AI Boolean Search Generator for Recruiters: Prompts and Checks

Give ChatGPT or Claude the role's must-haves, job-title synonyms and exclusions, ask for a separate string for each site in that site's syntax, then test each string on its own site, read the first 20 results and log the false positives. A hiring manager should approve the logic before anyone relies on the results.

A free generator hands you a string and no way to judge it. Strings fail silently: a misspelt term, a lower-case NOT or an exclusion that is too greedy removes good people, and nothing on screen says so. The fix is a ten-minute manager review of the logic and a log that shows what each version got wrong.

Follow me on Instagram@sagnikteaches

One role, briefed so a string can be built from it

Take a specialist food importer hiring an import and logistics coordinator. Before any prompt, the brief is written in a form a string can use. The table is the worked role for the rest of this tutorial.

Connect on LinkedInSagnik Bhattacharya

The split between must-have and plus decides the whole string. Only must-haves belong behind AND. If four pluses are all ANDed, the search returns only people who have all four, which is usually a handful and not the best handful. Pluses go into an OR group, or stay out and become a ranking step when you read the results. The intake form feeds this table, and filling a job intake form from the call with AI shows how to capture it.

Subscribe on YouTube@sagnikteaches

A prompt that returns terms and risks, not only a string

Ask for the ingredients first and the string last, and make the model state its assumptions. A bare "write me a Boolean string" prompt gets a confident string and nothing to review. This prompt carries the brief and the site's rules. Indeed's own recruiter guide includes a similar prompt for ChatGPT, and adds that it still needs strategic prompts and human oversight.

I am a recruiter writing Boolean search strings for a hiring manager to approve.
Role: import and logistics coordinator at a specialist food importer. Must-have: hands-on customs clearance or import documentation. Strong plus: Incoterms, freight forwarders, perishable or cold-chain goods, food safety.
Give me: (1) 8 to 10 job-title synonyms people really use; (2) synonyms for each skill; (3) three exclusion terms, each with the risk of wrongly removing someone; (4) one string for LinkedIn people search in its syntax: AND, OR and NOT in capitals, quotation marks for phrases, parentheses only, no wildcards; (5) a shorter string for a general web search engine, under 10 terms.
Do not invent job titles you are not sure exist. List every assumption you made.

This is an illustrative answer with the review verdict beside each part.

Three of those five parts needed a human decision, and none needed more AI. The assumptions list is the most valuable part of the reply: it showed that the assistant had read "coordinator" as the level and "customs brokerage" as a related skill, and both are choices the manager should make, not the model. Asking for the risk beside each exclusion turned a hidden problem into a visible one.

After fixes, the pair of strings below is what goes to the manager. The wide string finds everyone who meets the must-have. The tight string adds the pluses behind an AND, so the people who appear in both are the first calls.

WIDE:
("import coordinator" OR "logistics coordinator" OR "supply chain coordinator" OR "freight coordinator" OR "customs coordinator" OR "shipping coordinator") AND ("customs clearance" OR "import documentation" OR Incoterms) NOT (recruiter OR recruitment OR staffing)

TIGHT (WIDE plus):
AND (perishable OR "cold chain" OR "food safety")

Counting operators and the words inside phrases, the tight string runs to about 40 words. LinkedIn's Boolean help page states no word limit, but a long string is still worth a check on your own seat: if the search fails or the count looks wrong, drop two of the titles and compare the counts.

The syntax differences that break a good string

The same string meets three different rule sets on three sites. This table records what each vendor's own page said in November 2026.

Two lessons follow. First, a string that works on one site tells you nothing about the others, so each site gets its own version and its own test. Second, the failures are silent. A truncated search, an ignored NOT and a typo all return results, just the wrong ones. That is why the checks below do not rely on the string looking right.

Before and after: five edits that changed the count

Here is a string as an assistant first returned it, followed by the five edits a careful reader makes. The original looks reasonable, which is the problem.

