LLMs and Prompt Engineering

LLMs in Sw Development
 To improve productivity (of delivered code), LLMs can be used to do
some of the tasks (or parts of some tasks) of engineers
 Benefits can be limited
 E.g. if in a 50KLOC delivered code, 40KLOC was reused code, 10K was written code,
then LLM can only be used to reduce effort for 10K
 It can also improve quality – if generated code is of good quality (which as of now is
not fully true)
 But LLMs can be used for a host of tasks, not just code generation –
help in requirements, design, testing, installation/deployment,
refactoring, re-engineering, …
 So companies have aggressively employed them in their work flows
 Outputs of LLMs depends significantly on how prompts are given
 As context from prompts significantly impacts what LLMs generate
 There is randomness built in LLMs generation algorithms
 Proper prompting can help in getting the most out of LLM for a task

2
Prompt Engineering
 For general queries, we can prompt as we wish
 When reliability of the output generated is important
– ad hoc prompting does not work well
 In Sw, we want precise output without errors
 Prompt engineering has been coined after genAI tools
came – to guide users to prompt the AI suitably to get
the desired results
 At the top level, can dividing prompting approaches
into two broad groups
 Single prompt approaches
 Multi-prompt approaches
 Both use prompts – let us understand an effective
prompt
 Goal of an effective prompt: To get the desired output

3
Parts of an Effective Prompt
 A well structured prompt can guide the LL Model
suitably to generate accurate and relevant output
 For somewhat complex tasks, a prompt should
contain the following parts
 Task description: Concisely and clearly state what you
want the LLM to do
 Input data: Any relevant input data, context, examples,
etc that can help an LLM do the task should be
provided; helps generate output consistent with inputs
 Output format: If desired format is clearly specified
(e.g. different parts / sections of the output, ordering,
..), lesser work to be done after it

4
Effective Prompt…
 Additional Instructions (if any): Specific
instructions on how to solve a problem, guidelines
or principles to be followed, ..
 Evaluation criteria: How the generated output will
be evaluated, i.e. what constitutes a good output

 In these parts, many tricks can be used to get the

5
Giving Examples
 Provide some examples to guide the LLM
 One example – one-shot
 Multiple examples – few-shot (not always better)
 E.g. if you want a function for computing HCF, you can
give one or two examples
 Inputs: 15, 25
 Answer: 5
 Inputs: 88, 66
 Answer: 22
 Can give examples as part of instructions - which
you can specify separately in ChatGPT

6
Role Prompting
 LLMs have digested a lot of information – also
about how different roles perform some tasks
 Hence, role prompting – asking LLM to play the
role of someone – can help getting desired output
 You are a requirements analyst, …
 You are a cybersecurity expert
 Prompt: Imagine you are an end user interacting with a
new e-commerce platform. Describe the features and
functionalities you would expect …
 Prompt: You are a systems architect responsible for
designing a scalable back-end for a social media
platform. Describe the components of the application
described below ….
7
Chain-of-Thought
 Treating LLMs as a black-box: no visibility on how
the answer has been reached – unsatisfactory
 Can ask the LLM to solve some problem in a step
wise manner, giving the output after each step –
better control, more visibility
 E.g. Generate code for this problem: …. Pls solve this
problem step-by-step as follows
 Step 1: …
 Step 2: ..

8
Discussion. Other techniques for
better prompting
 Pls share your experience - what can you do in a
single prompt approach to get a better response

9
Multi-Prompt Approaches
 Single prompt approaches work well for many
tasks – if the prompts are well structured and
provide the desired information
 Their scope will increase
 For complex tasks, single prompt insufficient – a
series of prompts needed to solve
 Discuss a few approaches

10
Prompt-Chaining
 The task is broken into a series of subtasks –
output of a subtask goes as input to the next
 LLM prompted for each subtask in sequence (each
prompt well structured)
 It simplifies prompt creation (often challenging) –
as simpler prompts for each subtask
 Provides control and visibility into problem solving
– incremental problem solving easier to verify

11
Prompt Chaining…
 Prompt 1: Generate python code for this problem:
ask for a file name, sum up the numbers,…
 Prompt 2: Now enhance the code to handle the
cases where the file does not exist
 Prompt 3: Now enhance the code such that if
inputs in the file are not numbers they are skipped

