Turn quantitative credit model outputs into plain-language narratives that non-technical stakeholders can act on — through iterative prompt refinement.
⏱ 35 minutes
Exercise Overview
AnyCompany's credit committees review PayLater and merchant financing applications daily. The quantitative risk models produce numbers — probability of default, loss given default, exposure at default — but committee members need plain-language narratives that explain what the numbers mean and whether to approve.
Currently, credit analysts spend 20-30 minutes per application writing these narratives manually. In this exercise, you'll build a prompt template that converts raw model outputs into consistent, balanced credit narratives in seconds.
⚙️ Setup: How to run this exercise
Use Kiro chat for this exercise. You'll paste prompts and observe how the output improves with each technique.
Keeping steps independent:
Start a New Session for each step (Steps 1–5, 7, 8)
Step 6 continues in the same session as Step 5
Each step saves to a unique filename (step1-..., step2-...) so you can compare outputs side-by-side at the end
A steering file is pre-configured to keep each step isolated — Kiro will not read previous step outputs when generating new ones
📦 First-time Kiro setup (do this once before starting)
If you haven't already (from Lab 6), download and extract this zip into your workspace root folder.
Extract into your workspace root → creates .kiro/steering/workshop-rules.md and .kiro/steering/exercise-isolation.md.
🎯 Exercise Approach
You'll refine a prompt through 6 steps — each applying a different technique from the Advanced Prompting curriculum. Then you'll validate the resulting template against a completely different applicant profile. The final deliverable is a reusable prompt template for credit risk narratives.
Techniques You'll Practice
Step
Prompting Technique
Curriculum Reference
Duration
Step 1
Zero-Shot Baseline
Module 1: Prompt Fundamentals
3 min
Step 2
Step-Back Prompting
Module 2: Chain-of-Thought (Step-Back variant)
4 min
Step 3
Audience Framing
Module 3: Role & Persona
4 min
Step 4
Multi-Perspective Analysis
Module 3: Multi-Agent Framing
5 min
Step 5
Structured Output + Length Control
Module 4: Structured Outputs
6 min
Step 6
Template Extraction (Meta-Prompting)
Module 5.2: Meta-Prompting
8 min
Step 7
Template Validation (New Data)
Production Testing
5 min
Step 8
Evaluate Your Template (Rubric + LLM-as-Judge)
Module 6: Evaluating Prompts
5 min
Sample Credit Application Data
You'll use this data throughout all steps. Copy it once — you'll paste it into each prompt.
CREDIT APPLICATION DATA — Copy this for use in all steps
APPLICANT: FreshBites Pte Ltd (Merchant ID: MRC-4412)
Application Type: PayLater Merchant Credit Line
Requested Amount: $150,000 SGD
Term: 12 months revolving
BUSINESS PROFILE:
- Registered: Singapore, 2021
- Category: Food & Beverage (cloud kitchen, 3 outlets)
- Annual revenue (2024): $1.8M SGD
- Annual revenue (2025 YTD, annualized): $2.4M SGD
- Employees: 28 full-time, 12 part-time
- Years on AnyCompany platform: 2.5 years
QUANTITATIVE MODEL OUTPUTS:
- Probability of Default (PD): 3.2%
- Loss Given Default (LGD): 45%
- Exposure at Default (EAD): $150,000 SGD
- Expected Loss (EL): $2,160 SGD (PD × LGD × EAD)
- Risk Rating: BB+ (internal scale: AAA to D)
- Debt Service Coverage Ratio (DSCR): 1.4x
- Current Ratio: 1.2
- Debt-to-Equity: 0.8
TRANSACTION HISTORY ON PLATFORM:
- Monthly AnyCompany Pay volume (last 6 months): $85K → $92K → $98K → $105K → $112K → $120K
- PayLater transaction share: 32% of total volume
- Chargeback rate: 0.3% (within benchmark)
- Settlement reliability: 99.8% on-time
EXISTING OBLIGATIONS:
- Bank term loan: $80,000 SGD outstanding (monthly repayment $4,200)
- Equipment lease: $12,000 SGD remaining (monthly $1,000)
- No existing AnyCompany credit facility
COLLATERAL / SECURITY:
- Proposed: Personal guarantee from director (Lim Mei Hua)
- No physical collateral offered
- AnyCompany platform receivables can be used as informal security
INDUSTRY CONTEXT:
- Singapore F&B sector growth: 8% YoY (2024-2025)
- Cloud kitchen segment growing at 15% YoY
- Average PD for F&B merchants on platform: 4.5%
- Average PD for all merchants on platform: 3.8%
RED FLAGS / NOTES:
- Director Lim Mei Hua has a previous business (LMH Trading) that was dissolved in 2019 — no outstanding liabilities found
- One late payment on bank term loan (7 days late, March 2025) — subsequently resolved
- Rapid revenue growth (33% YoY) — positive but requires sustainability assessment
Start with a minimal prompt to establish a baseline. This shows what the model produces with almost no guidance.
