July 17, 2024
An instance of LLM prompting for programming

Xu Hao makes use of chain of thought and normal data prompting with ChatGPT when writing self-testing code

My account of an inside chat with Xu Hao, the place he reveals how he
drives ChatGPT to provide helpful self-tested code. His preliminary immediate primes
the LLM with an implementation technique (chain of thought prompting). His
immediate additionally asks for an implementation plan slightly than code (normal
data prompting). As soon as he has the plan he makes use of it to refine the
implementation and generate helpful sections of code.

Lately I watched an interesting name on Zoom. Xu Hao, Thoughtworks’s
Head of Know-how in China, defined his explorations in utilizing ChatGPT to
assist construct Self Testing Code. He
went by the form of interplay that works for him.

He begins with a immediate that units the context for the applying and the way
he needs the code to be structured.

The present system is a web based whiteboard system. Tech stack:
typescript, react, redux, konvajs and react-konva. And vitest, react
testing library for mannequin, view mannequin and associated hooks, cypress element
assessments for view.

All codes ought to be written within the tech stack talked about above.
Necessities ought to be applied as react parts within the MVVM
structure sample.

There are 2 sorts
of view mannequin within the system.

1. Shared view mannequin. View mannequin that represents states shared amongst
native and distant customers.

2. Native view mannequin. View mannequin that represents states solely relevant
to native person

Listed below are the frequent implementation technique:

1. Shared view mannequin is applied as Redux retailer slice. Examined in
vitest.

2. Native view mannequin is applied as React element props or states(by
useState hook), except for world native view mannequin, which can also be
applied as Redux retailer slice. Examined in vitest.

3. Hooks are used as the most important view helpers to retrieve information from shared
view mannequin. For many the case, it can use ‘createSelector’ and
‘useSelector’ for memorization. Examined in vitest and react testing
library.

4. Don’t dispatch motion immediately to alter the states of shared view
mannequin, use an encapsulated view mannequin interface as an alternative. Within the interface,
every redux motion is mapped to a technique. Examined in vitest.

5. View is encompass konva shapes, and applied as react element through
react-konva. Examined in cypress element assessments 

Listed below are sure patterns ought to be adopted when implement and take a look at
the element

1. When write take a look at, use `describe` as an alternative of `take a look at`

2. Knowledge-driven assessments are most well-liked.

3. When take a look at the view element, faux view mannequin through the view mannequin
interface

Consciousness Layer 

Requirement: 

Show different customers’ consciousness data(cursor, identify and on-line
info) on the whiteboard. 

AC1: Don’t show native person

AC2: When distant person adjustments cursor location, show the change in
animation. 

Present an general answer following the steerage talked about above.
Trace, hold all consciousness info in a Konva layer, and an consciousness
data element to render cursor, and identify. Don’t generate code. Describe
the answer, and breaking the answer down as a process record primarily based on the
steerage talked about above. And we'll refer this process record as our grasp
plan.

There’s lots happening with this immediate, so he highlighted a number of
factors.

He is utilizing a generic utility instance in right here: one factor to be
cautious of when interacting with ChatGPT and the like is that we should always by no means
put something that could be confidential into the immediate
, as that may be a
safety threat. Enterprise guidelines, any code from an actual mission – all these should
not enter the interplay with ChatGPT.

A lot of the immediate is setting out the design tips that he needs
ChatGPT’s generated code to comply with. He refined this by placing prompts into
ChatGPT, wanting on the consequence, and crafting the immediate to make it generate
within the fashion he wished. As soon as he has context working, he can paste it into
each session he has with ChatGPT, and share it with the remainder of the
staff.

Xu Hao begins his session by prompting the
ChatGPT LLM with context in addition to the instruction to generate a plan

Though the ultimate objective of this session is to generate code and its
assessments, at this stage he particularly stated Don’t generate code.
At this level he needs ChatGPT to stipulate its method earlier than we glance to any
code technology. It is a immediate method referred to as Generated Data.

A typical manner of interacting with an LLM is to provide it an issue and ask
it to generate some last product, whether or not than be prose or (on this case)
code. With Generated Data we cut up our interplay into separate steps:
first ask the LLM to generate some helpful details about the issue, and
then we feed that info again into the LLM to generate the ultimate
product.

