WiderGood

Free to anyone

Resources

Everything here uses an LLM, a large language model. That is the kind of AI behind ChatGPT, Claude, and Gemini: you type a request in plain English and it writes back. These tips are how to get useful work out of one, whichever you use. Not all of them are mine. Where a tip came from a paper, the paper is under Articles and papers below.

Prompt tips, in order

Eight of them. Click any one for the why and the exact wording.

01Make the model write the prompt. Describe the job and tell it to ask you questions first.

You know your problem. You do not need to know how to phrase it for a model, because the model is better at that than you are. Describe what you need in plain language, tell it to ask questions before writing, answer them, then paste the prompt it gives you into a fresh chat and run it.

Wording to use

Copy this prompt

I need a prompt I can paste into an LLM to do this job: [describe the job in plain language]. Context: the output is for [who reads it], in [format and rough length], and it will be used to [what happens next]. It would be wrong if [the failure you most want to avoid]. Before you write anything, ask me every question you need answered to write a better prompt, one at a time, and wait for my answers. Then give me the finished prompt in a single block I can copy, with [brackets] around anything I will need to fill in each time.

02Give it a role before the task. "You are a grant reviewer" beats a longer instruction.

A role narrows the range of plausible answers more reliably than adding sentences of instruction. Name the job the output is for, not just the task.

Wording to use

Copy this prompt

You are [the role: a grant reviewer at a community foundation / a bookkeeper who has seen a thousand invoices / a program officer reading her fortieth report this week]. You are careful, direct, and you say what you would actually think. Read the material below and give me [the task: the three claims you would question first, and why / the two numbers that do not reconcile / the paragraph you would cut]. Stay in the role. Do not soften it. If something is fine, say it is fine. [paste the material]

03Show one worked example for anything multi-step. Skip it for simple lookups.

For a report, a letter, or anything with structure, paste one finished example and say what made it good. The model copies the shape, not the words. For a quick question, this is wasted effort.

Wording to use

Copy this prompt

I am going to show you one finished example and then ask for a new one in the same shape. Here is the example: [paste it]. What made it work: [one or two sentences: the tone, the structure, the thing it led with]. Now write one for this: [the new details]. Match the structure, tone, and length of the example. Do not reuse its sentences. If the new details are missing something the example needed, ask me rather than inventing it.

04Ask what it assumed before it wrote anything. Most rework is an unstated assumption.

Bad output is usually a guess the model made about something you never said. Make it list the guesses first, correct them, then let it write.

Wording to use

Copy this prompt

Before you write anything: list every assumption you are making about this task, the person it is for, the format, the length, and what counts as done. Number them. For each one, tell me whether you are confident or guessing. Then stop and wait. I will confirm or correct the list, and only then should you produce the work. The task: [describe it]

05Ask for a second version, built differently, before you pick one.

The first draft is rarely the best one. A second attempt with a different structure costs ten seconds and gives you a real choice instead of a single take-it-or-leave-it.

Wording to use

Copy this prompt

Give me a second version of this, built differently: a different structure, a different opening, the same facts and the same length. Do not just reword the first one. Then, in two sentences, tell me which of the two you think is stronger for [the reader], and why. If neither is good, say that. [paste the first version]

06Ask three personas separately, a skeptic, an expert, a numbers person, then have it synthesize.

One balanced answer from a generalist is weaker than three one-sided answers stitched together afterward. Interview each persona on its own, then ask for the synthesis.

Wording to use

Copy this prompt

I want three separate answers to the question below, then a synthesis. First, answer as a skeptical board member who has seen consultants come and go and is protecting restricted funds. Second, answer as the program director who will have to run whatever we decide. Third, answer as the finance officer who has to defend the spend at audit. Keep each answer in its own section and do not let them agree with each other just to be agreeable. Then write one paragraph that names exactly where they disagree and what would settle it. The question: [paste it]

07Keep the prompts that worked in one shared document, with a line on what each was for.

A prompt that worked is an asset. This is the entire difference between a team that uses AI ad hoc and a team with a workflow, and almost nobody does it.

Wording to use

Copy this prompt

Entry format for the shared prompt document, one entry per prompt that worked: Name: [Grant report intake] Use it when: [you have raw notes for a reporting period and need a first draft] What it returns: [outcomes tied to numbers, one story, open questions] Owner: [name] Last checked: [date] What to watch for: [it will round numbers if you let it; tell it not to] Prompt: [paste the prompt exactly as it worked]

08Never paste donor, client, or beneficiary data into a consumer tool. Policy first.

Anything you type into a free consumer tool may be stored and may be used to train it. Work from a description with names and identifying details removed, and get a written policy before anyone on staff does otherwise.

Wording to use

Copy this prompt

I am going to describe a situation with every name, address, date of birth, and identifying detail removed, and I will refer to people as Person A, Person B, and so on. Work only from what I give you. Do not ask me for more detail about who anyone is. If anything I have written could still identify a real person, stop and tell me which part before you do anything else. The situation: [describe it]

+More short prompts that earn their keep.

Fifteen questions, one at a time

Ask me 15 progressively productive questions, one at a time, to help me with [this task].

Always ask for the alternative

List the best alternative approaches, compare them in a matrix, then ask me which one I'd like to proceed with.

Setting up a personal AI advisory board

The miscellaneous prompt list

This week

Four steps to your first workflow.

01Pick the task that took an hour last week and came out mediocre.

