A civil engineer exploring how open-source and frontier AI change the way I work, create,
and solve problems. This page is an open journal of that process — the prompts I reuse,
the projects I'm building on a home server, the people I learn from, and what I'm discovering
about applying AI to engineering, writing, and everyday life. Everything here is free to copy,
remix, and improve.
Prompts Library12 prompts
ProductivityClaude
Weekly planning in 5 minutes
Act as my executive assistant. I'll paste my messy notes and unfinished tasks. Group them into: (1) must-do this week with deadlines, (2) should-do, (3) nice-to-do, (4) delegate/drop. For each must-do, suggest the first 10-minute action. Keep it short, no fluff.
Here are my notes:
<<paste here>>
ResearchGemini
Deep research briefing
I'm researching <<topic>>. Give me a structured briefing with: (1) the core problem in 3 lines, (2) 5 key facts a non-expert must know, (3) 3 credible sources with why each matters, (4) 3 counter-arguments or limitations, (5) 3 follow-up questions that would deepen my understanding. Cite everything.
ResearchGrok
Grok — real-time sentiment check
What is the current sentiment on X about <<topic>> in the last 48 hours? Break down by: dominant view, dissenting view, key accounts driving the conversation, and any misinformation circulating. Link to the most-engaged posts.
ResearchNotebookLM
NotebookLM — turn my PDFs into a study guide
Using only the sources I've uploaded, create a study guide with: (1) a one-page summary, (2) 10 flashcard-style Q&A, (3) a glossary of technical terms, (4) 5 exam-style questions with model answers, (5) a concept map showing how the ideas connect. Cite the source for every fact.
PresentationsClaude
Slide deck from messy notes
Turn the following notes into a 10-slide presentation outline. For each slide give: title, 3 bullets (max 8 words each), one image idea, and a 30-second speaker script. The audience is <<audience>>. Tone: <<tone>>.
Notes:
<<paste here>>
Video GenAny
ComfyUI — cinematic video prompt
Generate a detailed video prompt for a text-to-video model in this exact structure:
[Shot type] [Subject] [Action] [Setting] [Lighting] [Camera movement] [Style reference] [Duration].
My idea: <<describe scene>>. Make three variations: one grounded/realistic, one dreamlike, one documentary.
AutomationClaude
Automation — n8n/Zapier workflow planner
I want to automate <<task>>. Design a workflow with: (1) trigger, (2) each step including which tool/API, (3) branching logic, (4) error handling, (5) a test plan. Output as a numbered list plus a simple ASCII flow diagram.
CreativityClaude
Creative — story seed generator
Give me 5 story seeds combining these ingredients: <<theme>>, <<setting>>, <<constraint>>. Each seed should be 2 sentences, have an implicit conflict, and feel unlike anything I've read. Avoid tropes.
Prompt WritingClaude
Meta — improve this prompt
You are a prompt engineer. Here is my prompt:
<<paste prompt>>
Rewrite it to be: (1) clearer on role and goal, (2) structured with explicit sections, (3) with concrete examples, (4) with an explicit output format. Then explain in 3 bullets what you changed and why.
ProductivityClaude
Engineering drawing reader
I'll upload a civil engineering drawing. Extract: (1) drawing title and number, (2) scale and units, (3) all dimensions and annotations in a table, (4) key elements (pipes, manholes, grades), (5) any notes or legends. If any part is unclear, say 'unclear' rather than guess.
ProductivityAny
Excel formula explainer & builder
I need an Excel formula that does the following: <<describe goal in plain English>>. Give me: (1) the formula, (2) a step-by-step explanation of each part, (3) an example with sample data, (4) common errors and how to fix them, (5) an alternative using a different function.
ResearchAny
Local LLM evaluator
I'm comparing local LLMs (Llama, Mistral, Qwen, DeepSeek, Gemma) for <<use case>> on a machine with <<specs>>. Rank them by: speed on my hardware, quality for my task, VRAM needed, license. Flag anything that requires >8GB VRAM. End with a one-line recommendation.
