AI Prompt Pack for Computer Vision Engineers

// the problem

"Ask AI for help" produces tutorials. You need a colleague.

$ why --model-sucks

Generic prompts get generic answers

"How do I improve my mAP?" returns a blog post. These prompts feed the model your real config, your real curves, your real class distribution — and demand ranked diagnoses, discriminating experiments, and code with tests.

segfault at epoch 47

The expensive failures are silent

Train/eval transform mismatch. Labels that didn't survive augmentation. A val set you've overfit by iterating on it. There's a prompt that hunts each of these — written by someone who's been bitten by all of them.

works_on_my_gpu = True

The gap between training and production

Export parity, INT8 calibration, edge latency budgets, drift monitoring, rollback drills. Fifty of these prompts live after the model "works" — where CV projects actually succeed or die.

// inside the pack

Eight chapters. The whole pipeline.

Dataset Strategy & Collection25Dataset specs, sample-size math, long-tail sampling, synthetic-data decisions, licensing audits.

Annotation & Labeling QA25Instructions that survive annotators, agreement stats, audit sampling, label-error hunting.

Augmentation & Preprocessing25Domain-justified pipelines, label-safe transforms, train/serve parity, ablation plans.

Model Selection & Training25Architecture shortlists with honest trade-offs, transfer strategy, schedule tuning, scaling.

Evaluation & Error Analysis25Slice-based eval, failure taxonomies, threshold selection, seed variance, regression gates.

Deployment & Optimization25ONNX/TensorRT/CoreML export & parity debugging, quantization, edge budgets, canary rollouts.

MLOps & Pipelines25Data/model versioning, CI for ML, retraining triggers, incident response, cost governance.

Claude-Assisted CV Workflows25CLAUDE.md for your repo, ML-aware code review, vision-model labeling QA, agentic chores, your own prompt library.

# every prompt is a working session, not a one-liner — e.g.:
#177 — Review my training script for the silent bugs that don't crash
Act as a senior CV engineer doing a line-by-line review … Hunt specifically for the classic CV failure modes: train/eval transform mismatch, label handling ([boxes not transformed with their images, letterbox padding shifting coordinates]), loss wiring, optimizer state … For each finding: severity, the exact line, why it's wrong, the fix as a diff. Rank by expected metric impact.

Dataset strategy prompts page — Deep prompts with [PLACEHOLDER] slots for your real configs and curves

Deployment and optimization prompts page — Export parity, quantization, edge deployment — the after-training half

Claude-assisted workflows prompts page — Chapter 8: durable AI workflows, not just Q&A

CV project pre-flight checklist page — The 2-page Pre-Flight Checklist — every gate, mapped to its prompt

73-page PDF, US Letter + A4

All 200 prompts as .txt files — open, copy, paste

Works with Claude, ChatGPT & Gemini

Names real tools: COCO/YOLO, PyTorch, Ultralytics, ONNX Runtime, TensorRT, CVAT, W&B

CV Project Pre-Flight Checklist included

Buy once — yours across every job and project

One caught bug pays for this pack a hundred times over. #177 finds the train/eval transform mismatch. #185 finds the FP16 parity gap. #194 kills the doomed experiment before the GPU bill.

// faq

Questions

What exactly do I get?

A ZIP with the full 73-page pack as two PDFs (US Letter and A4) plus all 200 prompts as plain .txt files organized by chapter, delivered instantly after checkout. The PDF includes the 2-page CV Project Pre-Flight Checklist; the .txt files are for fast copy-paste into your AI of choice.

Who is this for?

Working CV/ML engineers and serious practitioners — people training detection, segmentation, or classification models and shipping them. The prompts assume you know what mAP is. If you're just starting out, they'll stretch you (chapter order = a curriculum); if you've shipped models, they'll read like a checklist of scars.

Do the prompts require Claude?

No. Every prompt works with Claude, ChatGPT, or Gemini. Chapter 8 covers Claude-specific workflows (CLAUDE.md, coding agents) because that's where agentic tooling is strongest right now, but the patterns port to any capable assistant.

Aren't AI answers unreliable for engineering work?

Unverified ones, yes. That's why these prompts are built the way they are: they demand ranked hypotheses with discriminating tests, code with unit tests, dry-run modes, and explicit assumptions you can check. The pack treats the model like a sharp junior engineer whose PRs you still review — which is exactly the posture that makes it useful.

Can I get a refund?