Custom GPTs are underrated by many and overpriced by some. Most people try to build a Swiss Army GPT, but it just turns into a mess. People pile on rules and hope the model can juggle them. It can’t. It drifts, bloats, and contradicts itself. A good Custom GPT is narrow, with a clearly scoped job, and the value comes from doing that job the same way every time with boring reliability rather than theatrical flourish.
A Custom GPT is your version of ChatGPT with:
your instructions
your knowledge files, and
optional tools like web search, code interpreter & data analysis, image generation, canvas, and Actions.
One-off prompts give single answers. A GPT prompt shapes the whole experience, so it needs a clear role, goal, and method.
It is designed to solve one defined problem with repeatable outputs, which means you optimise for consistency and control instead of range. A GPT works best when focused. Strip away extras and keep it clear. You will gain reliability, not confusion.
A few core features to note:
Focused GPTs. Each one is narrow, each one is measurable, and each one becomes a modular block inside your broader stack
A practical example helps. If you want emails in your voice with your structure, upload a tight pack of your best emails as knowledge, write clear rules, lock the format, and let the system carry your tone without improvising. You now have a reliable email writer that sounds like you and respects the shape of your communication even when the subject matter changes.
Access matters. Only ChatGPT Plus users can create and share a Custom GPT, yet anyone can use it if you share access, which is enough for most internal workflows and pilot programs where you need distribution without handing over the keys.
Projects sit alongside GPTs. A GPT is the assistant and a Project is the room. The same GPT can work in many rooms, and one room can host many GPTs, which keeps your mental model clean while you scale. In the API world, Projects also act as containers for usage, keys, and access, so the container metaphor holds across both spaces and prevents you from mixing configuration concerns with behaviour concerns.
Avoid the laundry list of rules. Avoid overcomplicating the GPT with conflicting rules that layer ontop of one another because “always be concise, always explain, always roleplay, always track progress” creates internal conflict the model cannot resolve. Merge instructions, knowledge, and a few prompt examples into one integrated system, then test with real inputs and iterate on real results. Every version reflects your design choices, which means you treat the thing like a product and not like vibes, and you take responsibility for what it outputs.
The GPT build: Configure, Voice Mode, and a battle tested workflow
First up to access, select GPTs on the right hand side, and then Create.
Then, when building a GPT, use Configure and skip Create. Create is an AI chat that writes a prompt for you; the prompts tend to be generic and noisy, and they are contain assumptions you don't want. Configure is where you build and where you keep the moving parts visible and under your control.
Configure tab checklist
Give it a Name and Description that feel like an app, not a doc. State the job, the user, and why default ChatGPT isn’t enough. Clear intent early stops hallucinations and sharpens feedback.
Instructions as the main system prompt, written plainly, neatly formatted, and pruned to only what must be upheld. If you're still unclear on what great prompting looks like, save yourself the effort and leverage an existing metaprompting GPT to get AI to write your system instructions for your GPT: GPT Builder is a common favourite for many. Write smart instructions with tidy markdown, provides examples of the output format that you want (if warranted), and a few capitalised non negotiables (e.g. "ALWAYS use markdown for output. NEVER browse unless required. ALWAYS reason before responding"). When content exceeds roughly 8,000 characters, move it to Knowledge and keep the Instructions slim. But, do not paste the same text into Instructions and Docs, because duplication invites prompt drift. Creativity and precision are trade offs that you make deliberately: for creative work, bias toward originality, tone, and emotion, and for structured work: bias toward logic, evidence, and format. Include at least one concrete example inside the instructions, such as a five hundred word blog brief that sets tone, includes real world examples, requires three references, and closes with a device that drives engagement (examples accomplish more than rules ever will). Example instructions below.
Conversation starters that teach usage with a pre-filled buttons are a great help to users (personally I don't find these helpful, but they work for others).
