The conversation you are not having

Dana is an agency recruiter who carries about 18 open requisitions at a time across two client accounts, mostly mid-level software and product roles. She adopted AI for resume screening six weeks ago and her honest verdict is that it saves her maybe ten minutes a candidate, no more. The reason is not the tool. It is that Dana sends one prompt, reads the output, decides it is "fine," and moves on. She treats every AI response as a vending machine: put in a prompt, take out an answer, done. The recruiters getting real leverage from the same tool are doing something Dana is not. They are having a conversation.

The best recruiting outputs are rarely one-shot. A first response is a draft, and a draft is something you respond to. You ask, you read the output like a critic, you name what is missing, you ask a targeted follow-up, and the second or third version is genuinely better than anything you would have gotten by trying to write one perfect prompt up front. This lesson is the iteration loop, applied to the screening, drafting, and comparison tasks Dana does every day.

The four-step iteration loop

Iteration has four steps: review, identify the gap, ask a follow-up, and refine for next time. The first step is the one Dana skips. When the output comes back, she needs to read it like a critic rather than a customer, running it against five questions. Is it accurate, meaning does it match what she already knows about the candidate or the role? Is it complete, meaning did it answer every part of what she asked? Is it useful, meaning can she actually act on it? Is it biased, meaning does it lean on an unfair assumption or a proxy for a protected characteristic? And is it in the right format, meaning can she scan it in ten seconds or does she have to hunt? She is not looking for perfect. She is looking for the single most useful thing the output is missing.

Step two is naming that gap precisely, because a vague reaction produces a vague follow-up. "This isn't quite right" gets her nowhere. "This covers years of experience but says nothing about technical depth in the specific areas I care about" gets her a strong second pass. The precision of the gap statement determines the quality of everything downstream.

It helps to keep a few gap statements in your working vocabulary, because the same handful recur across screens. The output covers years of experience but not technical depth in the areas that actually matter for the role. The summary says what the candidate did but not why they did it or what they took away from it. The comparison ranks people on experience but says nothing about communication or how they work alongside others. The analysis names a red flag but does not offer a single question you could put to the candidate about it. Each of those is concrete enough to hand straight back to the model as the next instruction, which is the practical test of a well-named gap.

Step three is the follow-up itself, which is short and surgical. She is not re-running the whole analysis. She is pointing the model at one dimension and asking it to go deeper. Step four happens after the conversation ends: she notices what she had to ask for and folds it into the prompt template so next time the output arrives complete on the first try. That last step is how iteration compounds. A follow-up you ask once is a fix; a follow-up you bake into the template is an upgrade to every future output.

Make step four concrete and it stops being an aspiration. Suppose you screen engineers most weeks and you notice you ask the same follow-up every single time: what has this person actually shipped? Fold it into the opener. Instruct the model that for each role it should identify the specific projects the candidate shipped or contributed to, the technologies used, the size of the team, and any impact or time-to-market detail the resume states. From that point on, every engineering screen arrives with project detail already in it, and a follow-up that used to cost you a turn has quietly become part of your baseline.

A worked example: three turns on one candidate

Here is Dana screening a backend engineer named in her notes only as Candidate 7, for a platform role at one of her client accounts. Watch the output deepen across three turns.

Turn 1, the opener. "Summarize this resume in five bullets, focused on backend engineering experience." The model returns a clean summary: eight years of experience, three companies, Python and Go, a payments system, a couple of promotions. Useful, but generic. The gap Dana names: it tells her what the candidate did, not how deep any of it goes.

Turn 2, the depth probe. "Good. Now focus only on technical depth. Which technologies does this resume give real evidence of deep experience with, versus ones merely listed? For each, cite the specific line that supports it. If a skill is only listed with no supporting evidence, say so." This is the turn that earns its keep. The model now separates "Go, with three years building the payments matching engine described under the second role" from "Kubernetes, listed in a skills section with no supporting project." That distinction is exactly what Dana would otherwise spend fifteen minutes reconstructing by hand, and it is the difference between a real signal and resume keyword stuffing.

Turn 3, the red-flag follow-up done right. The summary flagged that the candidate held three jobs in eight years, which the model called possible "job hopping." Rather than accept that label, Dana probes it: "You flagged frequent job changes. For each role, give the exact tenure. For any short stint, note whether the resume offers context such as an acquisition, a contract role, or a relocation. Is there a pattern, or are these one-offs? Do not speculate beyond what the resume states." The model comes back with tenures of three years, four years, and one year, and notes that the one-year stint is labeled a contract. The "red flag" dissolves into a normal career, and Dana has turned a thin label into context she can actually use.

Three turns, about four minutes, and Dana now has a depth-verified, context-rich read instead of the generic five bullets she would have shrugged at and filed. Notice what she did not do at any point: dismiss the candidate on the strength of a label. A flag is a prompt to gather context, not a verdict, and the follow-up is how you convert one into the other.

