Engineering

Prompt Engineer

Also hired as: Context Engineer · AI Interaction Designer · LLM Engineer (Prompting)

The 2023 gold-rush title grew up: fewer dedicated roles, but the skill now lives inside every AI job.

US salary · 2026
$120k – $250k
Dedicated roles cluster mid-band; lab model-behavior roles pay more. Where prompting folds into a broader engineering job, expect standard SWE bands instead. Rough 2026 estimates.
Typical background
Writers and linguists who learned to ship code, product engineers who got obsessed with model behavior, and support or ops people who became their company's unofficial AI person.
This roadmap
5 stages · 24 nodes
5 proof-of-work checkpoints
What is a Prompt Engineer?

A prompt engineer designs the instructions, examples, and context that make a model behave — reliably, in production, at scale. The craft covers system prompts, few-shot example curation, output contracts, and increasingly the whole context window: what gets retrieved, what gets summarized, what order it arrives in, and what it costs in tokens. In 2026 the honest job description is closer to 'context engineer' — the words still matter, but the architecture of everything around the words matters more, and the best practitioners treat both as engineering.

Here is the uncomfortable truth: the standalone prompt-engineer role from the 2023 headlines — the postings that advertised $300k+ for prompting alone — mostly dissolved. Companies discovered prompting is a skill every product engineer and PM needs, not a separate headcount, so the junior prompt-whisperer listings disappeared. What survived is real but narrower: model-behavior work at labs, AI content systems running thousands of prompts in production, and prompt-heavy products where one specialist's improvements compound across millions of requests. Fewer roles, higher bar, better work.

What keeps the role distinct is ownership of the model-facing layer. An agent engineer owns the whole system — tools, state, orchestration. You own the part where intent becomes instruction: the prompts, the context assembly, the eval sets that prove a change helped, and the migration plan when the next model version shifts behavior underneath you. Treat prompting as engineering — versioned, tested, measured — and you will stay valuable whether or not your business card ever says prompt engineer.

What you'll actually do
  • Design and version the system prompts behind product features, with a changelog and rollback plan.
  • Curate few-shot example sets from real failure cases, not hypothetical ones.
  • Own context assembly: what enters the window, in what order, at what token budget.
  • Run evals on prompt variants and regression-test the library before every model upgrade.
  • Write model-behavior specs — tone, refusals, edge cases — that engineering can implement.
  • Harden prompts against injection and off-policy behavior in collaboration with security.
  • Migrate prompt libraries across model versions and providers without quality cliffs.
  • Teach the org: patterns, anti-patterns, and internal docs that raise everyone's baseline.
This role fits you if
  • You debug by reading transcripts, and you genuinely enjoy it.
  • You will rewrite one paragraph five times to remove a single ambiguity.
  • You think in edge cases: empty input, hostile input, weird Unicode, the almost-match.
  • You want to work with models every day without wanting to train them.
  • You can explain to an executive why the model failed — without jargon and without lying.
The toolbox
Claude / OpenAI / Gemini APIsSystem prompt designFew-shot curationStructured outputs (JSON / XML)Context engineeringRAG basicsPrompt versioning in gitGolden setsLLM-as-judgePromptfoo / BraintrustToken accountingInjection defense
The roadmap — 5 stages, 24 nodes
Stage 1

Prompting Fundamentals

Weeks 1–3

Learn what the model actually is, then learn the anatomy of a prompt that survives production traffic.

How LLMs work — the honest versionCore
Tokens, context windows, temperature, and why the model has no memory you didn't give it. Every prompting technique is downstream of this mental model.
Anatomy of a production promptCore
Role and goal, hard constraints, delimited input, output contract with an example. Write ten prompts using the full anatomy until the structure is automatic.
Few-shot engineeringCore
Two or three well-chosen examples beat paragraphs of instruction. Learn to pick examples that encode your hard cases: the ambiguous input, the empty input, the near-miss.
Checkpoint: publish a prompt library✓ Checkpoint
Ten production-grade prompts across extraction, classification, rewriting, and structured output — each with delimited inputs, an output contract, three test cases, and notes on the failure modes you designed against.
Stage 2

Context Engineering

Weeks 4–8

The window is a budget. Learn to control what fills it — retrieval, memory, structure — because that is the modern job.

