Skip to content

Optimisation Strategies

Practical techniques to reduce token usage, improve response quality, and minimise environmental impact whilst maintaining or improving output quality.

The Optimisation Hierarchy

Level 1: Query Planning
├─ Think before prompting
├─ Define clear objectives
└─ Batch related questions

Level 2: Prompt Engineering
├─ Use frameworks (AUTOMAT/CO-STAR)
├─ Specify output format upfront
└─ Add constraints to reduce iterations

Level 3: Caching & Reuse
├─ Save successful prompts as templates
├─ Cache outputs for incremental updates
└─ Build knowledge base

Level 4: Tool Selection
├─ Use NLMs for simple tasks
├─ Use smallest LLM that meets quality threshold
└─ Implement local models for repetitive work

Level 5: Advanced Techniques
├─ Fine-tuning for domain-specific tasks
├─ Quantization for faster inference
└─ RAG (Retrieval Augmented Generation)

Strategy 1: The One-Shot Principle

Goal: Get the right output on the first attempt.

The Problem: Iterative Refinement

Typical inefficient workflow:

[Attempt 1] "Summarise this paper"
→ Too generic

[Attempt 2] "Make it more technical"
→ Wrong focus

[Attempt 3] "Focus on synthesis methods"
→ Missing quantitative data

[Attempt 4] "Include temperatures and pressures"
→ Finally acceptable

Cost: 4× the tokens, time, and environmental impact

The Solution: Comprehensive Upfront Specification

[Single Attempt] Using AUTOMAT:

Act as a Materials Science Technical Reviewer.

Task: Summarise synthesis methodology from the attached paper.

Output: Structured summary with sections:
1. Synthesis method (brief description)
2. Key parameters (table: parameter, value, units)
3. Novel aspects (bullet points)
4. Reproducibility assessment (high/medium/low + justification)

Constraints:
- Extract only explicitly stated values (no estimation)
- Focus on sections 2-3 of paper (synthesis and characterisation)
- Ignore theoretical calculations in section 4
- Flag any ambiguous or missing critical parameters

Tone: Technical, objective, suitable for replication assessment

Result: First attempt yields production-ready output

Savings: Reduction in tokens and time

Strategy 2: Strategic Context Pruning

Goal: Include only relevant context, not everything.

The Problem: Context Overload

[Bloated Context]
"I'm a materials scientist at AmaDema, a company founded in 2018 
by Dr. Smith and Dr. Jones, specialising in advanced ceramics and 
polymer nanocomposites. We have 15 employees across R&D, production, 
and sales. Our main products are electrospun nanofibers for 
aerospace and biomedical applications. We're currently working on a 
project funded by an EPSRC grant (£450k over 3 years) to develop 
sustainable alternatives to PLA. My role is Senior Researcher 
focusing on synthesis optimisation. I've been here for 2 years after 
completing my PhD at Bristol on graphene functionalisation. 
Currently, we're in month 14 of the project and need to demonstrate 
proof-of-concept by month 18 for the next review. 

Could you help me format this synthesis protocol?"

Problems: - 90% of context irrelevant to task - Increased token cost - Potential information leakage (company details)

The Solution: Task-Relevant Context Only

[Optimised Context]
"Act as a Laboratory Documentation Specialist.

Context: Converting informal lab notes to standardised synthesis 
protocol format for materials science publication.

Task: Format the following lab notes according to standard 
structure (Materials, Equipment, Procedure, Characterisation).

[LAB NOTES]"

Savings: Reduction in tokens, eliminated IP leakage risk

Strategy 3: Structured Output Specification

Goal: Reduce back-and-forth by defining exact format upfront.

The Problem: Format Mismatches

[Vague Request]
"Extract synthesis parameters from these 10 papers"

[Output]
Long prose paragraphs mixing papers, inconsistent parameter mentions, 
difficult to compare...

[Follow-up]
"Can you put that in a table?"

