The Prompting Company is building a staff from scratch
A year-old San Francisco startup has five roles live on its Y Combinator job board, and every one is labeled "Founding" and based in San Francisco — a clue that the founders think the next layer of the AI economy is still being staffed. The company was founded in 2025 by Kevin Chandra, Michelle Marcelline, and Albert Purnama, and its pitch is blunt: "We help companies get mentioned in ChatGPT by rebuilding the internet for AI agents." The product analyzes and creates content shaped for AI assistants so brands get cited inside answers from ChatGPT, Gemini, and Perplexity — a roster that already includes NVIDIA, Rho, and Exa. The framing in the postings is unvarnished about the shift the company is selling: "You optimized for users experience. Now optimize for Agent Experience."
The five roles split into two pay clusters:
| Cluster | Roles | Salary | Equity |
|---|---|---|---|
| Engineering (3 seats) | Founding Forward Deployed Engineer; Fullstack Engineer — Go, TypeScript, and Next.js; Founding Forward Deployed Engineer (deployment) | $100,000–$150,000 | 0.50% |
| Go-to-market & Operations (2 seats) | Founding Ops; Founding GTM; Founding Account Executive | $80,000–$120,000 | 0.50% (Account Executive: 0.10%–0.50%) |
The spread between the two clusters ($20,000 at the top, $40,000 at the bottom) signals how seriously a seed-stage startup treats its non-engineering hires.
The Prompting Company closed a $6.5M seed round in October 2025, with Peak XV Partners, Base10, Y Combinator (S25 batch), Kearny Jackson, Firedrop, and Logan Kilpatrick on the cap table. The team is listed at 7, so each new hire roughly doubles headcount. The engineering work gets research-flavored language: new hires will "design experiments and systems that uncover how AI agents crawl, retrieve, and cite the web" — the kind of framing you usually see at a lab, not a 10-person startup with paying enterprise customers.
The wider market is hard to ignore. LinkedIn lists more than 10,000 AI prompt engineer openings in the United States; Indeed shows 7,726 "AI Prompt Engineer" roles and 1,474 "AI Prompting Engineering" roles. Major model labs are staffing up too: OpenAI posted an Applied AI Engineer role on its Codex Core Agent team, Anthropic listed an Applied AI Engineer for Enterprise Tech, Perplexity opened a Member of Technical Staff position on Model Behavior, and Adobe posted a Senior Prompt Engineer on its Brand Concierge product. Even Bain & Company has started hiring Lead AI Engineering roles.
What makes The Prompting Company's five open seats unusual is the concentration. Adjacent startups scatter one or two "prompt engineer" or "AI engineer" postings across 200-person org charts. The Prompting Company is building an entire founding bench around a single bet: the next hiring war won't be over who can train models, but over who can make products visible to them.
What the screening process actually tests
The Prompting Company's screen asks for something more specific than "good prompting instincts." Reviewers want evidence that a candidate treats prompt design as an engineering discipline. Interview Query's 2026 prep guide frames the assessment as the inverse of a LeetCode problem: the deliverable is a text artifact (an instruction) and the score comes from how reliably that instruction produces the right output when run against a large language model.
Most prompt engineering screens blend three components, per The Interview Guys' January 2026 interview-prep breakdown: technical knowledge about how LLMs work, a hands-on prompting exercise, and behavioral questions about real-world complexity. Some employers add a take-home piece where the applicant submits a full prompt system with documentation. For senior and lead roles, Greenroom's June 2026 guide notes the screen increasingly includes a design round structurally similar to a system-design interview, except the "system" being designed is the prompt and model layer.
The hands-on segment is where most candidates separate from the pack. Almost every prompt engineering interview includes a live exercise: write a prompt for a stated task, debug a prompt that's misbehaving, or iterate on a prompt given new failure examples. Interview Query documented an assessment at a data storage startup where a candidate received 23 prompts back-to-back, each one a JSON-to-JSON transformation: convert {"user_id": "u_482", "scores": [84, 91, 78], "status": "active"} into {"id": "u_482", "average_score": 84.33, "is_active": true}. The prompts were chained, so the output of question 1 fed directly into question 2. A weak early prompt didn't just cost one answer; it degraded every subsequent one. Greenroom warns that freezing during this kind of exercise is the most common failure mode, and recommends narrating observations in real time: "I'm noticing the model is being too verbose here, so I'm going to add a length constraint" shows analytical thinking even if the first prompt isn't perfect.