Context as a budgetCore
Order effects, lost-in-the-middle, and what happens when retrieval floods the window. Learn to ration tokens the way a performance engineer rations milliseconds.
Structured outputs and schemasCore
JSON mode, XML tags, and tool schemas as output contracts. Build one pipeline where a downstream program consumes model output with zero cleanup.
Retrieval as context assemblyCore
RAG is prompting at scale: chunking, ranking, and deciding what earns a place in the window. Build a small pipeline so you know where relevance actually breaks.
Memory patternsElective
Conversation summarization, persistent instruction files, and when to store vs re-derive. Claude Code's CLAUDE.md is a working case study in instruction memory.
Checkpoint: ship a context-engineered feature✓ Checkpoint
An app where you control retrieval, assembly order, and token budget — with before/after transcripts demonstrating a failure the naive version makes and yours doesn't.
Stage 3

Evaluation and Iteration

Months 2–4

Vibes don't scale. Prompts you can't measure are prompts you can't defend in a model migration.

Why 'it looks better' failsCore
Nondeterminism means single examples prove nothing. Build a golden set from real traffic and score against it — the habit that separates professionals from prompt hobbyists.
LLM-as-judge for prompt comparisonCore
Use a judge model with a written rubric to compare prompt variants at scale, and spot-check it against your own labels to catch judge bias.
Prompts as codeCore
Prompts live in git, changes go through review, and CI runs the golden set. Write a model-upgrade playbook: what you test before switching versions, and what triggers a rollback.
Red-team your own promptsElective
Attack your prompts with injection attempts and instruction overrides before someone else does. Defense starts with delimiting untrusted input.
Checkpoint: an eval-backed improvement✓ Checkpoint
Publish a repo with a 100+ case golden set, two prompt versions, judge scores with human spot-checks, and a writeup naming exactly which changes moved which failure categories.
Stage 4

Specialize or Integrate

Months 3–5

Pure prompting roles are scarce; prompting plus a surface is employable. Pick your surface.

Pick a surface and go deepCore
AI content systems, support automation, coding assistants, or model behavior. Each has its own failure modes and its own hiring pipeline — depth in one beats familiarity with all.
Prompting for agentsCore
System prompts for tool-using agents are a different discipline: when to act, when to ask, how to describe tools. Study how production agents are instructed.
Safety behavior and injection defenseCore
Understand the threat surface well enough to write refusal specs and layered defenses. Model-behavior roles are the best-paid corner of this field, and this is their core skill.
Know when to stop promptingElective
Some behaviors need fine-tuning, not longer prompts. Learn the decision boundary so you can recommend the right tool instead of defending your hammer.
Checkpoint: a spec someone adopted✓ Checkpoint
Produce a model-behavior spec and prompt suite for one real surface — support bot, content system, coding assistant — and get it adopted in a shipped project, an OSS repo, or your team's product.
Stage 5

Get Hired (or Get Absorbed on Purpose)

Months 5–7

Two winning paths: a dedicated role at a prompt-heavy company, or the AI-strongest engineer on a product team. Prepare for both.

Package the portfolioCore
Prompt library, eval repos, the adopted spec. Interviews here are often live: debugging a transcript or improving a prompt on the spot, so rehearse narrating your process.
Target the real openingsCore
Search for model behavior, AI content, LLM engineer, and conversation design — not just 'prompt engineer'. The role hides under a dozen titles.
Close the code gapCore
If you came from writing: Python, API calls, and git are the price of admission. If you came from engineering: ship one polished AI feature end to end to prove product sense.
Teach in publicElective
Prompt-failure breakdowns and migration postmortems build inbound interest. This field is young enough that one good post can make you the visible expert on a niche.
Checkpoint: convert applications to live interviews✓ Checkpoint
Run ten targeted applications with the portfolio attached. Land at least two interviews that include a live prompting or transcript-debugging exercise, and fold the feedback into the portfolio within a week.
Build your portfolio

Nobody hires a Prompt Engineer off a certificate. They hire off proof. Ship these and put them where people can click them:

