Week-by-Week Breakdown

• Set up AWS environment (account, region, IAM roles for deployment)
• Deploy auth stack — Cognito User Pool + App Client
• Deploy data stack — DynamoDB tables, S3 buckets
• Deploy API stack — API Gateway HTTP v2 + JWT authorizer + route stubs
• Deploy agent stack — SQS queue, Lambda stubs, IAM execution roles
• Set up local dev environment — configure .env.local pointing to real AWS dev account
• Confirm local login works end-to-end (Cognito → JWT → API Gateway)

• Write 5 seed estimation sheets covering: web app, mobile, system integration, data pipeline, AI feature addition
• Tag each document with metadata: project_type, tech_stack, scale, duration_weeks, total_person_hours
• Set up Bedrock Knowledge Base — S3 data source, embedding model, hybrid search enabled
• Load seed documents into Knowledge Base; confirm retrieval returns results
• Confirm RAG search returns relevant results for a test query against Bedrock Knowledge Base

Research and propose reusable workflow/skill patterns across three areas. Output is a short proposal doc per area — not code. Implementation follows in Week 7 once patterns are proven.

• Infrastructure — research patterns for AWS architecture and scaffolding (Claude Code skills, IaC generators, community prompt libraries). Propose: what artifact helps a new team go from product brief → CloudFormation stubs + local dev setup fastest?
• Backend — research patterns for Node.js/TypeScript Lambda projects (code generation, API scaffolding, test generation, handler conventions). Propose: what workflow helps a backend dev go from API spec → working tested Lambda handler fastest?
• Frontend — research patterns for Nuxt 3 / Vue component work (component generation, page scaffolding, composable patterns). Propose: what workflow helps a frontend dev go from design/spec → working page fastest?
• For each area: recommend whether the artifact should be a Claude Code skill, a workflow, a template, or a combination — with rationale
• Share proposals with Minh (infra + backend review) and Hoat (frontend review)

• Write estimation agent system prompt — structured output, citation instructions, assumptions, confidence rating
• Write RAG search sub-prompt — extracts search queries from user input
• Test prompts against Bedrock (Claude Haiku) — do not finalise on Groq alone
• Build prompt test suite: 3 inputs with expected output shape
• Seed SSM parameters: prompt text, model IDs, guardrail ID

• Define keyword blocklist — client names, internal rate cards, NDA terms
• Configure Bedrock Guardrail with keyword blocklist only
• Run test inputs through guardrail — confirm legitimate content passes
• Iterate blocklist based on false positives

• POST /sessions — create session, store in DynamoDB
• POST /sessions/{id}/run — validate input, enqueue to SQS
• GET /sessions/{id}/status — poll for result
• GET /sessions — list sessions for tenant
• Consumer Lambda — dequeue SQS → call estimation agent → write result to DynamoDB
• Estimation agent Lambda — RAG search → Bedrock invoke → return structured output
• POST /proposals — convert result to downloadable .md, store in S3, return presigned URL
• getLLMClient() factory — wraps Bedrock InvokeModel
• getVectorStoreClient() factory — wraps Bedrock Knowledge Base
• Unit tests for all handlers and lib modules
• Confirm full pipeline works against AWS dev: form input → SQS → agent → result in DynamoDB

• Login page — email/password, error handling, redirect on success
• Session list page — table of past sessions, status indicator, link to result
• New session page — input form (project name, type, tech stack, scale, features, integrations, deadline, context)
• Session result page — renders estimation output (summary, phase breakdown, role breakdown, assumptions, confidence, citations)
• Download button — fetches presigned URL, triggers .md file download
• Loading/polling state — show progress while agent runs
• "AI-suggested — requires human review" label on every result page
• Mobile-responsive layout

• End-to-end test: submit form → poll status → view result → download .md
• RAG citations appear on result page and match documents in knowledge base
• Guardrail blocks a test input containing a blocklisted keyword
• Guardrail does not block legitimate estimation content
• Multi-tenant isolation: user from tenant A cannot see sessions from tenant B
• Test with minimal input — confirm agent still produces output
• Test with detailed input — confirm assumptions list is shorter

• Fix issues found in integration testing
• Prompt iteration — improve output quality based on real Bedrock test runs
• Guardrail iteration — adjust blocklist based on false positives or misses
• Load at least 5 realistic estimation documents into Knowledge Base
• Prepare 3 demo scenarios: web portal (high confidence), AI feature addition (medium), legacy migration (low confidence)
• Run all 3 scenarios end-to-end, review outputs
• Deploy to production AWS (all stacks)
• Confirm production Bedrock + Knowledge Base works (not local Groq/ChromaDB)

• Update README.md — local dev quick start, env vars, deploy steps
• Document prompt design decisions — what was tried, what was rejected, why
• Document guardrail keyword list and rationale
• Document RAG retrieval strategy — metadata tags, hybrid search config
• Write architecture summary — patterns proven, what comes next

Implement the three workflows agreed in Week 1–2. Do not start before Week 7 — patterns must be proven end-to-end first.

• Infrastructure workflow — implement and validate against a hypothetical new product brief
• Backend workflow — implement and validate by scaffolding a sample Lambda handler outside this project
• Frontend workflow — implement and validate by scaffolding a sample Nuxt page outside this project
• For each: confirm output is actionable without manual cleanup
• Store all artifacts in .claude/commands/ or the agreed location; document how a new project adopts them