12
Prompt Chaining with Chain of
Thought
 Prompt chaining, but each subsequent step
determined based on progress so far
 Following steps can be followed
 Initial prompt
 Evaluate models response – identify how the response
should be improved
 Refine prompt for better response – provide more context,
information, …
 Repeat till step done well
 Then move to the next step in problem solving
 Problem is being solved step-by-step
 Output by LLM can be reviewed and corrected (at each step)
 Corrected output is part of the context for next subtask

13
Exploratory
 Can be useful for learning and exploration
 Initial prompt: Explain the bubble sort algo in simple
terms
 Evaluation: Assess the answer and see where you want
more clarity
 Refined prompt: Elaborate on time complexity …
…

14
Multiple Prompts for Consistency
 Self-consistency approach – keep refining the
prompt till you get the desired output
 Goal is to gradually converge
 Sometimes used to check the model output by
giving variations of the same input / prompt to
check the consistency

15
Exploratory Prompting
 Initial prompt can be to ask the model to give
multiple possible solutions / answers to the given
task / problem
 Then can ask for more information; select one and
then ask for more details on that
 Also possible to make it a tree-of-thought
 E.g. I have to develop an application: …; Please suggest
at least two mores suitable architecture styles for
building this
 This prompt style is expensive for sw tasks as user
has to evaluate different replies
 Copilot has one such option

16
Discussion – other approaches
 Other approaches you may have tried for sw
related task and which worked well

17
Prompting for Requirements
 LLMs can be used for almost any task done by
engineers in a software project
 For the requirements activity also, they can be
used – for specification as well as verification
 Use in analysis / elicitation – not clear as this is still
largely a human activity
 For standard type of applications – might help even in
eliciting requirements

18
For Generating SRS (Specification)
 After initial analysis, you have the problem title
(which has semantic information), user categories,
some notes on purpose, scope, goals, … Some
constraints may also become clear
 Compile this information (so it is suitable for
prompting) – use LLM help in SRS preparation
 For SRS in a specific format – specify the output
format (e.g. ChatGPT allows this)
 With these two context information, chatGPT can
be asked “produce an SRS for this application in
the format provided”

19
Generating SRS…
 One prompt approach– unlikely to work
 Desired output is too much (an SRS may be long)
 Better to have LLM generate different sections of the
SRS separately by using prompt chaining
 Allows LLM to provide a more detailed output for each
section / part
 Allows the analyst to correct the LLM’s hidden assumptions
by correcting the outputs
 Eg. Give the context, structure of the output,…
1. Generate the list of all functional requirements as a table
of use cases – one for each user category
2. Generate the non-functional requirements
3. Provide details of use case – give one prompt for each UC

20
For Requirements Validation
 Once SRS has been produced, it has to be validated for
key properties of correctness, completeness,
consistency, unambiguity
 LLMs can help in this
 Single prompt approach unlikely to work – better to
ask LLM to validate each property separately
 Some properties are per-requirement: Correctness,
unambiguity
 Check for these for portions of the requirements – e.g. first
for functional and then for non-functional
 Some are on overall SRS: Consistency, completeness
 For this the complete SRS has to be given

21
Looking for Errors…
 Providing examples of quality attributes help
 E.g. give examples of ambiguous statements along with
unambiguous equivalent
 Example 1: Ambiguous text
 Unambiguous text: … text
 Example 2: Ambiguous text
 Unambiguous text: … text
 Then give it some of the requirements (numbered)
and ask it to check for ambiguity
 Besides identifying errors, can also ask LLM to
suggest fixing the errors / suggest improvement

22
SRS-wide Errors
 Consistency and completeness are SRS-wide –
cannot do requirement-by-requirement
 Give the whole SRS and ask for inconsistencies – a
single prompt is suitable here
 For completeness, give the whole SRS
 It will help if LLM is asked to act as a stakeholder,
so it can think from that perspective
 Background info on the application
 The SRS of this application is: … SRS…
 Find the missing requirements for this SRS, think step-
by-step acting like each of the stakeholders
 Or you can ask it to check for one type of user, then
another type, ….

23
Which LLM to Use
 Interactive like ChatGPT – general purpose and can
be used for any task
 Built into the tools used by engineers to assist in
the task
 E.g. CoPilot in VS Code – programmer continues
programming, the LLM helps in code completion
(actually can do more)
 Focused bots for specific tasks – they are evolving
– often fine tuned for the specific task and hence
are more reliable (and may be “lighter”)

24