The anatomy of a strong follow-up

Turn 2 and turn 3 above share a structure worth making explicit, because it is what separates a productive follow-up from a frustrating one. A strong follow-up does four things in order. It acknowledges what the previous output got right, which keeps the useful material in play instead of triggering a full rewrite. It names the specific gap, so the model knows exactly where to dig. It asks for the missing piece in concrete terms, ideally with a numbered list of what to address. And it adds a constraint that keeps the model honest, most often the instruction to flag when the source material does not support an answer rather than inventing one.

In practice that reads as a single short paragraph. Thank the model for the summary of technical experience. Say that you want to go deeper on leadership and team impact. Ask it to address, in order, the teams the person has led or influenced, any feedback from direct reports that the source mentions, and any evidence of developing junior colleagues. Then close with the constraint: if the resume does not state it, say so rather than inferring. Four moves, four sentences, and the model has everything it needs to give you a second pass that is genuinely additive.

That fourth element is the recruiter's safeguard. "If this information is not explicit in the resume, note that rather than inferring it" is the single most valuable sentence Dana adds to her follow-ups, because the failure mode that hurts a hiring decision is not a missing detail, it is a confidently fabricated one. The constraint converts the model from a tool that fills gaps with plausible fiction into one that marks the gaps for her to investigate.

Iteration across the tasks you actually do

The loop looks slightly different depending on the task, but the move is always the same: deepen, do not discard. For screening and analysis, Dana opens broad and narrows: summarize, then probe technical depth, then ask for cross-functional evidence with examples, then ask what to verify in the interview. Each turn adds a layer.

For drafting and revision, such as an outreach email or a job post, iteration targets tone and specificity rather than information. A first draft comes back competent but generic; her follow-up is "personalize this to the candidate's specific work on the payments matching engine, and drop the phrase 'rock star,' which we never use." A second pass nudges tone: "make it read like a note between two engineers, not a recruiter pitch." Two turns and the message sounds like a person.

For comparison and calibration, where she is weighing three finalists, the opener produces a side-by-side and the follow-ups pull out decision-relevant judgments: "based only on the evidence in these resumes, which candidate shows the most experience operating under ambiguity, and cite the specific evidence." Then: "which would need the most onboarding ramp, and which could contribute from the first week." In every context the principle holds. She is not rejecting the output, she is using it as the first turn of a conversation that gets sharper each round.

When to iterate and when to move on

Iteration has diminishing returns, and the recruiter's discipline is knowing where the curve flattens. The honest rule of thumb: the first follow-up usually adds a lot, the second adds a useful amount, and by the third or fourth the gains are marginal wording changes. Dana iterates when the output is mostly useful but missing a dimension she genuinely cares about, when it is close but the tone or format needs a turn, or when she wants a second angle on the same evidence. She moves on the moment the output is good enough for the decision in front of her, when another round would not change her call on the candidate, or when she is spending more time iterating than the output is worth.

The leverage test settles most cases. If one follow-up will save her thirty minutes of manual analysis, she asks it without hesitation. If she has already spent fifteen minutes polishing something that was eighty percent done after the second turn, she stops, because the remaining twenty percent is not worth a third of an hour. And a separate warning sign: if two follow-ups have not meaningfully improved the output, the problem is usually the original prompt, not insufficient iteration. That is the moment to start over with a better opener, not to ask a third time.

One thing Dana keeps for herself: she never asks the model "what else should I consider about this candidate?" She decides what matters for the hire. She identifies the gaps and directs the follow-ups. The AI deepens her analysis on the dimensions she chooses; it does not get to set the agenda for what a good hire looks like. The judgment stays human, and the tool serves it.

Three anti-patterns that waste your turns

Vague feedback. This is telling the model that the output is "not quite right" without saying what is missing, as in "this summary feels off, can you try again?" It fails because the model has no idea what to improve, so you get the same content in different words and you have burned a turn to learn nothing. The fix is the discipline from step two: always name the gap. "This tells me about technical skills but nothing about leadership experience" is a working instruction. "This feels off" is a sigh.

Over-iterating. This is asking for revision after revision after revision, each one improving the output a few percent while you keep going. It fails because the returns flatten fast; after two or three turns you usually have the great majority of the value available, and everything after that is effort you will not get back. The fix is to step back after two or three turns and ask one question: is this output actionable? If it is, move on. If it is not, you probably need a different prompt approach entirely rather than another round of the same conversation.

Letting the AI drive the iteration. This is asking the model to suggest its own follow-ups, as in "you have summarized the resume, what other questions should I ask about this candidate?" It fails because you have handed the agenda of your hiring decision to a tool that does not know your team, your role, or your obligations. You are the one who should know what information matters for the hire. The fix is to keep the loop asymmetric on purpose: you identify the gaps, you ask the follow-ups, and AI supports your decision-making rather than the other way around.

Practice

Work through these with real outputs from your own requisitions rather than invented ones, because the whole skill lives in noticing what your particular roles need.