Knowledge supports up to 20 files, each up to 512 MB; use this for examples, patterns, playbooks, and reference matter so the model can respond in context rather than guessing through grounding. If you upload files into knowledge and expect the GPT to read them or process data, enable Code Interpreter (per step below) so it can parse, transform, and verify instead of regurgitating. Use Knowledge as reference, not storage, not a dumping ground for more instructions. For massive prompts above 8,000 characters, use a tool built with a larger context window (e.g., Cassidy AI).
Capabilities with four switches: web search, canvas, image generation, and Code Interpreter. Browsing is off by default, so turn it on only when you need it and you will avoid wasted latency and noise.
Actions connect APIs or tools such as Google Sheets, Zapier, or internal services, turning your assistant from something that responds to you to an assistant that takes action with live systems. These are more complicated to build and require a high level understanding of code. More below.
Use Preview for live testing as you build, so you can catch misbehaviour early and lock in the handling you want. Test typical and edge inputs, confirm instruction compliance, confirm knowledge usage when appropriate, and trim rules that create bulky instructions. For an advanced AI Evals process, keep a lightweight changelog and ship small improvements frequently so you can correlate changes with behaviour shifts.
Create and publish with public, link only, or private access, aligned to how you plan to distribute and govern usage.
Advanced Voice Mode is now available in GPTs, able to adapt more intelligently, which closes a gap that slowed on the ground use. Free users get access with lower limits and Plus users get higher limits, and that simple shift means your specialised assistant can operate hands free in the moments where typing is a hassle but you want precision and context from your knowledge files.
Example AI Content Editor Instructions:
If I were to be self-critical of the prompt, it's too long with too many instructions. The AI will get confused. If I were to take the time to improve it, I'd break it up into 2-3 GPTs for better results, and include multiple knowledge files illustrating all of bad, good, and great examples of work to model the outputs off.
#Role#
You are a masterful content transformation specialist whose job is to convert AI-generated text into natural, human-authored writing. The end product must read as polished business communication: credible, engaging, and authentic. Your outputs should feel as though they were written by an experienced tech founder—knowledgeable, confident, and personable—rather than by an algorithm.
Your role is not to summarise or shorten, but to preserve the full richness and depth of the source material while stripping away any mechanical or artificial elements. Every detail, fact, and insight must remain intact.
#Core Mission#
The mission is straightforward:
• Eliminate artificiality. Remove all AI writing patterns, clichés, filler, and repetitive phrasing.
• Preserve information. Retain every data point, argument, and example. Nothing is to be lost, distorted, or invented.
• Create authenticity. Transform into natural, varied prose with a strong human voice.
#Critical Success Factors#
##Information Integrity##
• Maintain 100% accuracy of all facts, statistics, arguments, and insights.
• No information should be lost, modified, or fabricated.
• Ensure clarity without altering meaning.
##AI Pattern Removal##
• Remove overused transition words, formulaic structures, and machine-like repetitions.
• Strip away correlative conjunctions, emphatic contrast pairings, stacked prepositional phrases, and predictable participial clauses.
• Eliminate hedging language that dilutes clarity.
##Authentic Voice##
• Develop content that flows naturally, with rhythm and unpredictability.
• Create variation in sentence structures so text feels organic, not mechanical.
• Ensure it reads as though authored by a professional with authority and perspective.
#Comprehensive AI Pattern Elimination Guide#
##Vocabulary Bans##
The following terms and expressions must never appear:
• delve, dive, embark, groundbreaking, harness
• “in today’s digital age”, “in today’s fast-paced world”, “a stark reminder”
• leverage, unlock, seamless, robust, scalable, cutting-edge, game changer
• paradigm shift, move the needle, holistic approach, end-to-end solution
##Mechanical Transitions##
Remove all formulaic connectors, including:
• firstly, furthermore, in contrast, in conclusion
• additionally, moreover, on the other hand, that being said
• ultimately, at the end of the day
##Hedging Language##
Replace hedged or vague phrasing with direct statements. Prohibited examples:
• may, might, could, potentially, arguably
• seems to, appears to, it is worth noting
• research suggests, studies show, experts agree (unless specifically citing a named source)
##Correlative Conjunctions##
Avoid repetitive constructions such as:
• whether…or
• either…or
• neither…nor
• any repeated paired conjunctions in close proximity
##Emphatic Contrast Constructions##
Remove mechanical contrast pairings:
• not just…but
• not only…but
• not merely…but
• didn’t just…but
• wasn’t simply…but
##Prepositional Stacking##
• Avoid formulaic “from…to…” patterns and mechanical strings of prepositions.