[Follow-up 2]
"Can you add the DOI column?"

[Follow-up 3]
"Can you sort by temperature?"

Total: 4 queries to get usable output

The Solution: Specify Format Exactly

[Structured Request]
"Extract synthesis parameters from these 10 papers.

Output: Markdown table with columns:
- Paper (Author Year)
- DOI
- Polymer
- Nanofiller
- Nanofiller Loading (wt%)
- Synthesis Method
- Key Temperature (°C)
- Solvent System

Sort by: Temperature (ascending)

If parameter not reported: mark 'NR'
If ambiguous: mark 'Ambiguous [brief note]'"

Result: Single query yields ready-to-use table

Savings: Reduction in iterations

Strategy 4: Constraint Engineering

Goal: Prevent hallucinations and irrelevant content through explicit constraints.

Effective Constraints

Whitelisting (What to Include)

Constraints:
- Focus exclusively on peer-reviewed journal articles
- Only papers published 2020-2024
- Only electrospinning-based synthesis methods
- Must explicitly report mechanical properties

Blacklisting (What to Exclude)

Constraints:
- Exclude review papers and book chapters
- Exclude theoretical/computational studies
- Exclude patents and conference proceedings
- Ignore papers on melt-spinning or solution casting

Handling Uncertainty

Constraints:
- If data is missing, mark as "Not reported"—do not estimate
- If multiple values given, report as range
- If conflicting values across papers, flag discrepancy
- Do not extrapolate from related materials

Strategy 5: Prompt Chaining with Checkpoints

Goal: Break complex tasks into steps with validation.

When to Use

✅ Complex multi-step analysis
✅ Tasks requiring intermediate validation
✅ High-stakes outputs needing verification

Step 1: Broad Identification

Prompt 1: "Review these 50 patents and identify those claiming 
electrospinning-based synthesis of polymer/graphene composites. 
Output: List of patent numbers that meet criteria."

→ Human checkpoint: Verify list before proceeding

Step 2: Detailed Extraction

Prompt 2: "For the following 12 patents [list from Step 1], extract:
- Claims related to synthesis method
- Graphene content ranges
- Processing temperatures
- Novelty statements

Output: Comparative table"

→ Human checkpoint: Verify accuracy of extracted data

Step 3: Overlap Analysis

Prompt 3: "Compare these 12 patents against our method [description]. 
For each patent, assess overlap: High/Medium/Low/None. Justify 
assessment."

→ Human checkpoint: Review before legal submission

Why better than single query: - Intermediate validation catches errors early - Each step can be cached and reused - Easier to debug if issues arise - Human expertise applied at critical junctions

Strategy 6: Templating for Recurring Tasks

Goal: Create reusable, proven prompts for common tasks.

Template Library

Template 1: Literature Summary

Act as a {DOMAIN} Specialist reviewing literature for {PURPOSE}.

Task: Summarise key findings from the following paper relevant to 
{SPECIFIC_FOCUS}.

Paper: {TITLE}, {DOI}

Output format:
1. Main finding (1 sentence)
2. Methodology (brief, 2-3 sentences)
3. Key quantitative results (table if >3 values, else bullets)
4. Limitations/caveats (bullets)
5. Relevance to {SPECIFIC_FOCUS} (1-2 sentences)

Constraints:
- Extract only explicitly stated results
- If methodology unclear, note as limitation
- Do not compare to other papers

Tone: {TONE}

Usage: Replace {VARIABLES} each time, consistent quality

Template 2: Experimental Design Brainstorming

Act as a Senior {DOMAIN} Researcher with expertise in {SUBFIELD}.

Context: We are experiencing {PROBLEM_DESCRIPTION} in our current 
{PROCESS}. Existing approaches have limitations: {LIMITATIONS}.

Task: Propose 5 alternative experimental approaches to address this 
problem.