01
A production-grade prompt library
Ten prompts across extraction, classification, rewriting, and structured output — each with delimited inputs, an output contract, three test cases, and notes on the failure modes you designed against.
Proves: You treat prompts as engineered artifacts with specs and tests, not incantations.
02
A context-engineered feature
An app where you control retrieval, assembly order, and token budget — shipped with before/after transcripts of a failure the naive version makes and yours doesn't.
Proves: You own the whole context window, not just the instruction text — the modern version of the job.
03
An eval-backed prompt improvement
Two prompt versions run against a 100+ case golden set, judge scores spot-checked by hand, and a writeup naming which change moved which failure category.
Proves: You can demonstrate a delta with numbers — the line between professionals and prompt hobbyists.
04
An injection-hardened prompt suite
A prompt system attacked with your own injection set — instruction overrides, hostile retrieved content — with the delimiting and defense layers that held, documented case by case.
Proves: You think about the adversarial case before production forces you to.
05
A model-behavior spec someone adopted
A behavior spec and prompt suite for one real surface — support bot, content system, coding assistant — adopted by a shipped product or OSS repo, with migration notes from one model version to the next.
Proves: Your work survives contact with real users and other people's codebases — adoption is the credential.
Position your profile
Headline formula

I make [support bots / AI content systems / coding assistants] behave in production: prompt and context systems with measured deltas — [X]% to [Y]% on a [N]-case golden set.

Resume bullets to earn
  • Shipped the versioned prompt library behind [feature]; cut format-break failures from [X]% to [Y]% on a [N]-case golden set.
  • Rebuilt context assembly — retrieval, ordering, token budget — for [product], cutting cost per request [X]% while judge-scored quality rose [Y] points.
  • Ran the eval-gated migration from [model A] to [model B] across [N] production prompts with zero quality regressions.
  • Built the injection test set for [surface]; [N] attack cases now run in CI and block bad deploys.
  • Wrote the model-behavior spec — tone, refusals, [N] edge cases — adopted by [team or repo] as the source of truth.
Where to be visible
  • A public prompt-library repo with test cases attached — most prompting content is screenshots, so a repo with evals instantly outranks it.
  • Prompt-failure breakdowns and migration postmortems: the transcript, the diagnosis, the fix, the numbers. This genre gets shared; tip threads don't.
  • Contribute prompts and eval cases to OSS projects with LLM features — a merged spec is public proof of the exact skill.
  • The practitioner corners: Promptfoo and Braintrust communities, provider-docs discussions, model-behavior threads where people compare real notes.
  • One definitive deep-dive per surface you claim — the post on injection-proofing support bots, or on surviving a model migration. Niche authority compounds.
What interviews actually test
  • Live prompt surgery: improve a weak prompt on the spot while narrating. Rehearse thinking aloud through the full anatomy — role, constraints, delimiting, output contract.
  • Transcript debugging: here's a bad output, explain it. Name the failure mode, locate the cause, propose the fix — your error-analysis reps are the entire prep.
  • Eval literacy: how would you prove version B beats version A? Golden sets, judge validation, and variance handling should be reflexes, not vocabulary.
  • Scenario design: a system prompt for a support bot or content pipeline, including refusals, edge cases, and injection defense — they're testing whether you design past the happy path.
  • The portfolio walkthrough: lead with before/after numbers and let the deltas talk. 'Measured' is the word that separates you from the field.
Who's hiring
Anthropic (model behavior)OpenAINotionIntercomDuolingoHarveyJasperAI content platformsEnterprise AI teams
Free courses that map to this role
Related roles
FAQ

Is prompt engineering dead?

As a standalone junior job title, mostly yes — those postings were absorbed into product, engineering, and PM roles. As a skill, it is more valuable than ever, and a smaller set of dedicated senior roles (model behavior, AI content systems) survived and pay well. The skill outlived the gold-rush title.

Do prompt engineers still make $300k?

The famous 2023 postings at that level were outliers, and most were at labs. In 2026, dedicated roles typically land in normal product-engineering bands, with lab model-behavior roles at the top. High comp now comes from prompting plus something: evals, engineering, or a surface you own.

Do I need to know how to code?

Increasingly, yes. Context engineering is programming — retrieval, token budgets, structured outputs, CI on golden sets. You do not need to be a strong software engineer, but Python, API calls, and git are the floor for the roles worth having.

What's the difference between a prompt engineer and a context engineer?

Same craft, wider scope. Context engineering covers everything that enters the model's window — retrieval, memory, summarization, budgets — not just the instruction text. The industry largely renamed the job as it matured; this roadmap teaches the wider version.

Will better models make prompting obsolete?

Better models need fewer tricks but more specification. The work is shifting from incantations to specs: precisely defining behavior, encoding it in prompts and examples, and proving it with evals. That work grows with model capability rather than shrinking.

Start building, today

Every stage above maps to free lessons on this site. No signup, no paywall — open the first course and ship your first checkpoint this week.

Browse the courses →