Reasoning gets graded at least as heavily as the final prompt text. Greenroom reports that reviewers weight clarifying questions, named edge cases, and explained tradeoffs as much as the artifact itself. The Interview Guys flags the same pattern: candidates who can define chain-of-thought prompting but can't explain when they wouldn't use it lose points. The "why not" carries as much weight as the "why."
Six core dimensions show up repeatedly across screening rubrics, per Ace My Interviews' February 2026 analysis: methodological rigor, technique breadth and depth, evaluation sophistication, safety awareness, production thinking, and communication. The first of those is the most decisive: anyone can write a decent prompt, but a dedicated prompt engineer is expected to measure quality and iterate systematically. Vague methodology is an explicit disqualifier. Ace My Interviews warns that treating prompting as an art rather than an engineering discipline is the most common failure mode. Saying you "categorize failures" signals the maturity reviewers want.
Production thinking is the dimension most often missed. Screens probe whether candidates understand that instructions are a probabilistic nudge, not a guarantee, and that production systems need a deterministic backstop. Greenroom's reviewers specifically look for candidates who can write even a short defensive-parsing function when asked, since shipping prompts to production requires that kind of code. They also want to see cost awareness: an extra 500 tokens of instructions on a prompt called a million times a day is a real line item, not a rounding error.
What a portfolio that clears the screen looks like
Candidates who make it past The Prompting Company's screen share a habit: they bring receipts. Per the PE Collective analysis of hiring managers across the AI industry, applicants with strong portfolios land interviews at three times the rate of those without one — and the portfolios that work share a recognizable shape.
Depth beats breadth. Three to five solid projects outgun a dozen shallow ones. PE Collective is blunt: a project that amounts to "I wrote a prompt and it worked" is not portfolio-worthy. GPT Prompts spells out the range companies actually hire for: a text classifier, a retrieval-augmented question-answer system, a structured-extraction tool, and a content-generation task with guardrails. A code-review prompt chain that caught 78% of common Python bugs, or an LLM evaluation harness tested across four different models, signals production-grade thinking — not tutorial regurgitation.
Metrics separate contenders from the rest. Vague claims collapse under review; specific numbers survive. The sentence that wins interviews, per GPT Prompts, sounds like this: "The baseline prompt scored 71% exact-match on 50 labeled tickets; after adding few-shot examples and a stricter output schema it scored 89%, with the remaining errors concentrated in ambiguous multi-issue tickets." That one line bundles a test set, a measured delta, and an honest error analysis — the three things reviewers in 2026 actively scan for. PE Collective's framing is the same: "Saying 'the prompt performed well' means nothing. Saying 'the prompt achieved 94% accuracy across 150 test cases, up from 71% in v1' is specific and credible."
The process matters as much as the outcome. Hiring managers are buying how a candidate thinks, not just what they shipped. PE Collective notes that a polished final prompt reveals nothing about problem-solving; five iterations with clear reasoning reveal everything. Reviewers ask: can this person take a vague requirement, turn it into a testable specification, measure model behavior against it, and improve it deliberately? A 2,000-token prompt trimmed to 400 tokens while preserving output quality shows exactly that mindset.
The format is part of the signal. A version-controlled GitHub repository of prompts, evaluation scripts, and API integration reads as developer-grade; a PDF of static outputs reads as hobbyist. GPT Prompts warns that "pure prompt text without surrounding code reads as hobbyist in 2026, because production roles expect API integration, evaluation scripts, and version control." Buried-in-a-PDF resumes fail the click-through test before the work is even seen.
Reasoning, not credentials, closes the gap. PE Collective draws a sharp line: certifications tell a hiring manager a course was completed, but a portfolio proves the work can actually be done. When two candidates are comparable on paper, the one with three well-documented projects beats the one with three AI certifications "almost every time." Refonte Learning adds that a clear portfolio can educate hiring managers who don't yet fully understand prompt engineering, and the artifact itself teaches the reviewer how to value the candidate.