In the Kiro chat panel, start a New Session and paste:
PROMPT — Step 1: Zero-Shot
Write a credit risk narrative for this application. Save the output as "step1-zero-shot.md" in a "lab7-credit-risk" folder.
[PASTE CREDIT APPLICATION DATA HERE]
CREDIT APPLICATION DATA
🔍 Observe the output: The response likely dumps all the numbers back at you, uses jargon a committee member wouldn't understand, and doesn't clearly recommend APPROVE or DECLINE. Note what's missing.
Instead of diving straight into the narrative, ask the model to first identify the most important factors. This "step back" forces the model to prioritize before writing — producing a more focused, insightful narrative.
Start a New Session in Kiro and paste:
PROMPT — Step 2: Step-Back
Before writing anything, first identify the 3 most important risk factors and the 3 strongest mitigating factors from this credit application data. Rank them by significance.
Then, using those prioritized factors, write a credit risk narrative for the application.
Save the output as "step2-step-back.md" in the "lab7-credit-risk" folder.
[PASTE CREDIT APPLICATION DATA HERE]
CREDIT APPLICATION DATA
🔍 Compare with Step 1: The narrative should now lead with the most important points instead of listing everything equally. The step-back forces prioritization — the same skill a senior analyst uses instinctively.
💬 Why step-back works for credit narratives: Credit committees don't want to read every data point — they want to know "what matters most?" Step-back prompting mirrors how experienced analysts think: assess the landscape first, then write the story.
Step 3: Add Audience Framing
📖 Technique: Audience Framing (Module 3: Role & Persona)
Instead of assigning a persona to the AI, you define the audience. This shapes the vocabulary, level of detail, and what gets explained vs. assumed.
Start a New Session in Kiro and paste:
PROMPT — Step 3: Audience Framing
You are a Senior Credit Analyst at AnyCompany Financial Group preparing a credit narrative for the Credit Committee.
Your audience is the Credit Committee — they understand business fundamentals and financial statements, but they are NOT statisticians. They need:
- Plain-language explanations of what the model outputs mean in practical terms
- Context: how does this applicant compare to peers?
- Clear "so what?" for each data point — don't just state numbers, explain their implications
- No unexplained acronyms on first use (spell out PD, LGD, DSCR on first mention)
First, identify the 3 most important risk factors and 3 strongest mitigating factors. Then write the credit narrative.
Save the output as "step3-audience.md" in the "lab7-credit-risk" folder.
[PASTE CREDIT APPLICATION DATA HERE]
CREDIT APPLICATION DATA
🔍 Compare with Step 2: The language should now be more accessible. Instead of "PD is 3.2%," you should see something like "The probability that this merchant defaults on the facility is 3.2% — lower than the 4.5% average for F&B merchants on our platform, indicating above-average creditworthiness for this segment."
Ask the model to present both the optimistic and cautious view. This prevents one-sided narratives and gives the committee a balanced picture — the same approach used in investment banking memos.
Start a New Session in Kiro and paste:
PROMPT — Step 4: Multi-Perspective
You are a Senior Credit Analyst at AnyCompany Financial Group preparing a credit narrative for the Credit Committee. The committee understands business fundamentals but not statistical models — explain implications in plain language.
First, identify the 3 most important risk factors and 3 strongest mitigating factors.
Then write the credit narrative with TWO clearly labeled perspectives:
**THE CASE FOR APPROVAL:**
Present the strongest arguments for why this credit facility should be approved. Reference specific data points that support creditworthiness, growth trajectory, and repayment capacity.
**THE CASE FOR CAUTION:**
Present the legitimate concerns and risks. What could go wrong? What assumptions are we making? What would make this application riskier than it appears?
After both perspectives, provide your balanced assessment — weighing both sides.
Save the output as "step4-multi-perspective.md" in the "lab7-credit-risk" folder.
[PASTE CREDIT APPLICATION DATA HERE]
CREDIT APPLICATION DATA
🔍 Compare with Step 3: The narrative now presents both sides explicitly. The committee can see the bull case AND the bear case, making their decision more informed. This is much more useful than a one-sided recommendation.