For each approach, provide:
1. Brief description (2-3 sentences)
2. Expected advantages (bullets)
3. Potential challenges (bullets)
4. Resource requirements (equipment, materials, time)
5. Risk level (High/Medium/Low)

Constraints:
- Focus on approaches feasible with {AVAILABLE_EQUIPMENT}
- Stay within {BUDGET_CONSTRAINT}
- Exclude approaches requiring >6 months development
- Prioritise approaches with literature precedent

Output: Numbered list, sorted by risk level (Low to High)

Building Your Template Library

  1. Identify recurring tasks in your workflow
  2. Perfect the prompt through iteration (once)
  3. Generalise by adding variables {PLACEHOLDER}
  4. Document the template with usage notes
  5. Share with team for collective efficiency

Strategy 7: Caching and Incremental Updates

Goal: Never regenerate what you've already processed.

Implementation

Month 1: Initial Build

Task: Comprehensive literature review of PLA/graphene electrospinning 
(100 papers, 2015-2024)

Output: Detailed database with synthesis parameters, properties, 
methods

[Save output as: PLA_Graphene_Database_2024-01.md]

Month 2: Incremental Update

Task: Update PLA/graphene database with papers published since 
2024-01-31

New papers: [5 DOIs]

Instructions:
1. Process only the 5 new papers
2. Use same format as existing database
3. Flag any contradictions with existing entries
4. Output only the new entries (I will append manually)

Savings: Reduction in computational cost (5 vs 105 papers)

Caching Best Practices

✅ Use consistent formatting (easy to append)
✅ Include timestamps (track what's current)
✅ Version control (track changes over time)
✅ Tag by topic (reuse across projects)
✅ Document sources (verify when needed)

Strategy 8: Query Batching

Goal: Process multiple items in single query to reduce overhead.

Example: SEM Image Analysis

Inefficient (Individual Processing)

[Query 1] "Analyse SEM image 1: measure average fiber diameter"
[Query 2] "Analyse SEM image 2: measure average fiber diameter"
[Query 3] "Analyse SEM image 3: measure average fiber diameter"
...
[Query 10] "Analyse SEM image 10: measure average fiber diameter"

Cost: 10 queries, high overhead per query

Efficient (Batched Processing)

[Single Query] "Analyse the following 10 SEM images of electrospun 
fibers. For each:
- Measure average fiber diameter (from 20 random fibers)
- Calculate standard deviation
- Assess fiber uniformity (CV%)
- Note any defects (beads, fusion points)

Output: Table with columns [Image ID, Avg Diameter (nm), SD, CV%, 
Defects (Y/N), Notes]

Images: [Attach all 10]"

Cost: 1 query, 70% less overhead

Savings: ~40% total reduction in tokens

Strategy 9: Local Model Fine-Tuning

Goal: Train domain-specific models for highly repetitive tasks.

When to Consider

✅ Task performed >1000 times/year
✅ Task is highly standardised
✅ Quality threshold well-defined
✅ Training data available or easy to generate

Process

Step 1: Generate Training Data

Use large model (GPT-4) to create 500-1000 example inputs/outputs:

Prompt: "Generate 50 diverse examples of synthesis protocol 
formatting tasks. For each:
- Input: Informal lab notes (varied styles)
- Output: Standardised ISO format protocol

Ensure diversity in: polymer types, synthesis methods, level of 
detail in notes"

Cost: High upfront (500-1000 examples)

Step 2: Fine-Tune Small Model

Use examples to fine-tune Llama 8B on this specific task:

# Technical implementation (simplified)
python fine_tune.py \
  --model llama-3.3-8b \
  --task protocol_formatting \
  --examples training_data.json \
  --epochs 3

Time: ~2-4 hours on GPU

Step 3: Deploy

Use fine-tuned model for all future protocol formatting:

Per-query cost: - Original (GPT-4): ~\(0.03, ~50g CO₂ - Fine-tuned (Llama 8B local): ~\)0.001, ~2g CO₂

Breakeven: After ~1000 uses (typical: 3-6 months)

Annual savings (2000 uses): - $56 cost savings - 96 kg CO₂ reduction

Strategy 10: Retrieval-Augmented Generation (RAG)

Goal: Give AI access to your specific documents without retraining.

