From vague prompts to production-grade automations

How I coupled the ICE framework with n8n + OpenAI to grade hundreds of CVs in minutes

Screening a stack of resumes is mind-numbing work. Yet it is one of the fastest ways to destroy candidate experience if you get it wrong. Today I walked a small cohort through my end-to-end workflow that automates 90 % of that grind while still surfacing the strongest talent.

The recording of the webinar is below.

Below is the distilled play-by-play so you can replicate (and improve) it inside your own stack.

1. The hidden cost of sloppy prompts

Large language models obey the letter of our instructions. Give them a blunt order like “Assess this résumé” and they spit back filler praise no recruiter can use. Craft a richer instruction set and you suddenly have a digital analyst that flags GPA gaps, rates leadership depth, and decides whether to advance the candidate. That jump happens because of prompt engineering – the discipline of designing clear, constrained guidance for the model to follow. ibm.comsendbird.com

---

2. ICE – the scaffolding that never collapses

My favourite structure is the ICE framework:

• Identity – tell the model who it is. Example: “You are a McKinsey-trained talent acquisition manager with ten years résumé screening experience.”

• Command – state the single outcome you need. Example: “Score the attached résumé against five criteria and recommend advance or reject.”

• Examples / Context – provide rubrics, style references, or sample outputs that show the desired format.

ICE works because it mirrors how humans give strong briefs – role clarity, explicit ask, guardrails.

---

3. Proof via sandwiches

To burn the lesson into memory I lead with food photography – everyone remembers a perfect grilled cheese.

• Prompt 1 – “Create an image of a sandwich.” Result: a plastic-looking BLT with suspicious lettuce.

Prompt: Create an image of a sandwich.

• Prompt 2 – Identity and command added. A professional food-photographer persona plus a clear request for oozy cheddar yields a far tastier scene.

Prompt: You are a professional food photographer. Please generate a high-resolution image of a gourmet grilled-cheese sandwich on a wooden board with melted cheddar visibly oozing from the sides.

• Prompt 3 – Full ICE treatment. We specify a James Beard award-winning stylist, precise lens data, light direction, color palette, and rustic props. Now the image could slide straight onto the cover of Kinfolk.

Here’s the prompt (yes, it’s a little bit intense):

Prompt, using ICE

IDENTITY:

You are a Donna Hay award-winning food stylist and photographer known for rustic-elegant magazine spreads.

COMMAND:

Produce a photorealistic hero shot of a sourdough-bread sandwich cut diagonally and stacked, displaying layers of smoked turkey, aged cheddar melting at the edges, peppery arugula, and apple-fig chutney. Frame it at a 45-degree angle, f-stop 2.8 look, shallow depth of field, 50 mm lens style. Use soft natural side-light from the left to emphasize crumb texture and gloss on the cheese.

EXAMPLES:

Background: weathered dark-oak tabletop with scattered sea-salt flakes and a vintage butter knife for scale.

Color palette: warm browns and muted greens with natural highlights—avoid harsh neon tones.

Mood reference: Kinfolk magazine food photography, emphasizing cozy artisanal ambiance.

Maintain realistic proportions—no exaggerated cartoon elements or lens distortions.

Result: The sandwich should evoke farm-to-table authenticity while remaining clean and appetizing for a print feature.

The same escalation applies to text tasks. More context equals fewer hallucinations and tighter consistency.

---

4. Wiring the résumé screener in n8n

Here is the flow that turns those prompts into a functioning pipeline:

1. Form Trigger – candidates drop a PDF and their email address into a simple web form generated by n8n.

2. PDF-to-Text Node – the file is parsed so the model receives raw text, not binary gibberish.

3. OpenAI Node – we inject the ICE prompt alongside the extracted résumé text. Temperature is set to 0.2 for deterministic scoring.

4. Email Node – the model’s JSON output carries both subject and body, mailed automatically to the applicant.

5. (Optional) Sheet Update – scores are logged to Google Sheets for sorting and analytics.

Total build time once you know the pieces: about fifteen minutes. The same pattern handles cover-letter triage, inbound RFP scoring, or any document review at scale.

---

5. Resources to replicate the demo

• Affiliate link for n8n – grab your own workspace and help support future breakdowns: https://n8n.partnerlinks.io/j1mut17ogftx

• DIY hosting guide – USD 5 per month – two ways to spin up n8n on a budget rather than pay vendor rates: https://www.aineversleeps.net/blog/from-us27-cloud-to-5-diy-n8n-in-two-ways

• Download the exact JSONs – import, add your API key, and run. The bundle sits behind the paywall below.

---

6. Why the paid community pays for itself

Inside the members-only area you unlock resources that would cost well over a grand if you tried to source them piecemeal:

• Live webinars, two to three each month – I share my screen, build workflows from scratch, and troubleshoot member projects in real time. Comparable small-group coaching typically runs about $200 a month.

• Full recording archive – every session is saved so you can binge whenever it suits you. The growing library already carries at least $500 in standalone value for on-demand training.

• Prompt vault – fifty-plus field-tested prompts ready to drop into sales copy, data cleanup, customer support, and more. Tap into it here: https://community.aineversleeps.net/t/the-ultimate-ai-prompt-library-and-how-to-use-it/23. A curated prompt pack of this depth would retail near $75.

• Workflow treasury – more than two hundred n8n flows covering CRM sync, lead enrichment, content repurposing, and AI agent orchestration, all pre-tested end-to-end: https://community.aineversleeps.net/t/welcome-to-the-n8n-workflow-vault/22. Hiring a freelancer to recreate even a fraction would set you back $500 or more.

• Included software seats – you get a ready-to-run n8n instance and a LobeChat account that lets you hit every major LLM from one chat window. Equivalent SaaS licences run roughly $55 each month.

Stack those numbers and the package tops $1,300 in monthly utility. Membership is still $49.95 per month for a few more days, but on 13 July it rises to $59.95 for new sign-ups. Join before the deadline and that lower rate is locked in for the life of your subscription.