Ask a chatbot, “What should I automate?” and it may propose a new app before it understands your working day. In Dylan Davis's walkthrough, the better starting point is evidence: find work you actually repeat, examine its moving parts, then try one small handoff. His 10-slide presentation supplies the two prompts behind that method. Below, we turn the slides into a usable sequence and add the access and review checks a real team needs.
Start with a 90-day, permissioned inventory of recurring work. For one candidate, count its steps, sources and judgment calls. Interview the person who does it, test a documented process once with AI, review the result, and only then make the instructions reusable or scheduled. Frequency is evidence of repetition, not proof that a task should be automated.
Watch Dylan's Walkthrough
Useful jump points: the automation trap (00:00), evidence sources (03:20), the first prompt (05:23), three numbers (07:10), the interview (08:15), and the feedback loop (10:48).
Map the Work Before the Tool
Davis breaks a job down as role → activities → tasks → steps. “Manage clients” is too broad to hand over. “Prepare the Monday client status note” has a trigger, inputs, decisions and a concrete output. Calendar events, sent email and previous AI chats can reveal repetitions that memory misses, but an event or email may represent only part of the work. Confirm the task with the person doing it.
His example looks across the previous 90 days. Start with only the sources you are allowed to inspect: an approved connection or a small, redacted export may be enough. Check the connected account, provider permissions, colleague or client consent, retention rules and your organization's policy before granting access. OpenAI's connection guide says availability varies by plan, region, account and workspace, and that the connected account's existing access still applies. Do not upload a full mailbox or team chat just to discover a handful of candidates.
Prompt 1: Find Recurring Work, With Receipts
The first prompt is for discovery only. Davis asks for a one-line description, counts and dates, specific source evidence, a maximum of ten tasks ranked by frequency, and no automation ideas yet. This copyable version is our adaptation of slide 5; use the presentation for Dylan's original wording.
Inventory the repeats
Evidence first; no tool recommendations.
Inspect at least two cited examples before trusting the ranking. A repeated meeting invitation might be a standing calendar block rather than a repeatable deliverable. Missing source access should reduce confidence, not be filled with invented evidence.
Choose a First Task With Three Numbers
For each promising repeat, Davis counts the number of steps, distinct information sources and human judgment calls. Fewer of each usually means a simpler learning pilot. The counts are an orientation heuristic, not an ROI score: a three-step payroll change may be riskier than a ten-step internal draft.
| Count | Ask the person doing the work | Why it matters |
|---|---|---|
| Steps | What actions happen from trigger to finished output? | More handoffs create more places to miss context. |
| Sources | Which inboxes, files, systems or people supply facts? | Each source adds access, freshness and reconciliation work. |
| Judgment calls | Where do you choose, interpret, approve or escalate? | Unwritten judgment is where a plausible draft can go wrong. |
Then apply a separate gate: is the task high-impact, regulated, confidential, client-facing, or able to spend money or change records? If yes, keep an explicit human approval step even if its three counts look low. Frequency helps find candidates; value, privacy and error cost decide whether to proceed.
Prompt 2: Turn One Task Into a One-Page Spec
Once you pick a candidate, Davis recommends showing a real finished example and having the AI interview you one question at a time, up to 15 questions. The output is a one-page specification covering steps, sources, judgment and what good looks like. This is our adapted version of slide 7, not a verbatim transcription.
Interview the task owner
Capture decisions before delegating them.
Read the spec yourself. If a judgment rule is missing, add it before testing. An example output is useful for calibration but is not permission to reuse private client material elsewhere.
Do One Run, Review It, Then Make It Reusable
- Run a bounded rehearsal. Open a new chat with the approved spec and a small, representative input. Have AI draft the result without sending messages, editing live records or publishing.
- Compare with your own work. Check factual sources, omissions, tone, decisions, exceptions and time saved. Record every correction as a rule or missing input, not just “try again.”
- Store the working instructions. After a successful rehearsal, keep the spec and revisions in a reusable prompt or a ChatGPT Project. Projects can hold chats, files and project instructions; another product's “skill” may be a different feature.
- Schedule only after the manual pilot works. If the task truly needs a timetable, check Scheduled task eligibility, app access, usage and approval behavior. A scheduled task created in a Project cannot access files uploaded to that Project, according to OpenAI's current help page, so test the exact data path before relying on it.
- Keep a review loop. Sample future outputs, track mistakes and drift, and define when the task stops for human review. Davis presents a path toward fewer incorrect judgment calls; that is a goal, not a guarantee they will reach zero.
The distinction matters: a repeatable instruction is not the same as an autonomous workflow. Letting a system read evidence, draft a note and ask for approval is a useful first deployment even when a fully unattended version would be unsafe.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Why a generic automation question misses the work |
| 03:20 | Calendar, sent email and chat evidence |
| 05:23 | List recurring tasks with source references |
| 07:10 | Count steps, sources and judgment calls |
| 08:15 | Interview for a one-page task specification |
| 10:48 | Refine the workflow from human feedback |
Two Companion Videos From the Presentation
I Gave ChatGPT 10,000 Emails. Here's What It Found.
Dylan links this email-analysis video from the slide about using calendar, sent email and chat history as evidence. Watch it if the hard part is finding work you actually do more than once. Its 10,000-email example is not a reason to grant broad inbox access for your own pilot.
Claude Asked Me 15 Questions. Now It Decides Like Me
The second companion video appears on the deck's judgment slide: make a rule explicit, compare AI's call with your own review, and revise the rule. Use it after you have a real task and a first output to compare; its title is a demonstration claim, not proof that judgment can be delegated without oversight.
Sources and Limits
Checked 9 November 2026. The account of Davis's method comes from the main video, user-supplied transcript and his presentation. The two copyable prompts above are JQ AI SYSTEMS adaptations, not Dylan's verbatim slide text. OpenAI's current connection, Projects and Scheduled tasks guidance supports the product-access notes. Features, permissions and limits can change; check your own account before building a workflow around them.