Speaker: Martin Rojas (@martinrojas)
See the DevNexus live blog table of contents for more posts
General
- Prompting is the new code switching [it took me a minute to realize he meant the ENglish language one]
Components
- System Message – sets behavior and role
- Instruction – what to do
- Context – background data
- Examples – pattern demonstration
- Constraints – output limits
- Delimiters – section separation
Markdown
- Most common prompting language. Still text but gives structure
- Headings, bold, list, code
Prompt types
- Zero shot – direct instruction – simple/fast but inconsistent quality
- One shot – format setting – consistent format, but limited pattern learning
- Few shot – pattern learning – adapts to context, but token intesive
- Role based – behavioral framing – consistent voice, but might override other instructions
Techniques
- Clarity and specificity. Need to define assumptions
- Chain of thought -make the model think like an analyst
- Format constraints – specify what want for output
- Prompt compression – use less tokens to say equivalent thing. Drop filler words like please. Use lists instead of sentences. Use a little quality, but worth it if minimal effect on output. Engineering tradeoffs.
- Progressive enhancement – Naked prompt (vagye, add role, add specificity, add chanin of thought, add constraints, add validation
AI as Coach
- Ask AI to improve your prompt; both with why and to rpdouce and improved prompt
- Ask AI to compress to make shorter
More notes
- Build prompt library that works for you – uses Obsidian and in the AI tools themselves (aka skills)
- Measure for your use cases
Advanced Patterns
- Tree of Thought (ToT) – explore multiple analytical approaches simultaneously then evaluate which version reals the most insight. This is why AI goes off for an hour; it is doing this behind the scenes
- Self consistency – try different approaches and then majority vote for accuracy
- ReAct pattern – Iterative reason > Act > Observe loops for multi step investigations
My take
Good start by defining vocabulary/components and good example. I’m really glad he shared the slides. The contrast between the text and background made the examples hard to read so I pulled up the deck on my computer for reading those.