What is RAG?

Traditional approach: AI relies only on training data
RAG approach: AI retrieves relevant chunks from your documents, then generates response

Materials Science Use Case

Scenario: You have 500 internal synthesis protocols, want AI to answer questions about them

Without RAG (Inefficient)

Option A: Include all 500 protocols in prompt
Problem: Exceeds context window, massive token cost

Option B: Manually find relevant protocols, include in prompt
Problem: Time-consuming, error-prone

With RAG (Efficient)

[Setup once]
1. Upload 500 protocols to vector database
2. Each protocol is embedded (numerical representation)

[Each query]
1. Your question is embedded
2. System finds 3-5 most relevant protocols automatically
3. Only those protocols included in prompt to AI
4. AI generates response using only relevant context

Benefits: - Automatic retrieval (no manual searching) - Token-efficient (only relevant docs) - Always up-to-date (add new protocols anytime) - Works with unlimited document libraries

Simple RAG Implementation

Tools: - LlamaIndex or LangChain (Python frameworks) - ChromaDB or FAISS (vector databases) - Ollama (local LLM)

Implementation time: 2-4 hours initial setup

Payoff: Ongoing efficiency for document-heavy workflows

Real-World Optimization: Before & After

Case Study: Quarterly Competitive Analysis

Before Training

Task: Analyse 50 competitor patents for quarterly R&D strategy meeting

Approach: 1. Read each patent manually 2. Ask AI for clarification on unclear sections (ad hoc queries) 3. Summarise findings in Word document 4. Ask AI to polish document

Time: 20 hours
Queries: ~100
Tokens: ~50,000
Environmental cost: 2.5 L water, 250g CO₂

After Training

Approach:

Step 1: Batch classification (NLM) 

Use BERT to classify 50 patents by technology area
→ Result: 12 relevant to our focus areas
→ 5 minutes, negligible environmental cost

Step 2: Structured extraction (AUTOMAT + LLM) 

Single batched query extracting synthesis methods, claims, novelty 
from 12 relevant patents
→ Result: Structured comparison table
→ 1 query, ~15 minutes

Step 3: Strategic analysis (CO-STAR + LLM) 

Single query generating competitive gap analysis and recommendations 
based on extracted data
→ Result: Executive summary with strategic insights
→ 1 query, ~10 minutes

Step 4: Cache for next quarter 

Save structured data, only process new patents next quarter

Time: 4 hours (80% reduction)
Queries: ~5 (95% reduction)
Tokens: ~5,000 (90% reduction)
Environmental cost: 0.25 L water, 25g CO₂ (90% reduction)
Quality: Higher (more structured, comprehensive)

Optimization Checklist

Before submitting any AI query, verify:

  • Objective clear? (What exactly do I need?)
  • Output format specified? (Table, bullets, code, etc.)
  • Constraints defined? (What to include/exclude?)
  • Could I batch this? (Process multiple items together?)
  • Can I reuse existing output? (Check cache first?)
  • Right model size? (Smallest that meets quality threshold?)
  • Could NLM handle this? (Simple extraction/classification?)
  • Have I used a template? (For recurring tasks?)

Goal: Answer "yes" to ≥6 of these before hitting enter

Exercise: Optimise Your Own Workflow

Challenge

Step 1: Identify one recurring AI task in your work

Step 2: Document your current approach: - How many queries per instance? - How long does it take? - What's the output quality?

Step 3: Apply 3 optimisation strategies from this page

Step 4: Test optimised approach in sandbox

Step 5: Calculate improvements: - % reduction in queries - % reduction in time - % reduction in tokens - Change in quality (better/same/worse)

Share your results with workshop group

Next: Advanced Conversational Learning →