PE Collective catalogs three patterns from actual placements: a career switcher who landed a $95,000 offer, a developer who closed at $140,000, and a domain expert at $160,000. The differentiator across all three: production-quality projects paired with the reasoning chain to explain them. Refonte Learning's advice (that what matters is showcasing creativity, problem-solving, and progress) lands hardest for career-switchers who lack shipped product experience. A simple classifier with documented iterations can outperform a flashy idea that isn't clearly demonstrated.
Why the bigger labs have walked away from the title
The Prompting Company's decision to staff up on prompting talent arrives against a backdrop in which most of the industry's biggest names have moved in the opposite direction. Anthropic, OpenAI, and Google DeepMind no longer publish the "Prompt Engineer" title on their careers pages at all. What they hire for now (AI Engineer, LLM Application Developer, AI Research Engineer) treats prompting as a single line item in a much longer skills list, the kind of role where a candidate's value is measured by their ability to ship a retrieval-augmented generation pipeline rather than to rewrite a clever instruction.
The compensation gap mirrors the rebrand. As of March 2026, the median base salary for an AI Engineer role sits at roughly $154,000 on Levels.fyi and $141,172 on Glassdoor. By comparison, the standalone "Prompt Engineer" title averages $128,625 on Glassdoor, based on just 29 reported salaries — a thin dataset that itself signals how rare the role has become at the firms setting industry pay. Anthropic's $335,000 "Prompt Engineer and Librarian" listing from 2023 still circulates online as a cautionary tale: it was widely cited and almost no one was actually hired into it.
The talent migration has reshaped what competitors ask for in the initial screen. LinkedIn data shows a 40% drop in profiles carrying the "Prompt Engineer" title between mid-2024 and early 2025, while profiles advertising "AI workflow design" climbed 25% over the same period. A 2025 Indeed Hiring Lab analysis by economist Allison Shrivastava found that listings explicitly titled "Prompt Engineer" peaked at 0.3% of AI-related postings in 2024 and have since retreated toward zero. ZipRecruiter's chief economist Julia Pollak pointed to a wave of AI hiring outside tech (in retail, finance, healthcare, and education) where the prompt title never took hold at all.
The hiring criteria themselves have hardened around a different stack. LinkedIn's Jobs on the Rise 2026 report ranks AI Engineer as the single fastest-growing role in the United States, with the top listed skills being LangChain, retrieval-augmented generation, and PyTorch. Microsoft's chief marketing officer of AI at Work, Jared Spataro, told The Wall Street Journal that companies no longer need dedicated prompt specialists because generative AI can, essentially, prompt itself, asking follow-up questions and requesting feedback without a human intermediary. Open-source tooling has reinforced that shift: Omar Khattab's DSPy framework, developed at Stanford NLP and now one of the fastest-growing AI repositories on GitHub, treats prompts as programs to optimize rather than text to handcraft, and published ICLR 2024 benchmarks showed accuracy gains of up to 49 percentage points on math reasoning tasks by doing so.
Salaries at the very top of the market have moved in step. Meta, Netflix, and Amazon have all hired AI builders and trainers at packages reported as high as $900,000, well above anything carried by a standalone prompt title. Even roles at a fintech like Stripe show the same pattern: a Machine Learning Engineer opening in South San Francisco posted at $212,000–$318,000, roughly two and a half times the median for a "Prompt Engineer" posting anywhere in the country.
The pressure on candidates is concrete. A standalone "Certified Prompt Engineer" credential on a resume now risks signaling that an applicant stopped at the layer senior AI hiring managers consider basic. What firms say they want instead, in job descriptions and recruiter conversations, is someone who has built a working RAG application, designed an automated evaluation framework using RAGAS or an LLM-as-judge pattern, and shipped an agentic workflow in LangGraph.
That is the field The Prompting Company's five open roles now have to clear — a labor market that has collectively decided prompting is a competency, not a career.
What recruiters say actually works
The recruiters and AI-hiring specialists who have published on this space in 2025 give remarkably aligned advice — and the through-line is that candidates who treat prompt engineering as a craft, not a checkbox, clear screens faster. Manatal's October 2025 guide, the ixceed-solutions and biometrictalent recruiter handbooks on prompt engineering with ChatGPT-4o, and Wonsulting's November 2025 job-seeker playbook all converge on a small number of high-leverage moves.