Projects in Flight6 projects
Active
Free LLM video generation with ComfyUI
Building a fully local text-to-video pipeline using ComfyUI + open-source video models (Wan, HunyuanVideo, LTX-Video). Zero subscription cost.
Hardware: home server with GPU. Exploring which open video models give the best quality per second of render. Next step: build reusable workflow JSONs for common shot types. Target: produce a 10-second clip entirely offline.
Running Llama, Qwen, DeepSeek, Gemma locally for privacy-sensitive work. Nothing leaves the home network.
Using Ollama as the runtime. Comparing models on: summarization of private docs, code assistance, and reading civil-engineering PDFs. Integrating with Open WebUI for a ChatGPT-style interface accessible only inside the LAN.
Testing how well vision-capable LLMs extract dimensions, annotations, and key elements from civil engineering drawings.
Current tests: uploading stormwater plan PDFs to Claude and Gemini, comparing extraction accuracy. Interested in combining OCR preprocessing + multimodal LLM for more reliable results. Long-term: automate QA checks on drawings.
Updated 2026-04-10
Exploring
3D animation in Blender with AI assist
Using AI to speed up Blender workflow: generating textures, auto-rigging, prompt-driven scene setup.
Investigating AI plugins for Blender (Dream Textures, AI-Render). Goal: produce architectural and civil visualisation clips faster. Next: combine with ComfyUI pipeline for hybrid 3D + video-model renders.
Building Model Context Protocol servers that let Claude drive HEC-RAS, AutoCAD, and other civil engineering software.
Early experiments: MCP server exposing HEC-RAS project files so Claude can read scenario results, compare runs, and draft flood reports. Also looking at MCP for reading/writing Excel hydrology calcs.
Using Copilot, Claude, and Gemini to speed up daily Excel formulas, Word report drafting, and document QA.
Prompts library (see Prompts section) is the backbone. Currently templating: stormwater design reports, cost estimate sheets, and site investigation summaries. Measuring time saved per document.
Updated 2026-04-16
Testing This Week
Claude Skills
Testing Anthropic's Skills feature for automating pptx/docx/xlsx creation.
Evaluating
Wan 2.1 on ComfyUI
Comparing Wan video model quality vs HunyuanVideo on the same prompts.
Evaluating
Open WebUI + Ollama
Self-hosted ChatGPT-style UI for local models. Running on the NextcloudPi.
Spec for giving LLMs standardised access to tools and data sources. Growing fast.
Anthropic
AI in Professional Life
Drafting stormwater design reports
Before
Writing a 20-page stormwater management report from scratch took 2–3 days. Much of it was reformatting calculations and rewriting boilerplate.
Prompt
Using the calculations I'll paste below and the project description, draft a stormwater report following MECP guidelines. Include: executive summary, site description, design criteria, calculations narrative, and recommendations. Leave <<INSERT>> tags where I need to verify manually.
After
First drafts now take 3–4 hours. I review and edit rather than write from zero. All numbers are still manually checked.
Lesson
AI is a drafting tool, not a replacement for engineering judgment. Always verify every number and regulatory citation.
Quick formulas for hydrology spreadsheets
Before
Googling Excel syntax mid-calculation broke focus. Simple things like 'lookup nearest rainfall intensity for a given duration' could take 20 minutes to get right.
Prompt
I have a table with storm durations in column A and intensities in column B. I need a formula that takes a duration and returns the closest intensity by linear interpolation. Explain the formula step by step so I can adapt it.
After
Near-instant formula generation with an explanation I can verify. My hydrology spreadsheets got cleaner because I finally use INDEX/MATCH and LET properly.
Lesson
Treat AI as a pair programmer for formulas. Ask it to explain the output so you learn, not just copy.
Weekly AI Log
2026-04-18
Launched ai.rajukhadka.ca
Spun up this page on the home server as a learning journal. Goal: make AI learning visible. Static JSON for now, admin panel for future updates.
2026-04-12
First HEC-RAS + MCP experiment
Sketched a plan for an MCP server that exposes HEC-RAS scenario outputs. Claude could then compare pre/post flood results and draft the discussion section of a report. Early days.