Advanced Prompting in Apps

The biggest upgrade isn't a clever one-off prompt — it's not re-typing the same context every time. Every major app stores standing instructions that apply to every conversation.

Where to find it in each app:

ChatGPT's Custom Instructions has two boxes: what the AI should know about you, and how it should respond. Fill them like a standing brief:

> About me: I'm a marketing manager at a B2B software company.
> Non-technical. I work mostly in plain English, not jargon.
>
> How to respond: Be concise and direct. Lead with the answer,
> then the reasoning. Use UK spelling. When you recommend
> something, give me the one best option, not five choices.

Claude's Projects go further: you set instructions and attach reference files (a style guide, last quarter's report) that every chat in the Project can see — telling the AI who you are once instead of pasting it into message one of every conversation forever.

Beginners treat each message as a fresh request. Power users treat a conversation as a workspace, refined over several turns. The AI remembers everything earlier in the same chat, so build on it instead of cramming everything into one giant prompt.

The refine-in-place pattern:

> Draft a 200-word announcement email for our new analytics
> dashboard, aimed at existing customers.

> Good start. Make the opening line punchier and cut the
> third paragraph entirely.

> Now give me two subject-line options: one curiosity-driven,
> one benefit-driven.

You never re-explained the product, audience, or goal — the AI carried that context forward. Each turn is a small, surgical instruction.

When a conversation drifts or gets cluttered, start a fresh chat rather than fighting the accumulated context. A clean slate often beats a tangled thread.

For anything involving reasoning, comparison, or multi-step logic, the quality jump comes from asking the AI to think step by step before it answers. A model that rushes to a conclusion makes more mistakes; asked to reason first, it's noticeably more reliable.

The lazy version — and what to do instead:

> Which of these three vendors should we pick?

> Compare these three vendors. First, list the criteria that
> matter for a 12-person team on a tight budget. Score each
> vendor against those criteria one by one with reasoning. Only
> after that, give your recommendation.

The trigger phrases are worth memorising: "think step by step", "reason it through before answering", "show your working", "first... then... finally...". You're not changing the model — only the order in which it does the work.

Use it for comparisons, trade-offs, decisions, maths, logic, multi-step planning, and diagnosing what went wrong. Skip it for quick facts, definitions, and simple rewrites.

Describing the tone you want works. Showing an example works far better. Few-shot prompting means giving the AI one or two finished examples to copy the pattern exactly. Compare "Write product descriptions in our brand voice" (vague) with this:

> Here are two product descriptions in our house style:
>
> EXAMPLE 1 — The Riverside Tote. Roomy enough for a laptop and
> a lunch, light enough to forget you're carrying it. Waxed
> canvas that only looks better with age.
>
> EXAMPLE 2 — The Field Notebook. Lies flat. Survives rain.
> Holds a year of ideas in 96 stitched pages.
>
> Now write a description for "The Commuter Bottle" in exactly
> this voice and length.

The model infers the rhythm, sentence length, and slightly dry tone — things almost impossible to describe in the abstract. Two examples are usually plenty; three locks it in tight. Few-shot wins for brand or personal voice (tone is felt, not described), structured output (one sample row defines the schema), and classification tasks (examples set the boundary cases).

The fastest way to a great prompt is often to ask the AI to write it for you. Meta-prompting uses the model to improve, expand, or critique your instructions before you run the real task.

Pattern A — improve your draft prompt:

> I want a prompt that gets a strong summary of a long research
> report. My rough attempt: "Summarise this report." Rewrite it
> to be far more effective, asking clarifying questions first.

Pattern B — make it interview you before doing the work:

> Help me plan a 3-day team offsite. Before you draft anything,
> ask me up to five questions that would most change your
> recommendation. Then wait for my answers.

Pattern C — have it critique its own output:

> Critique that summary as a demanding editor, then tell me what
> to add to my original prompt to get a better result next time.

Pattern B is quietly transformative. Instead of guessing what context the AI needs, you let it pull the context out of you — surfacing requirements you hadn't even articulated.

A prompt that sings in one app can fall flat in another. The techniques transfer, but each platform leans differently, so you adjust rather than blame the prompt.

A practical adaptation drill: run one strong prompt in two apps back to back.

> Read the pasted quarterly update below. Think step by step:
> identify the three most important changes, explain why each
> matters for a non-technical exec, then write a five-bullet
> summary an exec can act on. Keep it under 120 words.
> [paste the update]

Run it in Claude, then ChatGPT. You'll often find Claude leans long and structured while ChatGPT leans brisk and punchy. Neither is wrong — but knowing the lean lets you add one steering line ("be more concise" / "expand the reasoning") instead of rewriting from scratch. Keep your best prompts in a notes file with a line on how each app responds; that little table is tuned to your work.