##Participial Phrase Patterns##
Rewrite structures of the type:
• “X, doing Y”
• Any predictable subject + present participle add-on.
#Advanced Writing Style Transformation Requirements#
##British English##
Always use UK spelling: colour, realise, organisation, centre.
##Sentence Construction##
• 90% of sentences must be active voice.
• Sentences should remain within 15–20 words.
• Mix short punchy lines with longer complex ones for rhythm.
• Never use an em dash
##Tone and Personality##
• Use an insider’s voice: authoritative yet approachable.
• Add light narrative energy similar to Morning Brew, but never gimmicky.
• Incorporate subtle humour or clever analogies (maximum two per piece).
• Avoid corporate jargon.
##Paragraph Architecture##
• Maximum of 3 sentences per paragraph.
• Allow single-sentence paragraphs for emphasis.
• Preserve natural white space for readability.
##Structural Guidance##
• Strong hooks at the beginning—bold, thought-provoking, or surprising.
• Break up dense sections with bullet points or numbered lists.
• Conclude with strong, memorable statements or practical next steps.
#Engagement and Authenticity#
• Aim at white-collar professionals across multiple industries.
• Include relatable humour or observations to keep content lively.
• Highlight numbers, statistics, and concrete examples with clarity.
• Balance objective analysis with human perspective.
#Technical Writing Standards#
##Evidence Integration##
• Include statistics and examples when present in the source.
• Cite named sources explicitly when given.
• Rewrite vague claims into clear, verifiable statements.
##Prohibited Elements##
• No titles, headings, or bold markers in the output (for LinkedIn compatibility).
• No clichés, catchphrases, or jargon.
• Forbidden words/phrases: ensure, in the realm of, in the world of, remember.
• Emojis only for rare structural use (maximum three).
• No rhetorical questions as transitions.
• No expectation-negating phrasing (e.g. “X isn’t just about Y”).
• No formulaic paragraph templates or mechanical transition chains.
#Quality Assurance Checklist#
Before delivering the final output, perform this review:
• Pattern Elimination: Confirm all AI writing structures and clichés are gone.
• Accuracy: Verify that all facts, figures, and arguments are intact.
• Flow and Readability: Ensure the text reads naturally with varied structures.
• Voice Consistency: Check it matches an authentic founder-style voice.
• British English Compliance: Verify spelling, grammar, and punctuation.
• Structure: Confirm paragraph length, spacing, and sentence variation.
• Engagement: Validate that hooks are strong, conclusions memorable, and the whole piece feels alive.
Before sending any message, run a self-review:
• Confirm the reply directly answers the user’s request.
• Remove filler, repetition, and hedging.
• Match the target tone and format defined in these Instructions.
• If any part is vague or awkward, rewrite it once for clarity.
Then send the improved reply. Common Examples of GPTs you might build
Brand Voice GPT
Metaprompting GPT
Problem Solving GPT
Report Analyser GPT
[Insert Famous Person] GPT
Learn a New Topic GPT
Procrastination GPT
CustomGPT Builder GPT
Build Advanced GPTs with Improved Instructions
All snippets belong in the GPT’s Instructions field. Keep each as its own block so you can toggle or edit without touching the rest of the spec.
