Lurking in the Shadows: From One Big Prompt to Repeatable AI Workflows
A recap of Connor Callahan’s Below the Waterline talk on giving powerful AI agents smaller jobs, clearer handoffs, and workflows people can inspect and refine.
A visual field guide
From a prompt to a workflow
Follow one evolving diagram from crowded context to coordinated, repeatable work. Scroll at your own pace; move back to revisit a handoff.
01 / The context stack
One prompt keeps growing
A request starts with instructions. Then evidence, earlier outputs, exceptions, and another request accumulate in the same context. Relevant instructions can become harder to track as the stack grows.
One context holds the instructions and everything collected along the way. 02 / Focused work
Give each job a boundary
Separate the stack into repeatable jobs. Research receives a question and returns evidence. Compose receives evidence and returns a draft. Review and triage have their own inputs and outputs, too. A saved workflow describes the handoffs—not one ever-longer request.
Four focused jobs, each with an explicit input and a useful output. 03 / Coordination
Fresh context is not lost state
A coordinator hands the relevant work to a focused agent, then carries the result to the next step. A separate state record keeps evidence, drafts, and decisions between steps. Each agent can start with a smaller context without forgetting where the workflow has reached.
The coordinator routes work; the state record preserves the handoff. 04 / A chosen checkpoint
Make results—and pauses—visible
A draft is an output, not a completed external action. In this illustrative workflow, a chosen human checkpoint pauses the path before sending. Scrolling carries the example through review and into a receipt; it does not approve or send anything. Approval gates are a workflow choice, not a requirement for every Shadow.
An illustrative pause, review, and resume. No real approval or send occurs. 05 / The outreach loop
Carry what you learn into the next pass
Research → compose → approval → send → triage is one possible healthcare outreach pattern. Shared state carries the evidence, draft, review decision, and response forward. Triage feeds the next research pass, so the workflow repeats with what it has learned—not from an empty page.
A generic outreach pattern. Shared state connects one cycle to the next.
The workflow is the structure. The agents do the focused work. State connects the steps.
A capable AI agent can do impressive work, but getting it to repeat that work takes more than a longer prompt. In Lurking in the Shadows — Below the Waterline, Connor Callahan examines how to move from asking one agent to handle everything to smaller jobs with explicit handoffs and room for human judgment.
The September 29 video runs 50 minutes and 40 seconds. This recap, published October 5, follows its argument for structuring the work around powerful agents.
Why one big prompt becomes difficult to manage
Preparing a client proposal might require an agent to find relevant correspondence, read discovery notes, choose the correct template, assemble a draft, format a PDF, and eventually send it. Each task is reasonable. In one continuous conversation, the agent has to keep track of both the evidence and the process.
As the conversation grows, instructions can fade into the background. The agent may drift from the original goal or revisit a finished step. Old decisions and irrelevant material compete with useful source evidence for attention. A larger context window has more room, but it does not establish which instruction matters now or whether an action has already happened.
Carrying more context through repeated calls can also increase resource use, depending on the workload and implementation. That does not establish inevitable exponential cost growth. Retaining everything is not always the most economical or reliable way to finish a bounded job.
Save the workflow, scope the work
The talk proposes saving the workflow as distinct steps. A person defines what needs to happen, what each step receives, and what it should produce, instead of explaining the entire process each time.
A coordinator advances the process. Fresh, scoped agents can handle individual steps with the relevant instructions and inputs, without inheriting every detail of the preceding conversation. A step that gathers evidence can pass it to another for reasoning, followed by one that prepares the deliverable. The workflow keeps those handoffs together without requiring one agent to hold the whole project in conversational memory.
The accompanying slides credit Geoffrey Huntley’s Ralph Wiggum loops as an inspiration: plan the work, then tackle bounded tasks in fresh sessions. Shadows extend that idea beyond coding loops to everyday workflows.
Model selection depends on the task. Straightforward collection or classification may call for a different model than a nuanced proposal draft. Choosing a model to fit the work does not require using a smaller model for every step.
Make handoffs inspectable
Each step needs an identifiable output that the next can use: gathered correspondence, organized notes, a draft, or a record of an action. Splitting up the work is only useful if those handoffs are clear.
Keeping outputs and state outside a single conversation lets people inspect the workflow. Receipts help distinguish preparation from completion. A drafted email is not a sent email; an intended delivery is not evidence of delivery. Someone picking up the work needs to know which happened.
The coordinator can check progress and guide the next step. It cannot promise flawless or deterministic execution. Clear outputs make failures easier to locate, but people still need to check the relevance of the evidence, the accuracy of the draft, and whether the next action is appropriate.
The proposal example
In Callahan’s proposal example, the workflow gathers relevant emails and discovery notes, uses the correct template, and produces a draft proposal. Depending on the design, later steps can format a PDF and send the proposal after approval.
Callahan’s live demonstration walked through configuring an on-demand workflow. It took a client name as input and used a project to hold its context and drafts. The sequence loaded the template, then drew on Gmail, Outlook, and discovery notes before drafting. A recurring kind of work can use this structure without running on a daily schedule.
The demonstration covered setup; it did not establish that a proposal had been delivered. Preparing the workflow, producing a draft, and making an external commitment remain separate milestones.
A healthcare outreach loop
An accompanying slide illustrates a healthcare outreach loop with a shared Google Sheet tracker. The pipeline is an example, not a description of a currently running configuration or any particular organization’s contacts:
- 8:00 a.m.: Find and qualify. Identify potential prospects and assess whether they belong in the pipeline.
- 8:30 a.m.: Prepare outreach. Draft messages, review them, approve the intended sends, then send and log the results.
- 5:00 p.m.: Triage responses. Read incoming replies, classify them, and flag items that need attention.
The shared tracker gives people a record to review across prospecting, outreach, and responses, instead of leaving those activities in disconnected conversations.
The workflow’s risks and responsibilities determine where approval gates belong. This example includes review before sending. It does not establish an approval requirement for every step of every Shadow.
Feedback and human judgment
Feedback can address the step that needs refinement. If a draft misses the intended tone or qualification uses the wrong criteria, a correction can target that responsibility without rewriting the entire prompt. This supports deliberate iteration; it does not guarantee automatic improvement.
An inspectable process can handle repetitive assembly so people can spend more attention on relevance, exceptions, and commitments. Human judgment still belongs in the work; the workflow gives people clearer places to use it.
Sources: Keel3’s editorial recap of the September 29 video, Lurking in the Shadows — Below the Waterline, and the accompanying four-slide materials. The recording is 50:40; this article’s publication date is October 5, 2026.