Career roadmap
Prompt Engineer
Design, test, and maintain the instructions that drive a model — and prove the changes worked with evaluations.
Before you start Prompt Engineer
- Precise written English
- Systematic thinking: change one variable at a time
- Spreadsheets
- Enough Python or notebook comfort to run a script
How models behave
Understand the machine you are instructing.
The unit of cost and the ceiling on input, in one concept.
Ch 77 — LLM Fundamentals- Tokenisation quirks
- Context window budgeting
- Truncation behaviour
Know what makes output vary before you blame the prompt.
- Temperature and top-p
- Repeated runs on one prompt
- Seeds and why they are not enough
A prompt is a per-model artefact. Scaffolding written for an older model over-applies on a newer one.
- Capability tiers
- Instruction-following differences
- Refusal behaviour
- Re-baselining after an upgrade
BuildA side-by-side comparison of one task across three models, with notes on where each fails.
Prompt patterns
Move from lucky prompts to reusable structure.
The four-part skeleton behind almost every production prompt.
Ch 78 — Prompt Engineering- Ordering for cacheability
- Constraints over pleading
- Fallback instruction for missing data
Show the format you cannot describe. Vary the examples deliberately.
- Example selection
- Coverage of edge cases
- Cost of examples in every call
Two cheap focused calls usually beat one clever prompt.
- Splitting classify-then-write
- Routing with a cheap model first
- Where chains add failure modes
If code will parse it, constrain it with a schema.
- JSON schema and enums
- Validation as a normal error path
- Why prefill tricks are gone
Tag the data so instructions inside it read as content, not commands.
- XML tags
- Explicit 'this is untrusted' framing
- Escaping user content
BuildA prompt library of ten task templates, each with inputs, output schema, and known failure cases.
Evaluation
This is the whole job. Opinions about prompts are worthless without measurement.
Write scoring rules a second person can apply and reach your answer.
- Binary criteria over 1-5 vibes
- Inter-rater agreement
- Rubric drift over time
50-200 frozen cases, including every failure you have personally seen.
Ch 85 — Evaluation & Hallucination- Sourcing real cases
- Covering edge cases
- Keeping the set frozen for comparability
Humans and models both judge A-vs-B far more reliably than absolute scores.
- Blind comparison
- Position bias
- Win rate and ties
Scale scoring with a model, then audit the judge against human labels.
- Judge prompt design
- Spot-check sampling
- Known judge biases
A 2-point win on 40 cases is nothing. Know when to believe a result.
- Sample size intuition
- Repeat runs and variance
- Reporting uncertainty honestly
BuildAn eval report that recommends one prompt over another and shows why the difference is not noise.
Prompts inside systems
Prompts in production are code: versioned, cached, and budgeted.
Every prompt has an id and a version, and both appear in the logs.
- Prompt as a stored artefact
- Attributing a regression to an edit
- Rollback procedure
Order the prompt so the stable part caches and the volatile part comes last.
Ch 87 — Cost & Latency Optimization- Prefix stability
- Verifying cache reads in usage
- Trimming dead instructions
- Output token caps
In a RAG app, the retrieved context is most of the prompt. Its order and framing matter.
Ch 80 — RAG Pipeline- Chunk ordering
- Deduplicating context
- Token budget per section
A tool description is a prompt. Vague ones cause the wrong tool to fire.
Ch 81 — Tool / Function Calling- Naming and when-to-use guidance
- Parameter descriptions
- Testing tool selection
BuildA versioned prompt config in a real repo, with cost per call before and after your trimming.
Safety & handoff
Break your own prompts before a user does, then hand them over cleanly.
Assume any text from outside your system is trying to give orders.
Ch 86 — AI Security- Direct and indirect injection
- Data/instruction separation
- Output validation before action
Run a fixed attack suite on every prompt change, like a test suite.
- Common attack shapes
- Regression suite of attacks
- Reporting severity honestly
Prompts get logged. Decide what may never be in them.
- Redaction before send and before log
- Retention windows
- Least-context principle
A prompt only you can maintain is a single point of failure.
- Documenting intent and failure cases
- Eval instructions for the next owner
- Ownership and review cadence
BuildA red-team report against your own app: attacks tried, which worked, the fix for each.
Prompt Engineer tools on your CV
- Model playgrounds / Anthropic Console
- A notebook or script runner
- Git
- Promptfoo or a custom harness
- Langfuse
- Spreadsheets
What Prompt Engineer employers ask to see
- A public eval report with a recommendation and a noise analysis
- A prompt library with documented failure cases
- A red-team write-up
Read this honestly: the standalone title is shrinking. The work is being absorbed into AI engineer, applied AI, solutions engineer, and AI QA roles. Learn it as a specialism inside one of those, not as your only skill — it is still the most accessible on-ramp into AI work.
Content last reviewed 2026-08-31. Guidance only — no institute or paid placement is endorsed anywhere in this book.