Prompt engineering
The task contract
Instructions, constraints, examples, decomposition, tool descriptions, refusal behavior, and the required output schema.
24 weeks · Saturday start · 1.5–3 hours/day
Become an Applied AI Engineer who can actually ship.
One execution system for three overlapping roles: strong software engineering, prompt and context engineering, reliable LLM systems, and complete AI-powered applications. Progress is saved locally in this browser.
Finding this useful? A GitHub star helps other learners discover the roadmap and lets you find it again.
Umbrella: Applied AI Engineer
Specialization: LLM Engineer
Product track: AI Application Engineer
Core discipline: Prompt + context engineering
Differentiator: Arabic + bilingual NLP
Your engineer’s adventure
Apprentice Builder
Level 1 · 0 XP
0 / 500 XP to next level
Earn 25 XP per unique completed task. Unchecking removes its XP. Your rank follows your saved progress, with no daily streak to maintain.
Start here in five minutes
Built for juniors, fresh graduates, and career switchers
You do not need prior professional AI experience. Basic computer confidence and a willingness to practise consistently are enough to begin. If you already write Python and use Git, skim the setup gate and start at the first week that you can prove—not merely recognize.
- Choose your pace. Start with Core; increase it only when your schedule proves sustainable.
- Set your starting Saturday. The page will calculate your current week.
- Complete the setup gate. Do not skip the environment and baseline checks.
- Open only the current week. Learn, build, test, and ship its evidence before moving on.
- Export progress regularly. Progress stays in this browser unless you back it up.
1.5h Core2 hours/day3h Deep
Pick up where you left off
Your progress stays on this device
Each checkbox and setting saves automatically. Saved study sessions and insights stay here too. Reopen this same page in the same browser to continue; no account needed.
Checking local storage…
A backup is a JSON file you keep on your computer. Download one after each study session. Restore it to continue in another browser or on another device; restoring replaces that browser’s progress.
Progress does not sync between devices. Clearing browser data or ending a private browsing session can remove it. A PDF is for reading; use the JSON backup to restore checkboxes and logs.
How to use the pace system
One roadmap, three workloads
Core tasks are mandatory at every pace. Standard tasks activate at 2+ hours/day. Deep tasks activate at 3 hours/day. Progress is calculated only against the work appropriate for your selected pace.
Do not delay applications until 100%. Begin light applications after Milestone 3, stronger applications after Milestone 4, and use interview feedback to adjust Weeks 17–24.
Role overlap
Why one roadmap covers all three titles
Ordered by dependency
The 24-week execution roadmap
Finish the evidence for each week before collecting more courses. Setup is a one-time gate; the six phases then build toward three portfolio projects.
Exactly what to open next
The resource study sequence
Follow the numbered steps inside the current week. The free path is complete; paid items are replacements or later deep dives, never extra mandatory homework.
Master order — do not jump ahead
Complete the setup gate first, then move through the six dependency phases from left to right.
0 · SetupEnvironment + baseline
→
1 · Weeks 1–4Python, testing, Git, API
→
2 · Weeks 5–8Data, async, security, deploy
→
3 · Weeks 9–12ML + LLM + prompt/context
→
4 · Weeks 13–16RAG context + evaluation
→
5 · Weeks 17–20Agent context, MCP, Arabic
→
6 · Weeks 21–24Operations, portfolio, jobs
Step 1
Learn one concept
Use the named primary lesson on Saturday. Stop at the listed topics instead of finishing the entire course.
Step 2
Reproduce it
On Sunday, do a small exercise without copying. Explain the result and one failure in your own words.
Step 3
Use docs on demand
Monday through Thursday, open official documentation only when the current build creates a concrete question.
Step 4
Ship the evidence
The mini-project or flagship deliverable—not course completion—is the week’s finish line.
Step 5
Deepen only after
Optional and paid resources unlock only after the core output works, is tested, and can be explained.
Core · 1.5h/day
≈ 3h learning · 5.5h building · 2h testing/review. Complete only Core sequence steps.
Standard · 2h/day
≈ 3.5h learning · 8h building · 2.5h testing/review. Add Standard sequence steps.
Deep · 3h/day
≈ 4.5h learning · 12.5h building · 4h testing/review. Add Deep work only when the core output is stable.