AS RETURNED:
(coordinat* OR "logistics coordinator" OR "supply chain coordinator" OR "customs cordinator") and ("customs clearance" OR "import docs") not (sales OR intern OR driver)

None of the five edits needed deep sourcing knowledge. They needed a checklist, and a habit of testing every term before the whole string. That is the case for building the review into the routine: the cost is minutes, and the failure it prevents is a shortlist built on a search that never worked.

The hiring manager's ten-minute review

The manager does not need to read Boolean. They need to read the logic in plain English, which the recruiter supplies as a one-page summary: what the string requires, what it allows, and what it removes. These are the eight checks, in the order a manager can do them.

1. Must-haves. Is every must-have inside an AND group, and is nothing that is merely nice to have in one? Ask the manager to tick each must-have on the brief against the string.
2. Titles. Do the title synonyms cover the level you want, and no more? Senior titles slipped into the sample output above, and they would flood the results with the wrong people.
3. Exclusions. For each excluded word, who does it wrongly remove? The manager sees the risk written beside it, which is why the prompt asks for it.
4. Spelling. Does every term return results when searched alone? Indeed warns that typos can filter out relevant resumes, and a misspelt term fails without any message.
5. Location and pattern. Are these handled in the site's filters, so the string stays about skills and experience?
6. Fairness. Does any term describe who a person is rather than what they have done? Age, sex, ethnicity, religion, disability and nationality, and anything that stands in for them, stay out.
7. Known-good test. Does the string find three people you already know fit the role?
8. Sample. Of the first 20 results, how many would the manager want to speak to?

Checks six, seven and eight are the ones free generators never prompt for. On fairness, a bias test checklist for small agencies covers the screening side, and the same thinking applies upstream in the search string. For the sourcing tools that sit beside Boolean, AI sourcing tools compared on seat costs and consent rules is the next stop.

Here is the plain-English summary the recruiter sends with the strings. This is what the manager reads, so it avoids operators entirely.

This search finds people whose profile mentions one of six coordinator-type titles and at least one of customs clearance, import documentation or Incoterms. It removes profiles that mention recruiter, recruitment or staffing. It does not filter by place and it does not rank. The tight version also requires perishables, cold chain or food safety, and is a shortlist for first calls. It would miss a strong candidate who uses a title that is not on the list, so please tell me any titles I have missed.

Exclusions: what each one wrongly removes

Exclusion words are the easiest way to look precise and lose good candidates. On most sites a keyword exclusion applies to the profile or CV text as a whole, not only the current title, so test it on a known-good profile before keeping it. These examples come from the worked role.

The pattern is that single common words are risky and exact phrases are safer. A phrase such as "delivery driver" removes the person you meant to remove without taking out the coordinator who managed drivers.

Testing on three sites, with a log that shows what went wrong

Run the same routine on every site, one change at a time, so you can tell which change did what.

1. Term by term. Paste each term or phrase into the site alone. A term that returns nothing is probably a typo or a phrase nobody writes.
2. Quotes. Paste the string into a plain-text editor first, so curly quotation marks become straight ones. Then remove the quotes from one phrase and check that the count changes. If it does not, the quotes were not doing their job.
3. Whole string. Run it and note the total count. A count that is tiny or enormous is a warning on its own.
4. First 20. Mark each result relevant or not, and give every false positive a reason code from the table below.
5. Known-good check. Search for three people you already know fit the role, such as past placements or the manager's own team, and see whether the string finds them. Do not paste their details into an AI tool to do this.
6. One change, then log. Edit one thing, rerun, and record it.

This illustrative log runs the worked role across three sites. The counts are placeholders, and your own numbers will differ.

Version three is the entry that justifies the whole routine. The relevant share did not move, 14 of 20 both times, so the first-20 read alone would have kept the exclusion. The known-good check caught that it removed a real fit, and the log shows why the change was reverted. By version four, relevance in the first 20 reached 75%, which is a sensible place to stop. A simple stop rule is four versions or about 75% relevant, whichever comes first. The Indeed row also shows that a missing person may simply have no CV on that site, since Indeed notes that not every potential candidate has a resume in its databases.

False-positive reason codes, and the fix each one points to

Counting false positives is only useful if each one points at a repair. Five codes cover almost everything seen in the log.