Essentially the most fascinating a part of the immediate is the “implementation technique”
part which mixes two kinds of immediate:

  • Instruction prompting offers directions for the LLM to comply with
  • Chain of Thought prompting encourages the LLM to clarify its
    reasoning.

Within the Implementation Technique part, Xu Hao described the specified
structure sample as an anticipated “Chain of Thought directions” for
ChatGPT to comply with. Then he instructed ChatGPT to construct a process record (the
generated data) primarily based on this chain of thought.

Answer:

To implement the attention layer function, we are able to comply with the rules talked about in the issue assertion and use the present tech stack.

Process Checklist:

  1. Create a brand new element named “AwarenessInfo” to render cursor and identify for distant customers.

  2. Create a brand new Konva layer to carry all of the “AwarenessInfo” parts.

  3. Add the Konva layer to the primary whiteboard element.

  4. Create a brand new Redux slice to retailer the attention info for distant customers.

  5. Create an encapsulated view mannequin interface for the attention Redux slice.

  6. Implement the “AwarenessInfo” element to retrieve and show the attention info for every distant person from the view mannequin.

  7. Implement a technique within the encapsulated view mannequin interface to replace the attention info for every distant person.

  8. When a distant person adjustments the cursor location, dispatch the replace technique within the encapsulated view mannequin interface to replace the corresponding consciousness info.

  9. Within the “AwarenessInfo” element, use the brand new location to animate the cursor change.

  10. Write assessments for the brand new “AwarenessInfo” element, together with data-driven assessments to make sure it correctly shows the attention info for all distant customers.

  11. Write assessments for the encapsulated view mannequin interface to make sure it correctly updates the attention info for distant customers.

  12. Write end-to-end Cypress assessments to make sure the attention layer is correctly displayed and up to date when distant customers change their cursor location.

By following these duties, we are able to implement the attention layer function within the on-line whiteboard system.

The response from ChatGPT is non-deterministic, so this result’s simply
one run. In case you (or I) put this immediate in once more, we’ll get a unique response.

The important thing ingredient of the response is an ordered process record, with numbers for
every step. We are able to use these numbers to refer to those duties for the remainder of
the session.

The LLM responds with a plan.

Xu Hao can now have a look at the grasp plan and see if it is sensible. If
there are issues with it, he can ask ChatGPT to re-write a number of the duties
or all the grasp plan by offering extra info.

For instance, the present grasp plan was not grouped by differing kinds
of parts. One attainable immediate is perhaps:

Rewrite the grasp plan, and group duties as mannequin, view mannequin and examine.

This could possibly be helpful to confirm the alignment between the structure
imaginative and prescient and grasp plan.

These tweaks could possibly be added to the preliminary immediate to get an general
higher response.For this explicit immediate, the next instruction was
not within the preliminary immediate, however have been added in later iteration:

4. Don’t dispatch motion immediately to alter the states of shared view
    mannequin, use an encapsulated view mannequin interface as an alternative. Within the interface,
    every redux motion is mapped to a technique. Examined in vitest.

As soon as the grasp plan seems to be affordable, Xu Hao asks to regenerate it with
names for this system components.

Rewrite the grasp plan, this time with detailed element names,
strategies, and props

Certain, here is a revised grasp plan with detailed element names, strategies, and props:

Process Checklist:

  1. Create a brand new element named “AwarenessInfo” to render cursor and identify for distant customers. This element will obtain props reminiscent of person identify, cursor place, and on-line standing.

  2. Create a brand new Konva layer named “AwarenessLayer” to carry all of the “AwarenessInfo” parts.

  3. Add the “AwarenessLayer” to the primary whiteboard element.

  4. Create a brand new Redux slice named “consciousness” to retailer the attention info for distant customers.

  5. Create an encapsulated view mannequin interface named “AwarenessViewModel” for the “consciousness” Redux slice. This interface will include a technique named “updateUserAwarenessInfo” to replace the attention info for every distant person.

  6. Implement the “AwarenessInfo” element to retrieve and show the attention info for every distant person from the view mannequin. The element will use the “useSelector” hook to retrieve the related info.

  7. Implement the “updateUserAwarenessInfo” technique within the “AwarenessViewModel” interface to dispatch a Redux motion to replace the attention info for every distant person. This technique will probably be referred to as by the “MainWhiteboard” element.

