FROG in a Bowl: The Prompt Method You’ll Actually Remember

If you only remember one prompting checklist in 2026, make it FROG in a Bowl.

FROG is Format · Role · Objective · Goal. The bowl is Context — the water the frog sits in. Without the bowl, you have a cute acronym and a model guessing your background. With it, you have the same structured-prompting method every major lab documents — just named so you can recall it when you’re tired, shipping, or teaching someone else.

explainx.ai did not invent the ingredients. OpenAI, Anthropic, and Google all push some mix of role/persona, task, format, and context. Google Workspace popularized PTCF (Persona · Task · Context · Format). Anthropic’s Claude guides push a 4-block layout (instructions · context · task · output format). Classic Twitter threads recycle RTF (Role · Task · Format). We folded those into one sticky phrase: put the FROG in a bowl.

TL;DR — what people are asking

The five parts, clearly

F — Format

Tell the model the shape of the answer before it starts typing.

Examples: Markdown memo · JSON schema · table with named columns · three bullets under 20 words · PR diff summary · Lean-style proof sketch · email with subject line.

Missing Format is why you get a chatty essay when you needed a checklist. Our structured output guide covers JSON-mode and schema cases.

R — Role

Tell the model who it is for this turn: senior SRE reviewing an incident, staff product manager writing a one-pager, skeptical editor cutting fluff.

Role sets vocabulary and what the model treats as salient. It is not a free accuracy upgrade. Wharton’s Prompting Science Report 4 found expert personas do not reliably improve GPQA / MMLU-Pro scores — a caveat we unpack in Mollick: specs, not tricks. Keep Role for framing; put accuracy work into Goal, Context, and evaluation.

O — Objective

The concrete task — one clear verb.

• Summarize these three PRDs into risks.
• Rewrite this landing section for a technical buyer.
• Find the failing assumption in this agent loop design.

If you stack five verbs in one prompt, you don’t have an Objective — you have a backlog. Split turns or use an agent harness (loop engineering).

G — Goal

What success looks like — the acceptance bar.

• Reader can decide go / no-go in 60 seconds.
• Output must be pasteable into Notion without cleanup.
• Prefer false negatives over false positives on security findings.
• Match the tone of the two examples below.

Goal is where Mollick’s advice lives: specify desired output, what good and bad look like, and how you’ll test. Objective is the move; Goal is the scoreboard.

Bowl — Context

Everything the model cannot infer: audience, product constraints, prior decisions, pasted docs, “don’t do X,” tool limits, brand rules, the failing log.

Anthropic now frames much of this as context engineering — what the model receives matters as much as wording. Our deeper take: context engineering vs prompt engineering.

Mnemonic: the frog (FROG) is the instruction skeleton. The bowl is the world those instructions operate in. Empty bowl → hallucinated world.

Why labs keep rediscovering the same checklist

Different names, same bones:

Google’s prompt design strategies stress clear instructions, added context, and response format. Anthropic recommends distinct sections (background_information, tool guidance, output description) with XML or Markdown. OpenAI’s guidance has long pushed instructions at the front, delimiters around data, and explicit structure.

FROG in a Bowl is the memory aid for that consensus — not a proprietary algorithm.

Copy-paste template

# Role
You are a {specific expert} helping a {audience}.

# Goal
Success means {acceptance criteria}. Prefer {tradeoff} over {tradeoff}.
Bad output looks like {anti-pattern}.

# Objective
{One verb-led task}.

# Format
Return {exact shape}. Use {headings / schema / length}.

# Context (bowl)
- Audience: ...
- Constraints: ...
- Source material:
"""
{paste or link summary}
"""
- Out of scope: ...

Weak vs FROG

Weak:

Help me improve this landing page.

FROG in a Bowl:

# Role
You are a B2B SaaS conversion editor for technical buyers.

# Goal
A CTO skimming on mobile should understand the offer in 10 seconds
and know the primary CTA. No hype adjectives. Bad = generic AI voice.

# Objective
Rewrite the hero headline + one supporting sentence for explainx.ai’s
agent-skills directory.

# Format
- Headline: ≤12 words
- Subhead: ≤28 words
- 3 bullet alternatives for the headline only

# Context
Product: searchable directory of agent skills and MCP servers.
Audience: builders already using Claude Code / Cursor.
Current draft: "Unlock the future of AI agents today."
Brand: explainx.ai — precise, technical, no purple-hype tone.

Recommended order (and what to skip)

Not every chat needs all five. Use this ladder:

1. Objective alone — fine for trivia and tiny transforms.
2. + Format — when you need pasteable output.
3. + Goal — when quality is subjective or stakes are high.
4. + Role — when tone/domain framing matters.
5. + Context — whenever the model would otherwise invent your world.

For production systems, put stable Role / Format / Goal rules in the system prompt and keep turn-specific Objective + Context in the user message — that pattern also plays nicer with prompt caching.

How it maps to Claude’s 4-block pattern

If you already use our master Claude prompt guide:

FROG adds the missing Role and forces Goal as a first-class field instead of burying success criteria inside a vague instruction paragraph.

For Claude, wrap sections in XML:

<role>...</role>
<goal>...</goal>
<objective>...</objective>
<format>...</format>
<context>...</context>

Common failure modes

FROG vs tricks, tips, and “secret prompts”

FROG is compatible with Ethan Mollick’s line: management, not spellcraft. Specifying Goal and Format is writing a mini-spec. Threatening the model, tipping it $200, or pasting a 40-line persona novel is not.

When tasks get multi-step, graduate from one FROG prompt to:

• zero / few-shot / CoT for hard reasoning slices
• agent skills for reusable FROG packs
• context · prompt · loop · harness when the unit of work is a workflow, not a message

60-second practice drill

Take your last vague prompt and fill this aloud:

1. Format — “I need it as ___.”
2. Role — “Act as ___ for ___.”
3. Objective — “Your job is to ___.”
4. Goal — “I’ll accept it if ___.”
5. Context — “Here’s what you must know: ___.”

If you can’t fill Goal, you don’t know what you’re asking for yet — clarify the human problem before you re-prompt the model.

Bottom line

FROG in a Bowl = Format, Role, Objective, Goal, sitting in Context.

Labs teach those pieces under PTCF, RTF, 4-block, and system-instruction guides. explainx.ai gave the stack a name you can remember in a meeting. Use the mnemonic. Keep the honesty: Role is framing, Context is fuel, Goal is the scoreboard, Format is the shipping container, Objective is the single job.

Put the frog in the bowl. Then evaluate the output like a manager, not a magician.

Related on explainx.ai

• Master prompt engineering for Claude
• What is a system prompt?
• Context engineering vs prompt engineering
• Ethan Mollick / Wharton: specs, not tricks
• Zero-shot vs few-shot vs chain-of-thought
• Structured output & JSON mode
• Context · prompt · loop · harness stack
• Evaluating prompts — how to measure quality
• What are agent skills?

Primary references: Anthropic — Effective context engineering for AI agents · Google — Gemini prompt design strategies · Google Cloud system-instruction guidance (persona, format, goals, context) · OpenAI prompt-engineering structure guidance · Wharton Generative AI Labs Prompting Science reports (via Mollick, July 2026)

FROG in a Bowl is an explainx.ai teaching name for widely recommended prompt structure. Lab docs evolve; verify current official guidance for your model family. Follow @explainx_ai for updates.