Tomas Reyes is a recruiter at a 250-person edtech company, the only technical recruiter on a three-person talent team, and he is filling a backend engineer role that has been open for six weeks. He uses Claude and ChatGPT every day, but his results are wildly inconsistent: one prompt produces a sharp screening summary, the next produces generic filler he has to rewrite by hand. The difference, he eventually realized, was never the AI. It was the structure of what he typed. A vague request gets a vague answer; a well-built prompt gets something he can actually use. This lesson is the anatomy Tomas learned so that every prompt he writes is reliable rather than a coin flip, built around four pillars and one worked example he now reuses for every role.

The Four Pillars of a Recruiting Prompt

A strong recruiting prompt rests on four pillars: context, task, constraints, and specificity. Skip any one and the output degrades in a predictable way. Drop context and the AI guesses at your situation. Drop a clear task and it picks its own. Drop constraints and it wanders off-brand or into territory that creates compliance risk. Drop specificity and it gives you something technically responsive but too generic to use. Tomas learned to treat these four as a mental checklist he runs before he hits enter, the same way an experienced recruiter checks a job posting against the actual role before publishing it.

The pillars are not equally heavy for every task, which is the second half of the skill. A quick grammar fix needs almost no context and few constraints. A candidate-facing outreach message needs all four. Knowing which pillar carries the weight for a given task is what separates a recruiter who occasionally gets lucky with AI from one who gets usable output on the first try, consistently.

Context: Tell the AI Where It Is Standing

Context answers who you are, what you are hiring for, and what situation you are in. It is the difference between the AI writing for a generic company and writing for yours. When Tomas sets up a prompt, his context names his role, the company stage, the specific job, and the candidate when there is one: "You are a technical recruiter at a 250-person edtech company hiring a senior backend engineer for a team that owns our grading platform. The candidate, Lin, has six years of experience in distributed systems and is currently at a larger company."

Context comes in depths, and matching the depth to the task is its own discipline. Shallow context suits standalone tasks, like rewriting one sentence in a more professional tone or fixing the grammar in a job description, where the AI only needs the snippet in front of it and does not need your pipeline explained. Medium context suits candidate-specific work, like outreach, screening, or feedback, where the AI needs the candidate's background, their current role, and why they might be interested, because you are personalizing. Deep context suits strategic work, like defining what to look for in a new role, assessing team fit, or thinking about team dynamics, where the AI needs your company stage, your culture, your team composition, and what the hire must accomplish. More context yields more personalized output at the cost of a longer prompt; less context turns around faster but comes back more generic. Tomas spends the words where personalization actually matters and saves them where it does not.

Task: Name the Action With a Precise Verb

The task is the verb, and vague verbs are where most prompts fail. "Analyze this resume" is not a task, because analyze for what: technical fit, red flags, compensation signals, or seniority? The AI will produce something, but probably not the something Tomas needed. A strong task names a recruiting-specific verb (screen, extract, summarize, draft, evaluate, compare) and then states the target precisely. Tomas rewrites "write an email to the candidate" into "draft a concise, three-to-four-sentence outreach email to Lin that references her work on distributed systems and connects it to our platform's scaling challenges."

The strong version tells the AI the action, the length, and the emphasis in one move. Notice that task clarity also bounds the output: by naming three to four sentences, Tomas prevents the AI from returning a five-paragraph essay he would have to cut down. A precise task is not bureaucratic. It is the fastest path to output that needs editing rather than rewriting.

Constraints: Your Brand and Compliance Guardrails

Constraints define what the output must and must not do. They are where a recruiter's professional obligations enter the prompt, and they are the pillar most often skipped. For outreach and screening, Tomas attaches a standing set of constraints that protect both his employer brand and his legal footing: keep the tone warm but professional, avoiding enthusiasm that reads as insincere; do not assume gender or use gendered pronouns; focus on skills and experience only, never age, location stereotypes, or educational pedigree; do not promise compensation, benefits, or a timeline without approval; assume the candidate may never have heard of the company, so explain the mission and why it matters; and if information is missing, say so rather than inventing it.

The bias and compliance constraints are not optional polish. When AI assists a hiring decision, the recruiter remains responsible for the outcome, and an output that anchors on a protected characteristic such as age or national origin can produce the kind of disparate treatment the Equal Employment Opportunity Commission and laws like New York City Local Law 144 exist to catch. Writing "evaluate against the listed job requirements only, and ignore name, age, school, and any inference about protected characteristics" into the prompt is the cheapest compliance control available. It costs one sentence and keeps the AI focused on the qualifications that legitimately predict performance.

Specificity: The Details That Make Output Usable

Specificity is the level of detail about what you want and how you want it presented. "Summarize this interview" and "summarize this 45-minute backend interview with Lin, focusing on distributed-systems depth, code-quality signals, and communication clarity, as a bulleted list with one sentence per point" produce dramatically different results from the same recording. The specific version tells the AI what to prioritize, what to ignore, and what format to return, which is exactly what makes the output droppable into a hiring-manager debrief without reformatting.