Course-control rule: never have more than one unfinished general course. Official documentation does not count as a course; it is a searchable tool used during implementation. Keep passive resource time below roughly one-third of the week.
Learn = primary lessonPractice = active exerciseReference = open only as neededStandard/Deep = activated by your selected pacePaid option = replace or deepen, never stack
Continuous discipline · Weeks 10–23
Open the Week 11 lab
Prompt & context engineering track
Week 11 is the focused laboratory, but the discipline continues through RAG, agents, Arabic systems, evaluation, observability, and production governance. Every checkbox below is synchronized with the matching week or flagship requirement.
Context engineering
The complete working input
Which evidence, history, memory, tool results, permissions, metadata, and examples reach the model—plus their order, trust, size, and lifecycle.
Engineering discipline
Measured, versioned, reversible
Define success, keep a baseline, change one variable, evaluate quality/latency/cost, version the result, trace it, and preserve a rollback path.
The model’s working input
Build this pipeline explicitly
A prompt is only one layer. Context selection and assembly determine what the model can actually use.
0%
0/0 active requirements
Trusted instructionsSystem rules · task contract · permissions
+
Task & user inputIntent · constraints · requested output
+
Selected contextEvidence · examples · history · memory · tool results
→
Assembler & budgetFilter · dedupe · order · label · compact · truncate
→
Validated resultSchema · citations · safety · trace · metrics
Required experiments
Change one variable at a time
Use the same representative cases and report quality, latency, token use, cost, and failure categories—not just a favorite output.
Failure taxonomy
Diagnose the correct layer
“Improve the prompt” is not a diagnosis. Classify the failure before choosing a fix.
Move-on rule: complete the dedicated Week 12 prompt/context milestone only when you can reproduce a request from versions and a context manifest, compare a candidate with a baseline, and explain whether a failure came from prompt, context selection/assembly, retrieval, model behavior, tool/state, or validation.
Continuous practice · Weeks 1–23
Suggest a workflow
AI-assisted engineering workflow
Learn to use Codex, Claude Code, specification-driven development, reusable skills, subagents, and multi-agent work without outsourcing understanding or engineering responsibility.
Tool-neutral rule: learn the workflow with one coding agent first. Codex, Claude Code, Spec Kit, BMAD, Grill Me, and Multica are examples—not prerequisites. Add a tool only when it solves a measured problem in your current process.
Progressive adoption
From supervised assistance to controlled orchestration
Each stage adds authority only after the previous stage can be reviewed and verified reliably.
Portfolio rule: disclose how AI assisted the work, but lead with the decisions, tests, failures, trade-offs, and verification that you personally own. Generated code volume is not evidence of engineering ability.
Saturday → Friday
Your repeatable weekly operating system
The daily template protects independent thinking, project output, testing, security, and reflection. Friday is intentionally lighter and can become recovery time.
Non-negotiable session close
Reserve the final 5–10 minutes to update one checkbox, log your actual time, write one growth insight, and record any gotcha. That evidence is more useful than a vague feeling of “I studied.”
When life gets busy
Protect Monday’s build block, Tuesday’s no-AI/debug block, and Wednesday’s evaluation/testing block. Cut Deep tasks first, then Standard tasks. Never compress by skipping tests and reflection.
Proof-of-skill gates
Milestones that mean something
Mark a milestone only when its evidence is inspectable. Course completion alone never satisfies a gate.
Portfolio architecture
Three flagship projects, one coherent story
Mini-projects feed these flagships. Reuse infrastructure and a thin UI; spend your time on reliability, evaluation, and engineering decisions rather than three unrelated designs.
15 controlled builds
Mini-project ladder
Each build is deliberately small, but it must include acceptance evidence. A mini-project can become a module inside a flagship instead of a separate repository.
Free first, paid optional
Curated resource library
This library is sorted by first use in the 24-week sequence. Follow the Study Order section for exact topics and stopping points; use this grid for search and later reference.
Budget rule: a free-only path is sufficient. Avoid buying multiple resources that teach the same material, and verify current pricing before purchase.
Keep your judgment
Rules for learning with AI
AI should collapse typing and lookup time—not the productive struggle that builds engineering judgment.
Progress evidence
Session, growth & gotchas logs
Log actual work rather than intended work. The dashboard totals your hours and keeps the evidence portable through export/import.