💬 Why multi-perspective matters in credit: Regulators and auditors look for evidence that credit decisions considered both upside and downside scenarios. A narrative that only says "approve" without acknowledging risks is a red flag in an audit. Multi-perspective prompting builds this balance in automatically.
Step 5: Add Structured Output + Length Control
📖 Technique: Structured Output + Length Control (Module 4)
Define exact sections, add a decision recommendation with conditions, and control the output length. This ensures every credit narrative is comparable and scannable.
Start a New Session in Kiro and paste:
PROMPT — Step 5: Structured + Length Control
You are a Senior Credit Analyst at AnyCompany Financial Group preparing a credit narrative for the Credit Committee. The committee understands business fundamentals but not statistical models.
GROUNDING RULES:
- Base your narrative ONLY on the data provided. Do not add external information.
- Every claim must reference a specific data point from the input.
- If data is insufficient for an assessment area, state: "[INSUFFICIENT DATA: need X]"
- Spell out all acronyms on first use.
Produce a Credit Risk Narrative with EXACTLY these sections:
1. EXECUTIVE SUMMARY (3-4 sentences max)
- Who is the applicant, what are they requesting, and what is your recommendation?
2. APPLICANT OVERVIEW (1 paragraph)
- Business profile, platform history, revenue trajectory
3. KEY RISK METRICS (table format)
| Metric | Value | Benchmark | Assessment |
Include: PD, LGD, EL, DSCR, Current Ratio, Debt-to-Equity, Chargeback Rate
4. THE CASE FOR APPROVAL (3-5 bullet points)
- Each bullet: data point → plain-language implication
5. THE CASE FOR CAUTION (3-5 bullet points)
- Each bullet: concern → specific data point → potential impact
6. RISK MITIGANTS & CONDITIONS
- What conditions would reduce the identified risks?
- Suggested covenants or monitoring requirements
7. RECOMMENDATION
One of: APPROVE | APPROVE WITH CONDITIONS | DECLINE
- 2-3 sentence justification referencing the key evidence
- If APPROVE WITH CONDITIONS: list the specific conditions
8. MONITORING TRIGGERS
- What metrics should be watched post-approval?
- At what thresholds should the facility be reviewed?
TOTAL LENGTH: Keep the entire narrative under 800 words. Be concise — every sentence must earn its place.
Save the output as "step5-structured.md" in the "lab7-credit-risk" folder.
[PASTE CREDIT APPLICATION DATA HERE]
CREDIT APPLICATION DATA
🔍 Compare with Step 4: The output is now structured, scannable, and concise. The committee can jump to Section 7 for the recommendation, or read the full narrative for context. The 800-word limit forces the model to prioritize.
Step 6: Extract the Reusable Template
📖 Technique: Meta-Prompting (Module 5.2)
Meta-prompting asks the AI to analyze your conversation and produce a reusable artifact. The quality of the template depends on how well you instruct the extraction — what to include, how to structure it, and what makes it reusable.
In the same session from Step 5 (do not start a new one), paste this follow-up:
PROMPT — Step 6: Template Extraction
Now convert this into a reusable template that any credit analyst can use for ANY PayLater or merchant financing application.
Create a Markdown file called "credit-risk-narrative-prompt-template.md" and save it in a "prompt-templates" folder.
The template should:
- Be completely self-contained — a new analyst should be able to use it without additional training
- Use {{variables}} for all application-specific data (applicant name, amounts, model outputs, etc.)
- Include the persona, audience framing, grounding rules, output structure, and length control from our refined prompt
- Have clear usage instructions
Think about what makes a template truly production-ready and reusable at scale.
✅ Deliverable: Kiro will create a prompt-templates/credit-risk-narrative-prompt-template.md file. Open it and review the quality.
📤 Submit Your Template
Submit your template for automated scoring. How production-ready is it?
Resubmitting with the same name replaces your previous entry.
💡 How scoring works
Your template is sent to Amazon Bedrock which evaluates how production-ready it is — structure, reusability, guardrails, quality controls, and domain relevance. Scores appear on the leaderboard below.
🏆 Template Leaderboard
No submissions yet. Be the first!
🔐 Instructor Version — How would an expert extract this template?
After submitting, enter the passkey to reveal the production-grade extraction prompt. Compare it with what you used.
🎓 Instructor Version — Production-Grade Extraction Prompt
Compare this with the simplified prompt above. Notice the specificity in variable definitions, data format examples, and customization notes.