  8. When a distant person adjustments the cursor location, dispatch the “updateUserAwarenessInfo” technique within the “AwarenessViewModel” interface to replace the corresponding consciousness info.

  9. Within the “AwarenessInfo” element, use the brand new location to animate the cursor change. This may be completed utilizing the Konva “Tween” animation.

  10. Write assessments for the brand new “AwarenessInfo” element, together with data-driven assessments to make sure it correctly shows the attention info for all distant customers. The assessments will probably be written utilizing the React Testing Library.

  11. Write assessments for the encapsulated view mannequin interface to make sure it correctly updates the attention info for distant customers. These assessments will use the Vitest framework.

  12. Write end-to-end Cypress assessments to make sure the attention layer is correctly displayed and up to date when distant customers change their cursor location. The assessments will simulate person interactions with the whiteboard and confirm the state of the “AwarenessLayer” and “AwarenessInfo” parts.

At this level, Xu Hao is able to ask ChatGPT to supply code for certainly one of
the steps

present me instance assessments and implementation for the attention slice
and AwarenessViewModel. Check first.

Certain, let's begin with writing assessments for the attention slice and AwarenessViewModel:

Check for Consciousness Slice:

describe('consciousness slice', () => 
  let retailer: MockStore;
  beforeEach(() => 
    retailer = configureMockStore()(initialState);
  );

  it('ought to set the cursor location', () => 
    retailer.dispatch(setCursorLocation( x: 100, y: 200 ));
    const state = retailer.getState().consciousness;
    count on(state.cursorLocation).toEqual( x: 100, y: 200 );
  );

  it('ought to set the identify', () => 
    retailer.dispatch(setName('John Doe'));
    const state = retailer.getState().consciousness;
    count on(state.identify).toBe('John Doe');
  );

  it('ought to set the web standing', () => 
    retailer.dispatch(setOnline(true));
    const state = retailer.getState().consciousness;
    count on(state.isOnline).toBe(true);
  );
);
Check for AwarenessViewModel:

describe('AwarenessViewModel', () => 
  let retailer: MockStore;
  let awarenessViewModel: AwarenessViewModel;
  beforeEach(() => 
    retailer = configureMockStore()(initialState);
    awarenessViewModel = new AwarenessViewModel(retailer);
  );

  it('ought to set the cursor location', () => 
    awarenessViewModel.setCursorLocation( x: 100, y: 200 );
    const state = retailer.getState().consciousness;
    count on(state.cursorLocation).toEqual( x: 100, y: 200 );
  );

  it('ought to set the identify', () => 
    awarenessViewModel.setName('John Doe');
    const state = retailer.getState().consciousness;
    count on(state.identify).toBe('John Doe');
  );

  it('ought to set the web standing', () => 
    awarenessViewModel.setOnline(true);
    const state = retailer.getState().consciousness;
    count on(state.isOnline).toBe(true);
  );
);
Now, let's implement the Consciousness slice and AwarenessViewModel:

Consciousness slice:

interface AwarenessState 
  cursorLocation:  x: quantity; y: quantity ;
  identify: string;
  isOnline: boolean;


const initialState: AwarenessState = 
  cursorLocation:  x: 0, y: 0 ,
  identify: '',
  isOnline: false,
;

const awarenessSlice = createSlice(
  identify: 'consciousness',
  initialState,
  reducers: 
    setCursorLocation: (state, motion: PayloadAction) => 
      state.cursorLocation = motion.payload;
    ,
    setName: (state, motion: PayloadAction) => 
      state.identify = motion.payload;
    ,
    setOnline: (state, motion: PayloadAction) => 
      state.isOnline = motion.payload;
    ,
  ,
);

export const  setCursorLocation, setName, setOnline  = awarenessSlice.actions;

export default awarenessSlice.reducer;
AwarenessViewModel:

class AwarenessViewModel 
  non-public retailer: Retailer;

  constructor(retailer: Retailer) 
    this.retailer = retailer;
  

  setCursorLocation(cursorLocation:  x: quantity; y: quantity ) 
    this.retailer.dispatch(setCursorLocation(cursorLocation));
  

  setName(identify: string) 
    this.retailer.dispatch(setName(identify));
  

  setOnline(isOnline: boolean) 
    this.retailer.dispatch(setOnline(isOnline));
  


Word: On this implementation, I assumed that the AwarenessViewModel is used to
set the values within the retailer. If it's also used to retrieve the values from the
retailer, we are able to add corresponding strategies to the AwarenessViewModel class

(I’ve eliminated the styling from ChatGPT’s response.)