If code A dominates, the string is too loose on skills. If code B dominates, the titles are too broad. If code D dominates, add the exclusion that looks obvious, then retest it against a known-good profile. A log that shows the same code in three versions in a row is telling you to change the brief, not the string.

Free generators, privacy and the fairness line

A free generator usually takes a job title and returns a string. It cannot know your must-haves from your pluses, which site you are searching or what the manager rejected last time, so treat its output as one more draft and run the same review. Type only the role into it. A candidate's CV or profile never goes into a generator or a chatbot, because the string needs nothing about a person.

For chatbots, the vendors' own settings matter. Consumer ChatGPT and Claude plans let you switch off the use of chats for model training in privacy settings, while business plans do not train on business content by default. Check which account you are in before pasting even a role brief that names a client.

The log helps on fairness too, as long as it holds counts and reasons and no candidate names. It shows that every string went through the same manager review, which terms were excluded and why, and that no term described a protected characteristic. If a search string for one role carried a term such as a graduation year or a first-language requirement that has nothing to do with the job, the review is where it is caught, and the log is where you can show it was. For prompts on the intake and outreach steps either side of this one, ChatGPT prompts for recruiters: intake, outreach and shortlists picks up where the search ends.

How long it takes, and a routine you can repeat

For the first role, expect roughly 20 minutes to write the brief and run the prompt, 25 minutes to test across three sites, 10 minutes for the manager review and 15 minutes per revision, so about an hour and a half with two revisions. These are placeholders for planning; time your own first role. The second role of a similar type takes about half as long, because the exclusion list, the reason codes and the log template carry over, and only the titles and skills change.

Keep the output small and reusable. One page per role holds the brief, the final wide and tight strings, the site-specific variants, the reasons for each exclusion and the log. When a manager asks why a good candidate did not show up, that page answers the question, and when the same role opens again in a year, it is the starting point instead of a blank prompt.

Questions recruiters ask about AI-written Boolean strings

Can ChatGPT or Claude write a Boolean string for LinkedIn?

Yes, if you give it LinkedIn's rules in the prompt. LinkedIn's help says AND, OR and NOT must be in capitals, quotation marks mark exact phrases, parentheses are the only grouping symbol, and wildcards are not supported. The usual AI mistakes are asterisk wildcards and lower-case operators, so paste the string in, run it and read the results before trusting it.

Why does one string behave differently on each site?

Each site reads syntax its own way. LinkedIn's help documents capitals and parentheses and no wildcards. Indeed's guides show AND, OR, NOT, quoted phrases, parentheses and a skill: field for resume search. Bing's help says AND is the default and only the first 10 terms are used. That is why a string needs testing on every site, not just writing once.

Is it fair to exclude words like junior or intern?

Not automatically. An exclusion usually acts on the whole profile or CV text, so NOT intern can remove an experienced coordinator who interned years ago. Keep any term that stands in for age, sex, ethnicity, religion, disability or nationality out of the string altogether, and have the manager review each exclusion beside the people it would wrongly remove.

How many results should I read to check a string?

Read the first 20 on each site and run a known-good check with three profiles you already know fit the role. Twenty is a small sample, so treat the percentage as a rough guide and log it for every version. Stop refining when the share of relevant results stops improving across two versions, or after four versions.

Further reads

• Free AI Tool for Recruitment: What a Small Agency Gets for Nothing — What a small agency gets from free AI tools.
• AI for Recruitment Agencies: Re-Engaging Your Candidate Database — Searching your own candidate database before the open web.
• Recruitment CRM AI: Candidate Records, Duplicates and Follow-Up — Keeping candidate records and duplicates clean after sourcing.
• Using AI in Recruiting a Small Team: Job Ad to Shortlist, Step by Step — Job ad to shortlist for a small team, step by step.
• AI Agents for Recruitment Agencies: Sourcing, Screening and Review Points — Review points when AI agents source and screen.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: LinkedIn Help (Use Boolean search on LinkedIn); Indeed's employer guides on resume databases, Boolean search for Smart Sourcing and ChatGPT prompts for recruiters; Microsoft's Bing advanced search options page.