Guide for prompting GPT-5 with tools and optimizer

Prompting ChatGPT-5 Specifically

6 Moves That Consistently Work

written by

: an image by Ted Tschopp

Why this works: GPT-5 is a router model that can think, browse, code, write, and orchestrate tools. Results improve dramatically when you set the dials (reasoning + verbosity), name the tools, and force a quality loop. OpenAI’s own guides and product notes back this approach. (OpenAI Cookbook, OpenAI Platform, OpenAI)

1) Set the Reasoning Dial (every time)

Use explicit levels so the model doesn’t “guess.”

Snippet (drop in any task)

Reasoning: <think | think harder | ultra think>.
Before answering, produce an internal plan, resolve ambiguities, surface trade-offs, and justify choices.

Why: OpenAI’s GPT-5 prompting guide calls out insufficient reasoning and recommends prompt optimization to correct it; explicitly requesting deeper reasoning improves outcomes. (OpenAI Cookbook)

2) Set the Verbosity Dial (every time)

Pin the length and structure so you get the depth you want.

Snippet

Verbosity: <low | medium | high>.
Format: <bullets | numbered steps | sections + tables>.

Why: OpenAI’s prompt-engineering guides encourage clear format/length instructions; GPT-5’s optimizer also standardizes this in the revised prompt. (OpenAI Help Center, OpenAI Cookbook)

3) Name the Tools and Deliverables (multi-tool prompts)

Tell ChatGPT which skills to invoke and what artifacts to return.

Template

Tools to use in one pass:
1) Web browsing: gather 8–12 recent, reputable sources; note dates.
2) Data/Code: run Python to analyze and produce a CSV.
3) Document export: return a 2–4 page PDF.
4) Image generation: produce a simple logo (PNG + SVG).

Return: links to files/embeds for each output.
Reasoning: think harder. Verbosity: medium.

Why: ChatGPT can coordinate multiple tools/skills in a single run (web, code, files, images, etc.), a capability surfaced in OpenAI’s docs on tools and the ChatGPT Agent. (OpenAI Platform, OpenAI)

4) Force Self-Reflection (private rubric)

Make the model quietly grade and fix its own work before you see it.

Drop-in block

Self-reflection (private): Create a 5–7 category rubric for this task
(e.g., Goal coverage, Factual accuracy with citations, Completeness, Technical quality, Readability/UX, Testability, Constraints met). 
Iterate until the draft would score ≥90% overall, with no category <80%.
Do not show the rubric or scores—only the improved result.

Why: OpenAI’s GPT-5 prompting materials and optimizer emphasize evaluation loops; adding an internal rubric turns “one-shot” into “plan-draft-review-revise.” (OpenAI Cookbook)

5) Use Metaprompting when results miss

Fix the prompt, not just the output.

Repair prompt

Metaprompt:
Desired behavior: <what you wanted>.
Undesired behavior: <what you got that you don't like>.
Question: Provide minimal edits/additions to my original prompt so it consistently meets the desired behavior while preserving most wording. Return the revised prompt only.

Why: “Meta-prompting” (have the model improve your prompt) is a documented best practice and pairs well with OpenAI’s prompt optimizer. (OpenAI Platform, OpenAI Cookbook)

6) Finish with the Prompt Optimizer (free, official)

Paste your final prompt into OpenAI’s optimizer; use the tightened version going forward. (OpenAI Cookbook)

Copy-Paste Blueprints (fully specified)

A) Decision Brief (research + data + PDF + citations)

Goal: Produce a decision brief on <topic>.
Audience: execs.
Reasoning: ultra think. Verbosity: high.

Tools & outputs (one pass):
- Web browsing: find 8–12 diverse, reputable sources from the last 18 months; capture titles, publishers, dates, and URLs.
- Data/Code: analyze any quantitative figures and produce a CSV (sources × key metrics).
- Doc export: compile a 3–5 page PDF brief.

Structure:
1) Executive Summary (8 bullets max)
2) Findings with inline citations [#] and a reference list with dates
3) Options & Trade-offs (table)
4) Recommendation & Assumptions
5) Risks & Mitigations
6) Next 30/60/90 days

Self-reflection (private): coverage, credibility/recency, synthesis clarity, decision usefulness, risk balance, measurability. Iterate to ≥90%.
Deliver: Markdown brief + CSV + PDF + bibliography with live links.

Cite recency and credibility per OpenAI’s browsing/capabilities guidance. (OpenAI Platform, OpenAI Help Center)

B) Production-Grade Feature (design → code → tests → verification)

Goal: Implement <feature> in <tech stack>.
Reasoning: think harder. Verbosity: medium.

Outputs:
1) Design plan (data structures, API surfaces, error cases, perf/security notes).
2) Code (idiomatic, commented).
3) Tests (unit + one integration path).
4) Verification: explain how tests prove correctness; include documenting a runbook.

Self-reflection (private): correctness, robustness, readability, security, test coverage, DX. Revise until all categories ≥80%.

Backed by OpenAI tool docs (code execution/tools) and prompt-engineering guidance. (OpenAI Platform, OpenAI Help Center)

C) Brand Kit (logo + PDF + posts + competitor table)

Goal: “<Name>” mini brand kit.
Reasoning: think harder. Verbosity: medium.

Tools:
- Image generation: primary logo + mono variant (PNG, SVG).
- PDF export: 3-page brand guide (palette hex, typography, spacing, misuse).
- Web browsing: 8 peer brands with URLs and 1-sentence positioning.
- Copywriting: an announcement post + 3 alternates.