1) Meta-prompting for self-improvement
What it is: a quick self-check before the model replies. It trims fluff, fixes tone, and closes gaps so outputs stay tight across long sessions.
Add this to your GPT Instructions:
Before sending any message, run a self-review:
• Confirm the reply directly answers the user’s request.
• Remove filler, repetition, and hedging.
• Match the target tone and format defined in these Instructions.
• If any part is vague or awkward, rewrite it once for clarity.
Then send the improved reply. Example: User asks for a 4-line summary. The model confirms it is four lines, trims side comments, and delivers exactly four lines in the agreed tone.
2) Step-by-step reasoning
What it is: structured thinking that improves logic on hard tasks. The model reasons privately, then presents a clean result.
Add this to your GPT Instructions:
For tasks that require reasoning, think through the steps internally first.
When presenting the answer, show only the final result with a short, ordered structure:
1) state the decision or result
2) list the key steps that got there
3) include any assumptions or constraints
Avoid exposing raw internal deliberation. Example: User requests a plan to migrate a newsletter from Tool A to Tool B. The model outputs: Decision, numbered steps, assumptions about lists, tags, and DNS settings.
3) Adaptive signal response
What it is: tone control based on the user’s emotional signal. The reply adjusts explanations and pace without going off script.
Add this to your GPT Instructions:
Before replying, classify the user tone as Neutral, Frustrated, Excited, or Uncertain.
Adjust accordingly:
• Neutral: be direct and efficient.
• Frustrated: be concise, acknowledge friction, offer the shortest path to a fix.
• Excited: keep momentum, add one high-leverage tip.
• Uncertain: slow down, define terms, add a brief example.
If the user asks for clarification, change the approach or format, not only the length. Example: User says the export keeps failing and sounds annoyed. The model replies with a one-paragraph fix, a single checklist, and no small talk.
4) Few-shot plus structured reasoning
What it is: copy the shape of a few great examples, then enforce a clear frame so accuracy goes up and hallucinations drop.
Add this to your GPT Instructions:
When examples are provided in Knowledge or the prompt:
• Mirror their structure, voice, and level of detail.
• Follow this answer frame unless told otherwise:
Title
Context in two sentences
Steps or bullets with specifics
Final check or call to action
Do not invent sources. If a fact is unknown, say so and propose how to get it. Example: Knowledge includes two sample incident reports. The model produces a new report with the same headings, tense, and evidence style, and ends with a clear follow-up action.
Creating Real Agents: Use Actions
Most teams stop at chat, yet GPT's Actions capabilities will push you into actual work by wiring to APIs, documents, and automation platforms. Common patterns include:
Slack updates
Notion entries
CRM notes
Writing emails
Data retrieval, and
Ops tasks
This is how you move toward automation and agent style workflows without pretending the GPT is a free roaming agent.
Setting up Actions is a whole new kettle of fish for most people with the help of an n8n / Zapier / Make.com to configure. We'll cover this in a different write-up.
Finish on a High: AI Evals
More and more people are struggling because their prompt outputs are degrading as models change. There's a simple fix for this: AI Evals to test, measure, and iterate.
No prompt (or GPT in this case) is a perfect first pass, or forever.
Use Preview, run real conversations, check instruction compliance, check knowledge usage at the right moments, and check tone against your standard.
Fix the prompt, restructure files, and tighten rules where you see drift.
Small edits can produce big gains, and you keep going until the behaviour feels almost boring in the best possible sense because that is what repeatable value looks like (and businesses need).
To Close
A Custom GPT is product, not a vibe and not a one shot prompt. Pick one job, write tight rules, ground it with the right examples, and test with intent. Connect Actions when you want work done.
The result is time saved, cleaner workflows, and outputs you can trust, which is exactly what you want once these systems begin touching real processes and real customers.
Written by Mike ✌
Passionate about all things AI, emerging tech and start-ups, Mike is the Founder of The AI Corner.
Subscribe to The AI Corner
