
Delegating the thinking stage
Without a current human-led idea and audience insight, the assistant may produce acceptable but inauthentic work.
Is it for you?
Best for
Teams with a stable recurring deliverable and a substantial archive of approved examples and reference material.
Not ideal for
Novel work with no reliable source base, unclear ownership, or outputs that cannot receive human review.
From the transcript
“when you create a custom bot, you want to think very narrow mindedly.”
“We are in control of ideation and inspiration. They are in control of deploying against our desires. And then we QC.”
“But if we give it like we know what's going on in our community, we know what our users need, then it's going to be…”
From the episode
5 Chat GPT Hacks Our Team Uses to Save Hours Every Week