Prompts to use with the SPARKS method

Prompt 1: Speak it out

Best for: The first move on any new AI-assisted task—getting your messy thinking on the page before AI tries to help.

Use with: Any LLM (ChatGPT, Claude, Gemini)

Variables to fill in:

• [describe your project or challenge in 2–3 sentences] — e.g., ​“I’m planning a launch campaign for a new product feature aimed at solo marketers.”
• [dump it all without editing] — Everything in your head: what you know, what you’re unsure about, what you’ve tried, what’s at stake. Resist the urge to clean it up.

What to expect: The AI should reflect your thinking back with sharper structure—naming the audience you implied, the constraints you mentioned in passing, the questions you didn’t realize you were asking. If it jumps straight to writing copy, stop and re-prompt.

Cadence note: Run this once at the start of every new project. Don’t reuse it across projects—the messy dump only works when it’s specific to the current challenge.

Prompt 2: Pivot roles

Best for: Forcing AI to interview you before it generates anything. Stops it from filling in blanks you should fill yourself.

Use with: Any LLM

Variables to fill in:

• [3–5] — Number of questions. Five is the upper bound for most projects; three is enough for shorter tasks.

What to expect: A numbered list of questions. Good ones probe audience specifics, success criteria, and constraints you haven’t articulated. Answer each one before the AI moves forward.

Where it goes wrong: If you find yourself adding “…but also go ahead and write a draft” at the end, delete it. That defeats the purpose. Hold the line on the interview.

Prompt 3: Ask for more

Best for: Critiquing an initial output. AI’s first response is almost always competent and complete-looking—and almost always mediocre.

Use with: Any LLM

Variables to fill in:

• [paste it] — The full text of the AI’s previous output, copied back into the prompt.

What to expect: A list of weaknesses, hidden assumptions, and skeptic objections. If the AI says ​“this is great, here are some minor tweaks,” push harder—it’s defaulting to flattery.

Prompt 4: Reframe

Best for: Breaking out of a stale answer. A true reframe changes the viewpoint, not just the phrasing.

Use with: Any LLM

Variables to fill in:

• Pick one reframe at a time. Running all four at once produces a watered-down composite. The whole point is to commit to a single new angle of attack.

What to expect: A version of your output that approaches the same problem from a different position. The skeptical-customer reframe surfaces objections you missed; the competitor-embarrassment reframe usually exposes generic, safe phrasing.

Prompt 5: Keep asking

Best for: Iteration with surgical feedback. ​“Make it better” produces nothing—the more specific the note, the better the revision.

Use with: Any LLM

Variables to fill in:

• [name the exact problem with the output] — e.g., ​“the second paragraph reads like a brochure,” not ​“it feels off.”
• [describe with specifics, including what they’ve already tried or believe] — e.g., ​“founders who already tried HubSpot and found it bloated.”
• [name it] — Any constraint you held back—budget, regulatory, internal politics.
• [the specific section] — Name the part you want rewritten. Don’t ask for a full rewrite if only one paragraph is broken.

What to expect: A targeted revision that addresses the named problem. If the output still feels generic, the feedback wasn’t specific enough—diagnose the next round before re-prompting.

Prompt 6: Stop and think

Best for: The final quality check before you accept any output. This is where good work becomes exceptional.

Use with: Any LLM

Variables to fill in:

• [describe what excellent looks like for this task] — One sentence. e.g., ​“An excellent version of this brief is one that a junior strategist could execute against without asking a single clarifying question.”

What to expect: A self-graded critique followed by a revised output. The critique itself is often more useful than the revision—it tells you whether the AI actually understood the bar.

System-level use: Cassady’s longer-term move was to set a custom instruction at the account level: ​“Don’t compliment me. Challenge me. Before you give me any answer, give me a reason to look at this from a different direction.” That bakes the spirit of the final step into every session.

Putting it together

SPARKS works as a sequence on a single output:

1. S (Speak it out): dump messy thinking
2. P (Pivot roles): let AI interview you
3. A (Ask for more): force critique on the first draft
4. R (Reframe): try a different angle
5. K (Keep asking): iterate with surgical notes
6. S (Stop and think): grade against an excellence standard

You won’t always need all six. A short task might be S → P → S. A campaign brief might cycle through A, R, and K several times. The order is the discipline—the count is the project.

Tips for better results

• Run the Hallway Test before you start. Imagine asking the question to the smartest intern in your company. If they’d push back, the prompt isn’t ready.
• Wear the ​“bullshit hat” before reading any output. AI defaults to flattery. The skeptic’s mindset is yours, not the AI’s—deliberately put it on before evaluating.
• Hold the line on the interview step. The single biggest failure mode is asking AI to interview you and then telling it to generate anyway. Answer the questions first.
• Be specific with reframes. ​“Reframe this” without naming the new angle gives you a paraphrase, not a real reframe.