Prompting Foundations
How to talk to chatbots
CAVEAT
- The world is going crazy and tech choices matter. This talk is apolitical and does not represent my values or the values of anyone connected to me.
chatGPT, Claude, Gemini,
(also "the little box on ...")
What it's good at
- first drafts
- summarising
- brainstorming
- translation
- formatting
What it's bad at
- A decreasing number of things...
- Looking at EVERY ITEM on a list
- Providing guaranteed facts
- Reminding you that it has no idea what it's talking about because it sounds so smartie
Quick survey
- Who uses foundation chatbots?
- ChatGPT / Claude / Gemini / Perplexity?
- Do you have a policy?
- Is usage consistent with it?
Point 0: Safety
Follow policy (you have one, RIGHT ಠ_ಠ)
Create a free account
TURN OFF "helping train the model"
No personal data / case specifics
Point 1: Start new chats
When to start a new chat?
- Topic changes (your brain goes on a tangent)
- It drifts (it goes on a tangent)
- You want clean context
- You've been on one chat too long
- Fear the Compression!
Point 2: The "Looks Fine" Problem
We proudly support people that suffer from Asperger's —not only this, it’s an inspiring reminder of how much they overcome every single day.
Em dash
- Em Dash Basics (—) is that long, dramatic dash for breaks—like this—replacing commas or parentheses
- Normal Dash (-) or Hyphen; is the short one for compound words (e.g., well-known)
- en dash (–) is mid-length for ranges (2025–2026).
American grammar/spelling
- "people that"/ "people who"
- Most docs digested by LLMs are American
- uwotm8 - https://i-dot-ai.github.io/uwotm8/
"Unnaturally perfect"
- "feels too clean"
- Is my neurodivergent voice extinct?
Style guide
- do disabled people "suffer"?
Issues to deal with
- We want it to AVOID doing stuff you wouldn't do:
- em dash
- American/British grammar/spelling
- Sounding "too good"
- Violate your style guide
Fix 1: Style Control
Modify personal prompt (Settings -> Personalization -> Custom Instructions)
- “No em-dashes”
- “Use British English (emulate uwotm8 usage)”
Fix 2: Your Comms
- Define your population
- Clarify sensitive terms
- should we have a policy?
Point 3: Add Context!
In the chat window:
"Do the thing":
xxx/;;;
Content
xxx/;;;
... Or Via Custom Instructions:
- Settings -> Personalization -> Custom Instructions)
- Positive Controls!
... Or via formal context store
- To consider:
- one-time or dynamic?
- attack surface / other trust issues?
... Or flip the script and bring the "chatbot" into your knowledge base
- Microsoft copilot
- Claude Code
- Github Copilot
- Obsidian knowledge base with an AI front end containing whitelisted information semantically linked
CONTEXT IS VERY IMPORTANT
- Strategy docs?
- Current initiatives?
- Who you are and how you work
Point 4: Prompting Frameworks
- "Hey Claude, fix my report"
- ??? Um, I guess here you go?
- "😡Aaarg... AI is so stupid!"
METHOD
M – Motive - What’s the purpose of your prompt? What are you trying to achieve?
E – Expectations - What do you expect from the AI? Should it be detailed, concise, creative, technical, etc.?
T – Tone - What style or voice should the response have? Formal, friendly, persuasive, humorous?
H – How - How should the AI present the response? Lists, paragraphs, step-by-step guides, tables?
O – Output - What’s the ideal format? A summary, an email draft, a script, a blog post?
D – Details - Are there any specific requirements or constraints? Word count, key points to include, things to avoid?
CRISPMAD
C – Context- Provide background information to set the scene. This helps the AI understand the situation and generate relevant responses.
R – Role - Specify the role you want the AI to take (e.g., "Act as a marketing consultant" or "Be a Python tutor"). This helps tailor the response style.
I – Instruction - Clearly state what you want the AI to do. Be specific to avoid vague or generic answers.
S – Scope - Define the limits of the response, such as word count, format, or depth of explanation.
P – Purpose - Explain why you need this response—what outcome are you aiming for? This helps refine the focus of the reply.
Point 5: Iteration
- First draft ≠ final
- Small nudges improve it
- A few tries is normal
- Close the loop with "retros"?
Wrapping up
... hrm... I need an ending...
I'm doing a talk for an AI event. My slides are in obsidian. I need an ending for the following talk. How bout a summary?:
Thank you :)