First, demonstrate that you can scope a prompt before you ever write one. The ixceed recruiter playbook tells talent teams to "define who your audience is and what you need," "provide context," "define the format," and "iterate." The same four disciplines show up in Wonsulting's November 2025 advice as "clarity is king," "provide context — give the AI a role," and "use constraints." Candidates who walk into a screen with examples that show role-conditioning, format constraints, and a visible revision loop signal competence before the interviewer asks a question.
Second, show range, not just text. The Manatal guide frames prompt engineering as "multidisciplinary," requiring "a combination of technical, linguistic, and creative abilities." Hiring managers want evidence that you can direct LLMs to produce structured data, code, evaluation rubrics, and images. A portfolio that only contains marketing copy looks narrow. A portfolio that includes a Boolean-search generator, an interview scorecard template, and a structured evaluation rubric for an LLM looks like someone who has internalized what large language models can actually do.
Third, quantify outcomes. Recruiters themselves are being told in 2025 to "save hours on sourcing and admin tasks," "improve candidate quality through consistent, structured processes," and "personalize outreach at scale" — but those benefits only land when a candidate can show that their prompts moved a number. Advice from Farah Sharghi generalizes the rule: every story should follow the Google XYZ formula (accomplished X as measured by Y by doing Z) and if a candidate cannot quantify any part of their work, they will not get hired. A portfolio entry that says "I rewrote our prompt library" is weaker than "I rewrote our prompt library, cut average review time from 11 minutes to 4, and lifted extraction accuracy on our eval set from 78% to 93%."
Fourth, document your iteration. The Wonsulting guide calls out lack of iteration as a frequent failure mode, and Manatal's October 2025 piece describes prompt engineering as dynamic and evolving, requiring constant learning and experimentation. Hiring managers want candidates who can show what they tried, what failed, what they changed, and what the next version produced. A short before/after log attached to a portfolio artifact often matters more than the artifact itself.
Fifth, demonstrate the counter-moves, the things you did not let the model do. Wonsulting warns against over-reliance on AI without human review and generic prompts leading to generic outputs. Bringing a portfolio that visibly flags hallucination catches, bias mitigations, and human-review checkpoints signals the judgment senior prompt roles require.
Sixth, source beyond the obvious boards. Manatal's October 2025 piece notes that "you might not find many candidates on traditional job boards or platforms" and recommends social media, online communities, blogs, podcasts, newsletters, events, referrals, and talent pools. Recruiters hiring for prompt roles have begun mining GitHub gists, public LLM eval repos, and Discord communities for evidence of work. Candidates who publish their prompts, evals, or red-team notes in a findable place spend less time explaining what they can do and more time being interviewed for it.
What this means for the broader AI talent market
The Prompting Company's five-seat founding bench lands at a moment when the industry's largest employers are quietly rewriting their job descriptions to delete the word "prompt" entirely. Anthropic, OpenAI, and Google DeepMind now hire AI Engineers; LinkedIn's fastest-growing title in the United States is AI Engineer, anchored on LangChain, RAG, and PyTorch; even a fintech like Stripe is posting Machine Learning Engineer roles at $212,000–$318,000 that dwarf standalone prompt postings nationally. What the model labs have effectively conceded is that prompting has become table stakes, a skill subsumed into the broader AI Engineer stack rather than a role worth a separate hire.
That rebrand is forcing the people applying to The Prompting Company (and to every other seed-stage startup doubling down on the title) to make an explicit choice about how they present themselves. PE Collective's placement data shows the trade-off in dollar terms: $95,000 for a career switcher with documented projects, $140,000 for a developer, $160,000 for a domain expert. The common thread is not the "Prompt Engineer" label but the portfolio attached to it — version-controlled repos, eval scripts, iteration logs, and the quantified deltas that turn a clever instruction into evidence of an engineering discipline.
A year-old startup betting five founding seats on the title is, in effect, asking the labor market a question: is prompting a craft worth specializing in, or a baseline competency the rest of the AI stack has already absorbed? The bigger labs have answered. Rival startups are now watching whether The Prompting Company's screen — chained prompts testing reliability, six rubric dimensions, a portfolio gate before the first interview — produces candidates good enough to make the bet look prescient, or whether its new hires will need to rebrand themselves as AI Engineers within eighteen months of joining. The five open roles are the experiment, and the experiment has just gone live.
Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.