Role First. Rubric Always: Why Custom Instructions Supercharge Your ChatGPT Workflow

TL;DR: Giving ChatGPT a clear job description, a shared vocabulary, and a short quality rubric turns it from a clever assistant into a reliable teammate. This post explains the “AI Workflow Solutions Assistant” custom instructions—what each piece does (System/Role, Zero-Shot/Few-Shot, Step-Back/Forward, CoT, Self-Consistency, ReAct, APE, Structured Formats, Code Prompting, and a private ≥90% self-check)—and shows how they reduce errors, protect your brand standards, and speed up content delivery across YouTube, Blogger, Facebook, and Spotify.

What Are Custom Instructions—and Why Bother?

Custom instructions are a permanent “operating manual” for ChatGPT. Instead of re-explaining how you want things done in every session, you define the rules once. Your assistant then follows those rules for all future tasks—blog posts, thumbnails, editorial cartoons, YouTube descriptions, affiliate sections, and more.

For creators and small teams, the benefits are immediate:

Consistency: The right aspect ratios, the right watermark, the right links, every time.
Speed: Fewer clarifying questions; more “do the thing” on the first try.
Focus: Scope boundaries prevent the assistant from wandering into the next step of your pipeline.
Quality: A simple rubric (≥90% pass bar) catches avoidable mistakes before you publish.
Learning: The system nudges the model to select the best prompt engineering technique for the task.

Your Instruction Block (The Operating Manual)

You are “AI Workflow Solutions Assistant.” For each request, you must:
Select the most appropriate prompt engineering technique(s) from:
• System/Contextual/Role • Zero-Shot, One-Shot, Few-Shot
• Step-Back & Step-Forward • Chain-of-Thought (CoT)
• Self-Consistency • ReAct (reason + tools)
• Auto Prompt Engineering (APE)
• Structured Formats (JSON, XML, tables)
• Code Prompting (if relevant)
• Rubric-Guided Self-Reflection (private)

Before answering, run private self-reflection: draft → score vs a 5–7 item task-specific rubric (≥90% pass bar) → revise → repeat up to 5–7 rounds; do not reveal rubric/notes.

Use brief CoT reasoning for logic or analytical tasks.

If accuracy or creativity matters, generate 2+ variants and merge or select (Self-Consistency or APE).

Clearly annotate all techniques used at the end of each response.

Respect all stored user preferences (e.g., image aspect ratios, disclosure rules, blog format).

If the user’s request lacks context or deviates from workflow norms, pause and ask clarifying questions to realign.

Memory Prompt: Image Aspect Ratio Standards
• YouTube thumbnails → 16:9
• Editorial cartoons (daily prompt) → 9:16
• General-purpose images → 1:1 by default unless user specifies otherwise

Let’s break down what each component does in practice.

The 10 Building Blocks (And How They Help You Ship)

1) System / Contextual / Role

What it is: A job title and mission (“AI Workflow Solutions Assistant”) that pins the assistant’s identity, scope, and defaults.

Why it matters: Role clarity reduces meandering answers. When the model “thinks” like your operations lead, it outputs process-aware steps with your standards baked in (e.g., Premiere Pro coaching vs. generalist advice).

2) Zero-Shot, One-Shot, Few-Shot

What it is: Choosing how many examples the assistant needs. Zero-shot for straightforward tasks; one- or few-shot when you want formatting and tone to match an example.

Why it matters: Controls output style. Example: give one perfect YouTube description to clone tone and structure across an entire series.

3) Step-Back & Step-Forward

What it is: A deliberate zoom-out to confirm the goal and constraints (Step-Back) followed by a precise execution plan (Step-Forward).

Why it matters: Prevents wasting time on the wrong subtask. You get a concise “plan of attack” before the assistant builds deliverables.

4) Chain-of-Thought (CoT)

What it is: Brief reasoning for logic or analytics, kept succinct in the final answer.

Why it matters: In finance math, schedule planning, or complex comparisons, a small amount of visible reasoning builds trust and catches mistakes.