Excellent. Now convert this into a reusable template that any credit analyst can use for ANY PayLater or merchant financing application.
Create a Markdown file called "credit-risk-narrative-prompt-template.md" and save it in a "prompt-templates" folder. Structure it as follows:
## HEADER
- Title: "Credit Risk Narrative — Prompt Template"
- Version, date, purpose, usage instructions
## TEMPLATE USAGE GUIDE
A table listing ALL variables with: Variable name, Description, Expected Format, Example.
Variables should include:
{{applicant_name}}, {{merchant_id}}, {{application_type}}, {{requested_amount}}, {{term}}, {{currency}}, {{business_profile}}, {{model_outputs}}, {{transaction_history}}, {{existing_obligations}}, {{collateral}}, {{industry_context}}, {{red_flags}}, {{analysis_period}}
## DATA FORMAT EXAMPLES
Show the exact format expected for complex variables (model_outputs, transaction_history, existing_obligations) with realistic sample data.
## PREREQUISITES
What data to gather before using the template.
## ---START PROMPT--- / ---END PROMPT---
The full prompt containing:
- Persona and audience framing
- Grounding rules
- The 8-section output format with descriptions
- Length control instruction
- All {{variables}} in a clearly organized input block at the end
## CUSTOMIZATION NOTES (after ---END PROMPT---)
1. **Application Type Adjustments** — How to modify for PayLater credit lines vs. merchant term loans vs. working capital facilities
2. **Risk Appetite Calibration** — How to adjust recommendation thresholds for conservative vs. growth-oriented credit policies
3. **Regulatory Considerations** — Table of SEA markets with relevant credit/lending regulations
4. **Modifiable Sections** — What can/cannot be changed and why
The template must be self-contained — a new analyst should be able to use it without additional training.
💡 Teaching point: The simplified prompt produces a decent template. The instructor version produces a production-grade one. The difference? Specificity — naming every variable, defining data formats, including customization notes and regulatory tables.
Step 7: Validate the Template with New Data
📖 Technique: Production Testing
Test your template against a completely different applicant — different market, different product, different risk profile — to confirm it generalizes.
How to use your template
Open the file prompt-templates/credit-risk-narrative-prompt-template.md that Kiro created in Step 6
Find the section between ---START PROMPT--- and ---END PROMPT---
Copy that entire block — this is your reusable prompt
Start a New Session in Kiro
Paste the prompt, then replace the entire applicant data section (everything after the input/data variables block) with the test data below — no need to replace variables one by one, just swap the whole data block
Add this instruction at the end: "Save the output as run1.md in the lab7-credit-risk folder."
Test data — a deliberately harder case
TEST DATA — GoRide Motors (Malaysia, declining revenue)
APPLICANT: GoRide Motors Sdn Bhd (Merchant ID: MRC-7803)
Application Type: Working Capital Facility
Requested Amount: RM 500,000 (Malaysian Ringgit)
Term: 6 months revolving
BUSINESS PROFILE:
- Registered: Malaysia, 2019
- Category: Automotive Services (motorcycle rental & maintenance for ride-hailing drivers)
- Annual revenue (2024): RM 3.2M
- Annual revenue (2025 YTD, annualized): RM 2.8M (declining)
- Employees: 45 full-time
- Years on AnyCompany platform: 3 years
QUANTITATIVE MODEL OUTPUTS:
- Probability of Default (PD): 6.8%
- Loss Given Default (LGD): 55%
- Exposure at Default (EAD): RM 500,000
- Expected Loss (EL): RM 18,700
- Risk Rating: B (internal scale: AAA to D)
- Debt Service Coverage Ratio (DSCR): 1.1x
- Current Ratio: 0.9
- Debt-to-Equity: 1.6
TRANSACTION HISTORY ON PLATFORM:
- Monthly AnyCompany Pay volume (last 6 months): RM 280K → RM 265K → RM 250K → RM 240K → RM 235K → RM 228K
- PayLater transaction share: 18% of total volume
- Chargeback rate: 0.6% (at upper benchmark)
- Settlement reliability: 97.2% on-time
EXISTING OBLIGATIONS:
- Bank term loan: RM 200,000 outstanding (monthly repayment RM 12,000)
- Vehicle fleet financing: RM 350,000 remaining (monthly RM 18,000)
- Existing AnyCompany credit line: RM 100,000 (85% utilized)
COLLATERAL / SECURITY:
- Proposed: Fleet of 120 motorcycles (estimated value RM 600,000)
- Personal guarantee from director (Ahmad Razak bin Ismail)
- AnyCompany platform receivables
INDUSTRY CONTEXT:
- Malaysia ride-hailing market growth: 3% YoY (slowing from 12% in 2023)
- Motorcycle rental segment facing pressure from e-bike alternatives
- Average PD for automotive services merchants: 5.2%
- Average PD for all merchants on platform (Malaysia): 4.0%
RED FLAGS / NOTES:
- Revenue declining 12.5% YoY — director attributes to "seasonal adjustment" but trend is 6 months
- Current ratio below 1.0 indicates potential liquidity stress
- Existing AnyCompany credit line at 85% utilization — high dependency
- Director Ahmad Razak has 2 other businesses (both active, no flags)
- Fleet maintenance costs increased 22% in last 6 months
Run it 3 times for consistency testing
Start a New Session for each run. Use the same template + test data each time, but change the output filename:
Run
Add this to the end of your prompt
Run 1
Save the output as "run1.md" in the "lab7-credit-risk" folder.