General this code is neither dangerous nor excellent. It will not run because it stands.
It is lacking a configureMockStore and assumes it makes use of
redux-mock-store, which is not a part of the tech stack. Xu Hao can repair these
issues by asking the LLM to progressively tweak the code by a
sequence of rewrites. This is one to get rig of redux-mock-store.

rewrite the assessments with out redux-mock-store

Certain, here is an instance implementation of the assessments for the AwarenessViewModel and the attention slice with out utilizing redux-mock-store:

import createAwarenessSlice, AwarenessState, updateAwareness from './awarenessSlice'; import AwarenessViewModel from './AwarenessViewModel'; describe('AwarenessViewModel', () => { const mockDispatch = jest.fn(); const initialState: AwarenessState = customers: 'user1': identify: 'Consumer 1', cursorPosition: x: 0, y: 0 , on-line: true, , 'user2': identify: 'Consumer 2', cursorPosition: x: 100, y: 100 , on-line: true, , , ; let viewModel: AwarenessViewModel; beforeEach(() => viewModel = new AwarenessViewModel(initialState, mockDispatch); ); describe('getUserNames', () => it('returns the record of person names', () => const names = viewModel.getUserNames(); count on(names).toEqual(['User 1', 'User 2']); ); ); describe('getCursorPosition', () => it('returns the cursor place for a given person ID', () => const cursorPosition = viewModel.getCursorPosition('user1'); count on(cursorPosition).toEqual( x: 0, y: 0 ); ); it('returns undefined for an unknown person ID', () => const cursorPosition = viewModel.getCursorPosition('unknown'); count on(cursorPosition).toBeUndefined(); ); ); describe('updateCursorPosition', () => it('dispatches the updateAwareness motion with the brand new cursor place', () => viewModel.updateCursorPosition('user1', x: 50, y: 50 ); count on(mockDispatch).toHaveBeenCalledWith(updateAwareness( person: 'user1', cursorPosition: x: 50, y: 50 , )); ); ); });

And here is an instance implementation of the awarenessSlice with out utilizing redux-mock-store:

import createSlice, PayloadAction from '@reduxjs/toolkit'; export kind CursorPosition = x: quantity, y: quantity ; export kind UserAwareness = identify: string; cursorPosition: CursorPosition; on-line: boolean; ; export kind AwarenessState = customers: File<string, UserAwareness>; ; const initialState: AwarenessState = customers: , ; const awarenessSlice = createSlice( identify: 'consciousness', initialState, reducers: updateAwareness: (state, motion: PayloadAction< person: string, cursorPosition: CursorPosition >) => const person, cursorPosition = motion.payload; state.customers[user].cursorPosition = cursorPosition; , setOnline: (state, motion: PayloadAction< person: string, on-line: boolean >) => const person, on-line = motion.payload; state.customers[user].on-line = on-line; , , ); export const updateAwareness, setOnline = awarenessSlice.actions; export const awarenessReducer = awarenessSlice.reducer;

Word that since we’re not utilizing redux-mock-store, we have to go a mock dispatch operate to the view mannequin constructor, and within the take a look at for updateCursorPosition we have to test that the proper motion was dispatched utilizing jest.fn().

Additional rewrites can match in additional changes, reminiscent of asking it to
Use `configureTestingStore(‘consciousness’, initialState)` to initialize
the take a look at retailer
can get it to slot in with the staff’s take a look at idioms. Then
with some path changes and a small quantity of glue for the reducers, the
code can run simply high-quality.

As soon as that is working, Xu Hao can repeat the method for the remainder of the
duties within the grasp plan.

My take away from this dialogue was that utilizing chain of thought and
generated data prompting approaches generally is a considerably great tool
for programming. Particularly it reveals that to make use of LLMs nicely, we have to
discover ways to assemble prompts to get one of the best outcomes. This expertise
means that it is helpful to work together with the LLM like a junior associate,
beginning them with architectural tips, asking them to point out their
reasoning, and tweaking their outputs as we go.