Welcome back to your weekly corner of creativity, clarity, and AI. This week is a GPT‑5 special. I pulled the best prompt patterns I could find across the web, blended them with my own tests, and packaged them so you can put the new model to work immediately.
And please ignore all the launch hiccups and hype. You should test the latest GPT-5 and make up your own mind. Think of this Substack update as your personal prompt lab. If it’s here, it’s been tried and tuned. Grab a coffee (or a tea or something more substantial if you prefer) and let’s take this latest OpenAI creation for a spin.
BTW, these tips are ChatGPT-5 specific, but you can run them in any LLM—best, run in all of the models you can access and compare the results!
A rapid GPT‑5 overview (what changed and why it matters)
In a few short years, we’ve gone from “clever high‑school student” vibes to something that feels more like a PhD‑level collaborator. GPT‑5 pairs stronger reasoning with better task execution. Two practical implications:
It’s easier to nudge how hard it thinks (“think hard about this” vs. “fast take only”).
It’s better at doing, not just drafting: planning, researching, using tools, and delivering finished outputs.
You’ll also notice helpful upgrades for day‑to‑day work: more grounded answers, stronger code, longer context handling, optional memory for continuity, tone presets to shift style quickly, and agent‑style workflows that can plan, act, and report.
Yes, you may find GPT-5 less sycophantic than before… You be the judge if that’s a bad thing.
Four GPT‑5 prompt patterns to use this week
Each pattern includes a drop‑in block and a short “why it works.”
1) Route Selector: “Fast vs. Deep”
Use this when you want either a quick take or a slow, careful answer.
Fast take
You are an analyst. Give me the fastest useful answer.
Constraints:
- Be decisive and brief (5 bullets max).
- If info is uncertain, state one caveat once.
- Do NOT ask me questions unless essential.
Task: Summarize 3 takeaways from [TOPIC].
Deep take
You are an expert researcher. Think hard about this.
Constraints:
- Apply deep reasoning internally.
- Show only the final answer plus a 3‑bullet sanity check.
- Cite sources if used.
Task: Produce a rigorous, stepwise plan to achieve [GOAL] in [CONTEXT].
Why it works: GPT‑5 is better at routing between “smart & fast” and “careful & thorough.” Making that expectation explicit improves results.
2) Agentic Doer: “Plan, act, report”
Use this when you want an end‑to‑end result, not just notes.
Goal: Create a 1‑page competitive brief on [PRODUCT/COMPANY] for [AUDIENCE].
Workflow:
1) Plan: Write a 3‑step plan before doing anything.
2) Research: Gather 6–8 high‑quality sources (avoid duplicates).
3) Synthesize: Build a concise brief with sections: Snapshot, Differentiators, Pricing, Moat, Risks, Open Questions.
4) Deliver: One‑page brief (Markdown) + short sources list (title + link).
Status discipline:
- Post one short status line before each tool call.
- End with “What I did / What’s next”.
Guardrails:
- Respect robots.txt and terms of service.
- Skip paywalled or private data.
Stop when the brief meets the goal.
Why it works: GPT‑5 handles multi‑step tasks more reliably if you require a plan and status updates. It keeps long runs legible.
3) Verbosity Dial: “Short prose, long code”
Use this when you want concise explanations but richly commented code (or the reverse).
Global style: Low verbosity for prose. High verbosity for code comments only.
Task: Implement [FEATURE] in [STACK].
Requirements:
- Output a single diff or file set.
- Use thorough inline comments and docstrings.
- Keep non-code explanation to 5 bullets max.
Why it works: GPT‑5 respects natural‑language “verbosity” controls if you separate what should be brief vs. expansive.
4) Structure‑Locked Outputs: “No more messy JSON”
Use this when downstream systems need strict formatting.
Output EXACTLY this JSON shape:
{
"idea": "string",
"assumptions": ["string", "string"],
"first_step": "string",
"risk": "string"
}
Task: Generate 1 validated idea for [GOAL] with realistic risks.
If a field is uncertain, use "TBD".
Why it works: Tight schemas plus explicit failure handling (“TBD”) drastically reduce cleanup.
Best‑of‑web prompt ideas (curated and adapted)
A) Agentic eagerness controls
When the model is too proactive, cap its search budget; when timid, encourage persistence.
<context_gathering>
- Search depth: very low
- Max tool calls: 2
- Proceed under uncertainty if ≥70% confidence
</context_gathering>
<persistence>
- Do not hand back until the task is fully resolved.
- If uncertain, make the most reasonable assumption and continue.
</persistence>
Task: [YOUR TASK]
B) Tool preambles
Ask for a plan up front and status lines before/after tool use.
<tool_preambles>
- Restate the goal.
- Share a step-by-step plan before any tool use.
- Post a one-line status before each tool call.
- End with “What I did / What’s next”.
</tool_preambles>
Task: [YOUR TASK]
C) Minimal‑reasoning speed mode
For extraction/classification at volume, skip reflection.
Speed mode: Use minimal reasoning. One-word label only.
Task: Classify this review as positive | neutral | negative:
"[TEXT]"
D) One‑liner to prototype
Use a single, vivid instruction to scaffold a working demo, then iterate.
Make a procedural brutalist building creator where I can drag and edit buildings.
They should look like actual buildings. Think hard about this.
Tip: Follow up with “make it better” plus specifics.
E) Memory‑savvy setup
Tell ChatGPT what to remember about your preferences, then apply to a task.
Please remember my preferences:
- Writing voice: clear, no fluff
- Deliverables: one-paragraph executive summary, then bullets
Acknowledge, then apply on this task:
[TASK]
(You control memory in Settings.)
F) Personality stress‑test
Compare tones quickly using built‑in tone presets.
Topic: Explain [COMPLEX CONCEPT] for smart non‑experts.
Generate 4 versions styled as: Default, Cynic, Robot, Listener.
Each: ~120 words, same structure, different tone.
How I’d use GPT‑5 today (by role)
Startup founders
Validate ideas with rapid “fast vs. deep” paired prompts.
Generate competitor briefs and micro‑landing pages end‑to‑end.
Operators and ICs
Use minimal‑reasoning prompts for classification/extraction.
Set memory‑based templates so recurring docs come pre‑formatted.
Public‑sector and policy leaders
Agentic synthesis with transparent plans + sources for briefing notes.
Enforce structured outputs for downstream systems.
Developers
Split verbosity: concise status, richly commented code.
Use strict schemas to eliminate parsing errors; load longer context for repo‑wide reasoning.
Weekly Wisdom
GPT‑5 is both a thinker and a doer. Your job: tell it how hard to think and how far to go. Nudge the router, set stop conditions, and use preambles to stay in control.
Bonus Prompt Templates for Paying Subscribers
Keep reading with a 7-day free trial
Subscribe to AI-Driven Success to keep reading this post and get 7 days of free access to the full post archives.