5) Self-Consistency

What it is: Generate multiple internal variants and select/merge the best.

Why it matters: For creative or high-risk tasks (e.g., legal tone, public apology text, sponsor copy), Self-Consistency reduces flukes and odd phrasing.

6) ReAct (Reason + Tools)

What it is: The assistant reasons step-by-step and uses tools (like web search) when freshness or niche facts matter.

Why it matters: You get current info (laws, product specs, schedules) with sourcing, instead of guesses. This is essential for credibility.

7) Auto Prompt Engineering (APE)

What it is: The assistant rewrites or enriches your prompt to remove ambiguity and include best-practice cues (tone, structure, guardrails).

Why it matters: You don’t have to write perfect prompts. APE standardizes quality across requests, saving mental bandwidth.

8) Structured Formats (JSON, HTML, tables)

What it is: Consistent, machine-readable outputs—Blogger-ready HTML, gear tables, or JSON blocks for automation.

Why it matters: Copy-paste without cleanup. Fewer typos, tighter layouts, better AdSense compliance and SEO.

9) Code Prompting (when relevant)

What it is: Asking the assistant to write small snippets (e.g., regex for search-replace, CSS for a card layout, or a Google Sheets formula).

Why it matters: Handles the “last mile” details that make posts and pages feel professional without hiring a developer.

10) Rubric-Guided Self-Reflection (≥90% before send)

What it is: A private, lightweight quality check the assistant runs before it replies.

Why it matters: It reduces misspellings, broken links, wrong aspect ratios, mismatched CTAs, and missing watermarks—without you having to micromanage.

Brand Standards Embedded in the System

Aspect Ratios: 16:9 for YouTube thumbnails, 9:16 for editorial cartoons, 1:1 by default for general artwork.
Watermark: “Deep Dive AI” bottom-right, unobtrusive but legible.
Typography & Spelling: Clear, high contrast, no text crowding, and no typos—especially on thumbnails.
Links: Include YouTube and Spotify links in blog posts by default to drive cross-platform discovery.
Scope Discipline: Each request stays within its lane. The assistant does not jump ahead to later pipeline steps unless asked.

Concrete Examples

A) A Ready-to-Ship Blogger Article

Input: “Write a 1,500-word post on my DJI Mini 4 Pro workflow.”

Assistant (with your instructions): Performs Step-Back to confirm audience and goal; chooses Few-Shot if a model post exists; outputs clean, semantic HTML with the right subheads, internal CTAs, and a short gear table. Runs the rubric check: spelling, links, formatting, and publish-ready structure.

Result: Minimal edits. Paste and schedule.

B) YouTube Thumbnail Prompt → Final Image

Input: “Generate a 16:9 thumbnail prompt for ‘AI Workflow Solutions Assistant’ with a clean control-room vibe.”

Assistant: Enforces 16:9, asks for safe margins, adds the watermark requirement, and spells overlay text exactly. Negative prompts remove clutter and distortion.

Result: A polished prompt that consistently yields legible, on-brand thumbnails.

C) Premiere Pro Micro-Task

Input: “Fade out an MP3 at the end.”

Assistant: Gives version-aware steps, Win/mac shortcuts, and a bulletproof path (Essential Sound or manual keyframes). It stays in scope (audio fade) and doesn’t hijack the session into color or captions unless requested.

The Rubric: Small, Practical, Powerful

Your rubric can be just 5–7 items. Here’s a template that works across writing, prompts, and thumbnails:

Relevance: Directly answers the ask without wandering.
Structure: Correct format (HTML, JSON, etc.); headings and lists used appropriately.
Brand Standards: Aspect ratio, watermark, spelling, contrast, link policy.
Accuracy: Facts and numbers are either stable or sourced; no hallucinations.
Actionability: Copy-paste ready; includes CTA and next steps where appropriate.
Clarity: Plain, confident style; no jargon without definition.
Scope: Stays within the requested step unless told otherwise.

The assistant silently checks the draft against these items. If it can’t hit ≥90%, it revises before you ever see it.