Run 2
Save the output as "run2.md" in the "lab7-credit-risk" folder.
Run 3
Save the output as "run3.md" in the "lab7-credit-risk" folder.
🔍 Validate the outputs: This is a deliberately harder case — declining revenue, weak liquidity, high existing debt. Check across all 3 runs:
Does the template handle Malaysian Ringgit (not just SGD)?
Does it correctly identify this as a higher-risk application?
Does the multi-perspective section present a genuine "case for caution"?
Is the recommendation appropriate (likely DECLINE or APPROVE WITH CONDITIONS)?
Are the 3 runs consistent in structure and recommendation?
💬 If the template doesn't handle this well: Common issues include: not adapting to a different currency, being too optimistic despite clear warning signs, or not adjusting the recommendation threshold. Iterate on the template — this is how production templates get hardened.
Step 8: Evaluate Your Template
Now use LLM-as-Judge to score each of your 3 runs from Step 7. You have run1.md, run2.md, and run3.md in your lab7-credit-risk folder.
Score each run
Start a New Session in Kiro. Paste the rubric below and tell Kiro which file to evaluate:
PROMPT — Evaluate Run 1
You are a STRICT expert evaluator for credit risk narratives. You have high standards and rarely give perfect scores. A score of 5 should be genuinely exceptional — most good outputs score 3-4.
Read the file "lab7-credit-risk/run1.md" and score on 4 criteria (1-5 each):
1. **Completeness** (1-5): Are ALL required sections present with SUBSTANTIVE content?
- 1 = Missing key sections (no bull/bear case, no recommendation)
- 2 = Missing 1-2 sections or several are just headers with one sentence
- 3 = All sections present but 2-3 are thin (under 2 sentences)
- 4 = All sections present with good detail, minor gaps
- 5 = RARE — every section has exceptional depth, the bull/bear cases present genuinely different arguments (not just rephrasing)
2. **Data Grounding** (1-5): Does EVERY factual claim cite a SPECIFIC metric (PD, LGD, DSCR, etc.)?
- 1 = Most claims are generic ("the applicant has good financials")
- 2 = Some metrics cited but many vague claims remain
- 3 = Most claims cite data but 2-3 are generic or use "approximately"
- 4 = Nearly all claims cite specific metrics, 1 minor gap
- 5 = RARE — zero vague statements, includes calculated derived metrics (e.g., "DSCR of 1.1x is only 0.1x above the minimum threshold"), compares to peer benchmarks
3. **Actionability** (1-5): Is the recommendation clear with SPECIFIC conditions and monitoring?
- 1 = No clear APPROVE/CONDITIONS/DECLINE decision
- 2 = Decision stated but no conditions or monitoring
- 3 = Decision with some conditions but missing specifics (amounts, timelines, triggers)
- 4 = Clear decision with specific conditions, monitoring could be more detailed
- 5 = RARE — decision with precise conditions (exact amounts, covenant thresholds), monitoring triggers with specific metric thresholds, and escalation procedures
4. **Audience Fit** (1-5): Is the language appropriate for a NON-TECHNICAL credit committee?