Specificity is also where Tomas avoids the trap of conflicting constraints. Asking for a warm, personable note that is under 25 words yet mentions the mission, team size, growth rate, and a personalized hook is asking for output that cannot exist; the constraints fight each other. Being specific forces him to prioritize honestly. If brevity wins, personalization gets shallower; if depth wins, he allows more room. Naming the trade-off in the prompt is more honest, and more effective, than pretending the AI can satisfy mutually exclusive demands.

Worked Example: Building One Strong Prompt Step by Step

Here is the prompt Tomas builds for his open backend role, assembled one pillar at a time so the structure is visible. Step one, context: "You are a technical recruiter at a 250-person edtech company. We are hiring a senior backend engineer for the team that owns our grading platform, which is straining under growth. I am screening a resume for this role." Step two, task: "Screen the resume below against our requirements and produce a structured summary covering distributed-systems experience, evidence of scaling production systems, and any gaps against a five-year minimum."

Step three, constraints: "Evaluate against the listed job requirements only. Do not consider or infer name, age, gender, school prestige, or any protected characteristic. If the resume lacks evidence for a requirement, state that explicitly rather than assuming. Flag nothing as a red flag unless it relates directly to a stated requirement." Step four, specificity: "Return four labeled sections: Meets Requirements, Partially Meets, Gaps, and Open Questions for the phone screen. Keep each section to three bullets, one sentence per bullet."

Read the four steps together and the value is obvious. The context tells the AI whose role this is, the task names the verb and the target, the constraint embeds the bias and compliance guardrail directly into the instruction, and the specificity fixes the format so the output drops straight into Tomas's notes. He saved this assembled prompt as a template, swaps the role details per req, and now gets a consistent, defensible screening summary every time instead of the coin flip he started with. The anatomy is the whole point: once the structure is reliable, the output is too.

A Second Build: A Candidate Research Prompt

The same four steps assemble a completely different kind of prompt, which is how you know the anatomy is doing the work rather than the wording. Tomas is now researching candidates for a senior frontend role, and he wants to know what he can responsibly learn about a designer-turned-engineer named Priya before he reaches out. His context establishes the frame: he is a technical recruiter at an edtech company hiring a senior frontend engineer, and he is looking at people who have worked on large-scale applications and care about user experience.

His task names the verb and the targets: research and summarize what can be learned about Priya from her public code repositories, her professional profile, and any publicly available work samples, focusing on the technology stack she prefers, her open source contributions, any evidence of mentoring or leadership, and her stated geographic preferences. His constraints do the ethical work: include only information that is publicly available, do not speculate about her personal life or make assumptions about her motivation, say so explicitly if the information is limited, and keep the answer under 250 words. That third constraint is quietly the most important one, because a research prompt without it invites the model to fill silence with plausible invention, and an invented detail about a candidate is worse than no detail at all.

Then comes the step people skip: verify the specificity before you send. Tomas reads his draft prompt back and asks three questions. Does it tell the AI exactly what to look for? Does it tell the AI what to ignore? Does it say what format to return? If all three answers are yes, the prompt is ready. If any answer is no, the missing pillar is exactly where the output will disappoint him, and it is cheaper to fix now than to iterate later.

Three Anti-Patterns to Avoid

Vague context. This is providing so little background that the AI has to guess, as in "summarize this candidate's background." It fails because the model does not know what role this is for, what your company does, which skills matter, or what format you need, so you get a generic summary that may not surface anything relevant to the hire. The fix is a habit: always include the role, the candidate's background, your company and industry context, and what you are evaluating for.

Conflicting constraints. This is asking for mutually exclusive outputs, such as a warm and personable outreach email of no more than 25 words that also covers your mission, your team size, your growth rate, and why this specific person is a fit. It fails because the requirements fight each other and nothing can satisfy them all without sounding robotic. The fix is to prioritize and be honest about the trade-off. If brevity matters most, accept shallower personalization. If depth matters most, allow more room.

Unclear task. This is asking for something vague without defining what success looks like, the "analyze this resume" problem. It fails because analyze could mean technical fit, cultural fit, compensation expectations, or red flags, and the AI will pick one you did not want. The fix is to choose a recruiting-specific verb, screen, summarize, extract, identify, draft, evaluate, or compare, and then say explicitly what you are looking for.

Practice

Run these against real requisitions rather than hypothetical ones, because the pillars only become second nature when you feel the cost of skipping one.

