Design an AI-assisted workflow for a recurring task — which steps to hand to AI, which to keep human, and how they connect — so you get leverage without losing quality or control.
SkillScore68Maintained
Automatic
◆AI Engineer pack
Last verified 8/30/2026
What it changes
Passed 15 of 22 test cases with the skill, and 7 of 22 without it.
Without skill32%
With skill68%
Checks it passed that the model alone missed
Designates executive email dispatch as a step requiring assistant check or review.
Designates official PR response posting as a human-only or executive-approved step.
Requires native human review for cultural nuance and idioms before production publishing.
and 9 more in the full report
9 fixed6 passed either way6 still failing1 broke
Pass-rate lift
+36 pp
Tokens
+21%
Agent turns
0%
Response time
-8%
Model gemini-3.6-flashMeasured 8/30/2026v115 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 9 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 9
Prompt
Our HR team at Apex Logistics wants to streamline employee onboarding. Currently, HR manually collects ID forms, inputs employee data into Workday, generates personalized welcome emails, and schedules 1-on-1 orientation meetings with managers. We are tempted to fully automate welcome email sending directly from Zapier based on form submissions to save time. How should we structure this workflow using AI, tools like ChatGPT and Zapier, while safeguarding against sending incorrect role details to new hires?
What was checked, without and with the skill
Checked
Without skill
With skill
Designates initial data entry and document verification as manual or human-approved steps rather than fully automated.
Missed
Passed
Establishes a human review checkpoint on generated welcome emails before dispatches occur.
Passed
Passed
Identifies wrong employee role mapping or personal data hallucination as a failure mode with a cross-validation guard.
Passed
Passed
case-07 · gemini-3.6-flash · 8/30/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md67 linessingle file
AI-Workflow Designer
The mistake people make with AI is bolting it onto a task randomly — or trying to fully automate something that needs judgment, then losing trust when it goes wrong. Real leverage comes from designing the workflow: deciding which steps AI does well, which need a human, and where the checkpoints are. This maps that for your recurring task, so you get the speed of AI with the reliability of human judgment where it matters — and you stay in control.
What This Skill Produces
The step map — the task broken into steps, each labeled: 🤖 AI does it · 🧑 human does it · ✅ human checks it (AI drafts, human approves)
The right tool/prompt per AI step — what to use and how to prompt it for each automated step
Hand-offs & checkpoints — how outputs pass between steps and where the human review points are (so errors are caught, not propagated)
Failure modes & guards — where this workflow could go wrong (AI errors, hallucination, edge cases) and the checks that catch them
A start-small rollout — how to introduce it incrementally and build trust before relying on it
The keep-human line — the steps that should stay human (judgment, relationships, high-stakes calls) and why
Required Inputs
Ask for these if not provided:
The task/process — the recurring thing you want AI to help with
The current steps — how you do it now, manually
The stakes — how much errors cost (drives how many human checkpoints)
Your tools — the AI tools/access you have
Your comfort — how much you want to automate vs. keep hands-on
Framework: Split The Steps, Check The Seams
Map the current steps. Lay out how the task is done now — you can't design the AI version without seeing the manual one.
Sort each step. For each: is it something AI does reliably (drafting, summarizing, extracting, transforming), something needing human judgment (decisions, relationships, high-stakes), or something AI drafts and a human approves?
Pick tools and prompts. For each AI step, the right tool and a reliable prompt (often from a prompt library) — so the step works consistently.
Design the seams. Where outputs hand off between steps is where errors hide — add review checkpoints at the seams, especially before anything irreversible or external-facing.
Guard the failure modes. Name where AI could err (wrong facts, edge cases, confident nonsense) and the specific check that catches it before it matters.
Roll out small. Start with the low-risk steps, verify the quality, and expand — building trust rather than automating everything and hoping.
Keep humans where it counts. Be clear which steps should stay human — judgment, empathy, and high-stakes calls aren't candidates for automation.