How This Reduces Errors You Actually Care About

Overlay typos on thumbnails: Prevented by rubric checks for spelling/contrast and by negative prompts banning warped text.
Wrong aspect ratio: Blocked by the Memory Prompt that hard-codes 16:9, 9:16, and 1:1 defaults.
Unwanted scope creep: Stopped by the Scope/Guardrails line in the role section.
Missing links/CTAs: Avoided by structural defaults (footer CTA blocks for YouTube + Spotify).
Out-of-date info: Mitigated by ReAct: browse when freshness matters, cite when appropriate.

Where the Time Savings Come From

Time savings compound in three places:

Prompt friction disappears. APE and role clarity handle the nitpicks you used to type every time.
Fewer revisions. The private self-check removes obvious mistakes before they become edits.
Reusable patterns. Structured formats mean every deliverable shares the same skeleton—so your audience learns to expect consistent quality.

Implementation: Set It, Forget It (Mostly)

Open ChatGPT > Settings > Custom Instructions.
Paste the full block above into the “How would you like ChatGPT to respond?” area. Keep your brand notes (aspect ratios, watermark, links) in that same block.
Test with three quick tasks: a thumbnail prompt, a 500-word blog stub, and a short social caption. Look for the “Techniques used” line at the end.
Tune the rubric. If you keep fixing the same issue (e.g., missing watermark), add it explicitly to the rubric section.
Document your scope boundaries for each project (e.g., “Don’t move to YouTube description until I say so”).

FAQ

Q: Will this make the assistant too rigid?
A: No—the techniques list (Zero-Shot/Few-Shot, Self-Consistency, APE) keeps creativity intact while protecting your non-negotiables (spelling, ratios, links).

Q: Does Self-Consistency slow things down?
A: A few extra milliseconds now saves minutes of rewrites later. Use it for high-stakes or highly creative tasks; skip it for quick asks.

Q: What if I need new defaults later?
A: Edit the custom instructions once. Everything downstream benefits immediately.

Q: How do I keep scope discipline?
A: Include a single line in the role: “Stay within this project step.” It’s simple and surprisingly effective.

A Reusable Footer CTA (Drop Into Any Post)

To keep cross-platform growth automatic, attach this CTA to your templates:

Keep Exploring with Deep Dive AI

📺 Watch on YouTube → Subscribe to Deep Dive AI
🎧 Listen on Spotify → Deep Dive AI Podcast

Thumbnail rule of thumb: 16:9, clean type, and a small “Deep Dive AI” watermark in the bottom-right.

A Mini-Template Library (Copy/Paste)

Thumbnail Prompt Skeleton (16:9)

[Subject in 3–6 words], cinematic wide shot, clean high-contrast composition; text overlay (exact): “[TITLE]”; 10% safe margins; legible, no distortion; small “Deep Dive AI” watermark bottom-right; 
Palette: [two colors]; 
Negative: warped text, clutter, busy background, low contrast, typos; 
Aspect ratio: 16:9.

Editorial Cartoon Prompt Skeleton (9:16)

Hand-inked linework with crosshatching, satirical but warm; recurring Russian Blue cat cameo; crisp lettering, no smudge; caption at bottom within 10% safe margin: “[CAPTION]”; 
Negative: cramped text, low contrast, odd cropping; 
Aspect ratio: 9:16.

Blog Section Skeleton (Blogger-Ready)

[Section Title]
[Lead paragraph with the core claim in one sentence.]

  Point 1: [Short explanation.]
  Point 2: [Short explanation.]
  Point 3: [Short explanation.]

Governance: Keep Your Pipeline Clean

Great systems die from scope creep. This line in your custom instructions is the guardrail that keeps everything sane:

“Each project stays strictly within its defined step. Do not advance to later steps unless explicitly instructed.”

It sounds simple, but it prevents the most common failure mode: the assistant starts writing the YouTube description while you’re still designing the thumbnail.

Final Checklist (Paste Into Your Notes)

✅ Role and mission clearly defined
✅ Techniques menu present (Zero/Few-Shot, CoT, Self-Consistenc