• Deconstruct a prompt. A colleague sends you this: "Write an email to software engineers who might be interested in fintech." Identify what is missing, whether that is context, task clarity, constraints, or specificity, then rewrite it with all four pillars present.
• Build a screening prompt. You are screening resumes for a marketing manager role at a remote-first B2B company. Write a prompt that tells the AI what to look for, what to ignore, and how to structure its findings, naming at least three specific skills or experiences you want highlighted.
• Decide the context depth. For each of these, decide whether you need shallow, medium, or deep context and explain why: drafting a job posting for a backend engineer, writing a rejection message to a candidate, and evaluating whether a candidate fits a scrappy, fast-paced startup team.
• Add the constraints. You are asking AI to summarize an interview. Write the constraints that would stop it from making assumptions about the candidate's background, including irrelevant personal information, hallucinating qualifications the candidate never claimed, or producing output biased against a particular group.
• Iterate and improve. Take a recruiting prompt you have used before, or write a fresh one. Test it, review the output, and note what changed between your first attempt and the polished version. The differences are your personal weak pillars.

Terms Worth Knowing

• Context. The background that frames your request, including your role, your company, the job, and the candidate. Context is what turns generic output into personalized output.
• Task. The specific action you want performed, expressed as a clear verb such as draft, summarize, analyze, screen, or extract. Defining the task explicitly is what prevents ambiguity.
• Constraints. The boundaries and requirements you set on the output, covering tone, length, accuracy standards, and what to avoid. Constraints keep output aligned with your hiring brand and your compliance obligations.
• Specificity. The level of detail about what you are looking for, what format you want, and what good output looks like. Higher specificity produces more targeted, more usable results.
• Prompt iteration. Testing a prompt, reviewing the output, identifying gaps, and refining. Most strong prompts arrive through iteration rather than inspiration.

Putting It to Work This Week

Building effective recruiting prompts is part craft and part discipline. The craft comes from knowing your domain deeply: what information actually matters for a hiring decision, which questions reveal fit, what communication style works at your company. The discipline comes from taking the extra thirty seconds to structure the request before you send it. Who are we talking to, what do we actually want, what are our boundaries, and how specific do we need to be?

Over the next few days, start noticing the quality of your own prompts. When you get output you love, reverse-engineer it and ask what you did right. When you get generic or misaligned output, ask what was vague, what context was missing, and which constraint you should have written down. That reflection, repeated a handful of times, will make you a substantially stronger prompt-writer, and it costs nothing but attention.

Reflection

• Think about the last time you asked AI for help with a recruiting task. What worked and what did not? Would more context, a clearer task, or stronger constraints have improved the result?
• Which of the four pillars do you most often skip or underdevelop, and why? Name one concrete way you could improve on that pillar this week.
• Describe a recruiting decision you are making in the next few days. What would a well-structured prompt for that decision look like, written out in full?
• How would your context depth differ between a quick screening task and a strategic assessment of team fit? What would you add, and what would you strip out?

Related Lessons

The anatomy here is the foundation, and the next chapter applies it to specific work. Prompting for Resume Screening, Sourcing, and Research takes the four pillars into the highest-volume tasks you run, and AI-Assisted Outreach: Templates, Personalization, and Quality does the same for candidate-facing writing, where the constraints pillar carries most of the weight. Research Synthesis: Building Candidate Context from Multiple Sources extends the research prompt Tomas built above into a repeatable method for pulling a candidate picture together responsibly.

No prompt is right on the first attempt, which is why Iterating with AI: Following Up, Clarifying, and Refining is the natural companion to this lesson: it is the loop that turns a decent prompt into a reliable one. Avoiding Bias in Prompts: Language, Examples, and Assumptions goes deeper on the constraints pillar and on the wording choices that can smuggle bias into a request that looks neutral. And once your prompts are working, Hands-On Practice: Build Your Prompt Library is where you turn them into saved templates so the structure survives your busiest week.

Key Takeaways

• Every strong recruiting prompt rests on four pillars. Context, task, constraints, and specificity. Skip any one and the output degrades in a predictable way, so run the four as a checklist before you send the prompt.
• Match context depth to the task. Shallow context for standalone edits, medium for candidate-specific outreach and screening, deep for strategic role definition. Spend words on personalization only where it changes the result.
• Name the task with a precise recruiting verb. Screen, extract, summarize, draft, evaluate, compare, then state the target and the length. A precise task yields output you edit rather than rewrite.
• Put bias and compliance into the constraints. Instructing the AI to evaluate against job requirements only and ignore protected characteristics is the cheapest compliance control you have, and it matters because the recruiter, not the tool, owns the outcome under EEOC guidance and laws like NYC Local Law 144.
• Specificity makes output usable and prevents conflicting demands. State what to prioritize, what to ignore, and the exact format. When constraints fight each other, prioritize honestly rather than asking for output that cannot exist.
• Verify before you send. Read the prompt back and ask whether it says what to look for, what to ignore, and what format to return. A no to any of those three predicts exactly where the output will disappoint you.
• Build once, reuse forever. Assemble a prompt pillar by pillar, save it as a template, and swap the role details per requisition. A reliable structure turns inconsistent results into consistent, defensible ones.