Not the hardest thing you do. The thing you do every week that takes too long and comes out fine. The monthly board update, a quote for a job, the narrative section of a grant report, the follow-up email after a missed call.

Wording to use

Copy this prompt

Fill this in before you open any tool. One line: Every [week / month] I [the task], it takes about [time], and the result is [what is wrong with it: generic, late, error-prone, never quite right]. If you cannot fill in the last blank, the task is fine as it is. Pick a different one.

02Describe it to the model and make it write the prompt. Answer its questions.

Do not write the prompt yourself. Paste the line from step one into the wording below, answer whatever it asks, and it will hand you a prompt that is better than the one you would have written.

Wording to use

Copy this prompt

I need a prompt I can paste into an LLM to do this job: [paste your one line from step one]. Context: the output is for [who reads it], in [format and rough length], from [what I will paste in: notes, a spreadsheet export, last month's version]. It would be wrong if [the failure you most want to avoid]. Before you write anything, ask me every question you need answered to write a better prompt, one at a time, and wait for my answers. Then give me the finished prompt in a single block I can copy, with [brackets] around anything I will need to fill in each time.

03Run what it gives you. If it saved the hour, do the next four tasks the same way.

Open a fresh chat, paste the prompt it wrote, paste your raw material, and read the result as the person it is for. Then make it check its own work before you do.

Wording to use

Copy this prompt

Open a fresh chat. Paste the prompt from step two, then paste your raw material where it asks. When it finishes, send this: Now read what you just wrote as [the person it is for]. List everything they would question, push back on, or find generic. Fix each one, and show me a short list of what you changed and why. Do not add anything the source material does not support.

04Write the prompt down. That is your first workflow.

Put it in a document the whole team can see, with what it is for and who owns it. The next person does not have to rediscover it, and it is the first row of the list a Fit Check would build.

Wording to use

Copy this prompt

Name: [what the prompt does, in four words] Use it when: [the situation] What it returns: [the shape of the output] Owner: [your name] Last checked: [today] What to watch for: [the one thing it got wrong on the first run] Prompt: [paste the prompt from step two, with the fill-in brackets left in]

Start here, 04

Getting started with Claude Code

The doc I hand people on day one, then the tools I build alongside it.

Reference

Articles and papers

The reading behind the tips above, plus a few pieces worth the hour on their own. Where there is a PDF link, it is a copy I keep on Google Drive.

  • How I Built a Personal Board of Directors With GenAI

    Vipin Gupta, MIT Sloan Management Review · Credit: Charles V.

  • The Bitter Lesson versus The Garbage Can

    Ethan Mollick, One Useful Thing · Credit: Phil R.

  • The Best Way to Use AI for Financial Planning, Starting a Hobby, and More

    The Wall Street Journal, PDF

  • A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT

    Jules White et al., Vanderbilt

  • Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Jason Wei et al., Google Research

  • ReAct: Synergizing Reasoning and Acting in Language Models

    Shunyu Yao et al., Princeton and Google Research

  • ChatGPT Prompt Patterns for Improving Code Quality, Refactoring, Requirements Elicitation, and Software Design

    Jules White et al., Vanderbilt

  • Large Language Models Are Human-Level Prompt Engineers

    Yongchao Zhou et al., University of Toronto and Vector Institute

  • Large Language Models as Optimizers

    Chengrun Yang et al., Google DeepMind

  • The Fearless Future: 2025 Global AI Jobs Barometer

    PwC

  • Building Living Software Systems with Generative and Agentic AI

    PDF

  • Benchmark Design Considerations

    PDF

  • My favorite Google Gemini prompts after 100 hours

    Digital Trends · Credit: Don

  • This prompt optimizer learns from its mistakes, like DNA

    Every · Credit: Phil R.

Reference, continued

Newsletters

Reference, continued

People to follow

  • Chase Hitchens

    Austin AI community; runs rooms worth being in

  • Damon Bodine

    Clear Road Labs; founder of Code for Community

  • Dario Amodei

    Anthropic CEO

  • Gergely Orosz

    The Pragmatic Engineer

  • Claire Vo

    How I AI

Reference, continued

Austin events

Where the people doing this in Austin actually show up.

  • Austin Forum on Technology & Society

    Monthly speaker series

  • Austin Technology Council

    Trade association and events

  • Austin AI Events

    Community events calendar

  • AITX events calendar

    On Luma

  • Gauntlet AI

    AI training community

  • Gauntlet AI Wednesday Night School

    Recorded sessions on YouTube

  • Code for Community

    Damon Bodine's open-source civic-tech initiative. I was at the first one.

  • The Larger Half

    On Luma

  • Claude Community

    On Luma

  • Claude 9-1pm

    On Luma

Reference, continued

Courses, not recommended

Kept for the record, struck through on purpose. I would not send anyone to a course now: the models will walk you through everything these teach, on the task you actually have.

  • Prompt Engineering for ChatGPT

    Coursera

  • Agentic AI and AI Agents: A Primer for Leaders

    Coursera

  • OpenAI GPTs: Creating Your Own Custom AI Assistants

    Coursera

  • Agentic AI and AI Agents for Leaders Specialization

    Coursera

  • AI Agents and Agentic AI in Python

    Coursera

Want the full annotated list?

Everything, with my notes

Send an email and I’ll send back the annotated version — why each item is on the list and who it’s actually for.