Self-reflection (private): distinctiveness, legibility, consistency, practicality, copy clarity, differentiation. Iterate, then deliver links to assets.

Leverages images, docs, browsing, and copy in one run; see tools and agent references. (OpenAI Platform, OpenAI)

Tools ChatGPT Can Call (and what they’re for)

Availability varies by plan and region. Names shift as features evolve, but these are the major, documented tools/skills in ChatGPT as of Aug 30, 2025, with official references.

  1. Web Browsing / Agentic Web Actions Search, open pages, extract facts, and (with Agent) complete multi-step tasks on the web with user approval. (OpenAI Platform, OpenAI)

  2. Code Execution / Data Analysis (Python sandbox) Run code, analyze datasets, generate charts/tables, and create files (CSV, PPTX, etc.). (In product UIs this is often presented as “data analysis”/“code interpreter.”) (OpenAI Platform)

  3. File Handling & Projects/Canvas Upload files, co-edit in Canvas (side-by-side doc/code space), and organize work in Projects. Canvas supports inline edits, comments, and code fixes. (OpenAI Help Center, OpenAI)

  4. Image Generation & Editing Create or edit images from prompts (logos, diagrams, mockups). (In-app image tools draw on the platform’s image capabilities.) (OpenAI Platform)

  5. Document Creation/Export Compose and export docs (e.g., PDF) directly from chat/canvas workflows. (OpenAI docs show tool calls that generate artifacts; Agent can also orchestrate outputs.) (OpenAI Platform, OpenAI)

  6. Prompt Optimizer (official) Paste any prompt to get an optimized version tailored to GPT-5 behavior; recommended by the GPT-5 prompting guide. (OpenAI Cookbook)

  7. Automations / Tasks (scheduled & conditional) Create reminders or recurring checks (“tasks”) that notify you later (e.g., “search X weekly”). Manageable from Settings → Notifications → Manage tasks. (OpenAI Help Center)

  8. Voice Mode Speak naturally; get spoken responses (mobile/desktop). Useful for hands-free prompting. (OpenAI Help Center)

  9. Connectors & Account Integrations (Agent-driven) With your approval, ChatGPT Agent can use connected services (e.g., Gmail, Google Drive/Calendar, GitHub) to retrieve or act on information during tasks, pausing for consent. (Tom’s Guide, The Times of India)

Why I didn’t include third-party “tool catalogs”: they’re not authoritative for ChatGPT’s native capabilities. The list above maps to OpenAI’s own help docs, platform docs, and product announcements. (OpenAI Platform, OpenAI Help Center, OpenAI)

A One-Page “Do It Right” Prompt (fully specified)

Use this when you want reliable, citable research + a polished brief in one go.

Goal: Create a decision brief on <topic>.
Audience: senior leadership.
Reasoning: ultra think. Verbosity: high.

Tools & outputs:
- Web browsing: find 8–12 reputable sources published in the last 18 months; extract title, publisher, date, URL.
- Data/Code: compute any comparable metrics; output a CSV of sources × metrics.
- Document export: produce a 3–5 page PDF brief + Markdown version.

Structure:
1) Executive Summary (≤8 bullets)
2) Key Findings (with inline bracketed citations and a numbered References section with dates/links)
3) Options & Trade-offs (comparison table with 0–5 scores + rationale)
4) Recommendation (with assumptions)
5) Risks & Mitigations
6) 30/60/90-day plan

Constraints:
- Show publish dates after facts likely to change.
- Prefer primary/official sources; avoid low-credibility blogs.
- List any material uncertainties.

Self-reflection (private): coverage, credibility/recency, synthesis clarity, decision usefulness, risk balance, measurability. Iterate until ≥90% overall; no category <80%.
Deliver: links to PDF + CSV + Markdown + full reference list with live URLs.

Justification: aligns directly with OpenAI’s tools/browsing guidance and prompt-optimization practices. (OpenAI Platform, OpenAI Cookbook)

Rubric (to judge this guide and your own)

Score each 1–5; aim ≥31/35.

  1. Task Clarity & Goal Framing – Clear goals, when/why, before/after.
  2. Actionability & Structure – Copy-paste blocks, checklists, decision tables.
  3. Router Controls Coverage – Reasoning, Verbosity, Tools, Self-Reflection, Metaprompting, Optimizer.
  4. Quality Loop – Private rubric, iterate-until-good thresholds.
  5. Examples & Transferability – Research, code, branding, enterprise.
  6. Precision & Safety – Dates, citations, constraints, credible sources.
  7. Brevity-by-Design – Dense, skimmable sections; purposeful length.

Sources (official & reputable)

  • GPT-5 Prompting Guide – how to diagnose/optimize prompts; references to the optimizer. (OpenAI Cookbook)
  • Optimize Prompts (Cookbook) – official optimizer walkthrough and patterns. (OpenAI Cookbook)
  • Platform “Tools” Docs – how ChatGPT/Assistants call tools (code, browsing, files, functions). (OpenAI Platform)
  • ChatGPT Capabilities Overview – Voice Mode and Canvas overview. (OpenAI Help Center)
  • Introducing Canvas – official product intro and collaboration workflow. (OpenAI)
  • Tasks in ChatGPT – creating and managing scheduled tasks/automations. (OpenAI Help Center)
  • Introducing ChatGPT Agent – multi-step, agentic actions with approvals and connectors. (OpenAI)
  • News coverage on Agent capabilities – independent summaries of what the Agent can do (web actions, connectors). (Tom’s Guide, The Times of India)
  • Prompt-engineering best practices (Help Center) – explicit instructions, formats, and constraints improve outcomes. (OpenAI Help Center)