• Identify the gaps in an output. Take a summarized resume or set of interview notes that AI produced for you. Review it and write down four things: what is included and accurate, what is missing that you would want to know, what format change would make it more useful, and what assumption or bias it carries.
• Write a follow-up question. Using the gaps you just listed, draft a single follow-up that addresses two or three of them. Keep it specific and focused, and finish it with a constraint that tells the model to flag anything the source does not support.
• Run an iteration test. Pick a recruiting task you do often, whether screening, sourcing, or drafting. Write an opener, get an output, and run two or three follow-ups. Document what each turn added and the point at which you had enough to decide.
• Practise knowing when to stop. Ask AI to draft an outreach email. The first version is decent but generic, the second is more personalized and better, and the third is the same content in slightly different words. Decide where you would have stopped, and write down the reason, because that reason is your own diminishing-returns rule made explicit.
• Refine your template. Take a recurring task, write your initial prompt, test it, iterate twice, then write the refined prompt that captures what you learned. That refined version is your new template, and it should make the first two follow-ups unnecessary next time.

Terms worth knowing

• Iteration loop. The cycle of asking a question, reviewing the output, identifying the gaps, asking follow-ups, and refining the prompt. Most effective recruiting prompts develop through this loop rather than arriving fully formed.
• Follow-up question. A targeted, brief prompt that deepens one dimension of a previous output. Good follow-ups are specific and build on the context already established rather than restating it.
• Gap. Information or analysis missing from an AI output. Identifying gaps clearly and specifically is what makes a follow-up effective.
• Diminishing returns. The point at which additional turns produce minimal improvement. Smart iteration stops before the curve gets too flat.
• Prompt refinement. Updating your template based on what you learned while iterating, so future outputs arrive better without needing as many follow-ups.

Putting it to work this week

Effective recruiting with AI is iterative rather than one-shot, and that is good news, because it means you do not need a perfect prompt. You need a prompt good enough to start with and the willingness to have a conversation about it. That is a more human way to work than trying to compose the flawless instruction on the first attempt, and it produces more nuanced results.

So this week, pick one recruiting task you do regularly. Write a basic prompt for it. Use it twice, iterating each time, and pay attention to how the output improves between the first turn and the third. Then write the refined version that incorporates what you learned. That becomes your template for the next ten times you do the task, and the ten after that.

Reflection

• Think about a recruiting task you completed recently. If you had iterated with AI instead of accepting the first output, what would you have wanted to explore further?
• Which of your recruiting decisions require the deepest information, and which could you make with less? How should the depth of your iteration depend on how consequential the decision is?
• How do you currently know when you have gathered enough information about a candidate? How might a structured loop help you make that judgment explicit rather than instinctive?
• When you review AI outputs, do you find yourself concluding "good enough" too quickly? What would change in your work if you treated every output as a draft that could be improved?

Related lessons

Iteration begins with a decent opener, and Prompt Anatomy: Structure, Context, and Constraints is where that opener gets built, since a prompt carrying context, task, constraints, and specificity leaves you fewer gaps to chase in the follow-ups. When you are applying the loop to screening and research specifically, Prompting for Resume Screening, Sourcing, and Research supplies the task-shaped openers that make turn one worth responding to.

Step four of the loop, refining the template, is the whole subject of Hands-On Practice: Build Your Prompt Library, which is where the follow-ups you keep repeating turn into reusable assets. Iteration also has a limit: when the problem is not depth but reliability, Red Flags and When to Reject or Escalate AI Output tells you when to stop the conversation and discard the output instead of deepening it. And because the next high-stakes application of these prompting skills is candidate-facing writing, AI-Assisted Outreach: Templates, Personalization, and Quality takes the same loop into drafting authentic, personalized messages where tone matters as much as content.

Key takeaways

• Treat the first output as a draft, not an answer. The recruiters who get little from AI accept the first response and move on. The ones who get leverage read it like a critic and respond to it. Most strong recruiting outputs come from two or three turns, not one perfect prompt.
• Name the gap precisely, because vague feedback produces vague output. "This isn't right, try again" gets you reworded sameness. "This covers experience but not technical depth in the areas I care about" gets you a real second pass. The precision of your gap statement sets the ceiling on the follow-up.
• Structure every follow-up the same way. Acknowledge what was right, name the specific gap, ask for the missing piece in concrete terms, and add a constraint. The most valuable constraint is "if the source does not support this, say so rather than inferring," because a fabricated detail hurts a hiring decision more than a missing one.
• Deepen, do not discard. Across screening, drafting, and comparison, iteration means using the output as the first turn of a conversation, not rejecting it and starting over. Each turn adds a layer: depth, then evidence, then the specific judgment your decision needs.
• Know where the curve flattens. The first follow-up adds a lot, the second a useful amount, the third mostly marginal wording. Use the leverage test: iterate if a follow-up saves real time, move on once the output is good enough for the decision. If two follow-ups have not helped, fix the opener instead of asking again.
• Refine the template so the loop compounds. When you find yourself asking the same follow-up every time, fold it into your prompt. A follow-up you ask once is a fix; a follow-up baked into the template upgrades every future output.
• You drive the iteration, not the AI. Decide what matters for the hire yourself. Identify the gaps and direct the follow-ups. Do not ask the model what else you should consider. It deepens the dimensions you choose; it does not define what a good candidate looks like.