- 1 = Reads like a quant model output — jargon-heavy, no explanations
- 2 = Technical terms used without explanation in several places
- 3 = Mostly accessible, but 2-3 terms unexplained on first use
- 4 = Good plain language, all acronyms spelled out, 1 minor jargon slip
- 5 = RARE — every technical concept explained in practical terms ("the probability of default is 6.8% — meaning roughly 1 in 15 similar merchants would fail to repay"), analogies used effectively
IMPORTANT: Be honest. Most good AI outputs score 14-17/20. A score of 20/20 should be almost never given. If you find yourself giving all 5s, you are being too lenient — re-read the "RARE" criteria. There is ALWAYS something to improve.
Return your evaluation as JSON and save it as "eval-run1.md" in the "lab7-credit-risk" folder:
{"completeness": X, "grounding": X, "actionability": X, "audience_fit": X, "total": X, "strengths": "one sentence", "weaknesses": "one sentence — there is ALWAYS something to improve"}
💡 For runs 2 and 3: Start a New Session each time. Change run1.md → run2.md → run3.md and eval-run1.md → eval-run2.md → eval-run3.md in the prompt.
Record your scores
Completeness
Grounding
Actionability
Audience Fit
Total /20
Run 1
Run 2
Run 3
Average
✅ What to look for:
All 3 runs score 17-19/20: Your template is production-ready and consistent — this is the ideal outcome
All 3 runs score the same: Excellent consistency — exactly what you want for credit committee reports
Average total 14-16: Good but has room — check which criterion scored lowest and refine that part of the template
Scores vary by 3+ between runs: Needs tighter constraints — add more structure or decision rules
Want to see real score differences? Try the bonus challenge below — switch to a different model and compare
💡 This is LLM-as-Judge — you're using one AI to evaluate another AI's output. For credit narratives, this is especially valuable: you can score 50 narratives in minutes and catch quality drift before it reaches the committee.
🎯 Bonus challenge (if time permits): The 3 runs above test consistency (same model, same template). For a different test, try switching models in Kiro's model selector and generating a run-nova.md or run-haiku.md. Evaluate with the same rubric. You'll likely see score differences — a cheaper model might score 14/20 where the default scores 17/20. For credit narratives, this helps you decide: is the quality difference worth the cost difference? Maybe the cheaper model is fine for routine renewals, while the premium model handles new applications.
Reflection & Discussion
What You Built
Through 7 iterative steps, you evolved a basic prompt into a production-grade credit narrative template that:
Translates quantitative model outputs into plain language for non-technical stakeholders
Presents both optimistic and cautious perspectives for balanced decision-making
Produces structured, scannable narratives under 800 words
Includes a clear recommendation with conditions and monitoring triggers
Works across different markets, currencies, and application types
Technique Recap
Step
Technique
What It Fixed
1. Zero-Shot
Baseline
Established what "bad" looks like
2. Step-Back
Prioritize before writing
Focused on what matters most
3. Audience
Audience framing
Plain language, no unexplained jargon
4. Multi-Perspective
Bull case + bear case
Balanced, audit-ready narrative
5. Structured
Sections + length control
Consistent, scannable, concise
6. Meta-Prompt
Template extraction
Reusable at scale
7. Validation
Test with new data
Confirmed template generalizes
8. Evaluation
Rubric scoring
Measurable quality, consistency proof
💡 Key takeaway: The multi-perspective technique is especially powerful for credit decisions. Regulators expect evidence that both upside and downside were considered. A prompt template that automatically generates balanced narratives doesn't just save time — it improves the quality and auditability of every credit decision.
💾 Save Your Game — AI Memory for Long Projects
You just completed 8 steps across multiple sessions. In a real project, you'd want the AI to "remember" this work next week. But AI has no memory between sessions — every new chat starts blank.
The fix: maintain two files as your project's persistent memory:
project-status.md
Current state — what exists, what's remaining, key decisions. Load this every session. Keep it compact (~2 pages).
session-log.md
History — what was done each session and why. Load only when needed (e.g., "when did we change the threshold?").
End-of-session prompt:
Update project-status.md with what we built today:
- List all files created or modified
- Update the "What's Remaining" section
- Note any key decisions we made
Then append a summary of today's session to session-log.md.
💡 Why be specific? "Update the status" is vague — the AI might miss details or append instead of replacing. The more specific your save prompt, the better your load next time. Think of it like writing a handover note for your future self.
What You Accomplished
📝 Applied step-back prompting to prioritize risk factors before writing
👥 Used audience framing to make quantitative data accessible to non-technical stakeholders
⚖️ Built multi-perspective analysis (bull case + bear case) into every narrative
📋 Created structured, length-controlled output that's consistent across applications
🏗️ Produced a reusable template for credit risk narratives across markets and products