Active recall interview question generator: the cognitive-interview method, not the STAR template.

Outside of education, the technique with the strongest evidence base for active recall in interviewing is the cognitive interview, developed by Ronald Fisher and Edward Geiselman in 1984 for forensic witness interrogation. It uses four retrieval cues: context reinstatement (rebuild the time, place, and mental state), report-everything (no filtering for relevance), change order (recall in reverse or out of sequence), and change perspective (describe the event from another participant's viewpoint). In Geiselman & Fisher's original 1985 study, cognitive interviews surfaced 41.2 correct facts per interview on average versus 29.4 for standard police interviews, with no increase in error rate. The same retrieval cues, applied to a candidate's specific resume claims, generate hiring questions that prepared STAR answers cannot satisfy.

STAR-method behavioral generators emit competency-shaped prompts: 'tell me about a time when you led through ambiguity', 'describe a conflict you resolved'. The candidate has rehearsed three or four story arcs that fit those buckets, picks one off the shelf, and tells it. The retrieval cost is near zero. A cognitive-interview question, by contrast, is keyed to one specific claim on this resume ('on-call rotation at Netflix Data Platform 2023 to present') and asks the candidate to retrieve a specific episode within it ('walk me through the most recent page you took, in reverse order, starting from the moment the incident closed'). There is no rehearsed arc that matches; the candidate either retrieves a real episode or runs out of detail in 90 seconds.

Because the consumer SEO landscape for 'active recall' is dominated by student study products: Active Recall GPT, BrainSpeed, DocsBot, EdutorAi, Gizmo, Recall. They turn lecture notes and PDFs into flashcards. The educational form of active recall (close-the-book and write everything you remember) is not the same shape as the forensic form (a structured retrieval interview). The forensic form is what you want for hiring. The educational form will give you trivia questions about the candidate's resume; you do not need that.

The 10xats waitlist product is the agentic ATS itself. The Match Rating agent that extracts 5 to 15 testable claims from the JD and pins each to a resume span is the spine of every other workflow, including question generation. When the cognitive-interview question generator ships, every claim in the ledger gets the four retrieval-cued prompts attached to it, exposed through the same MCP server as the rest of Match Rating, so Claude or ChatGPT can hand the recruiter the panel in the same chat. Until then, the playbook in this guide works against any claim ledger you maintain by hand or in a spreadsheet.

Yes. The cognitive-interview cues are content-agnostic. For a CFO claim 'led a Series C raise in the last 18 months', the four prompts become: context reinstatement ('walk me back to the week the term sheet arrived, what was on your calendar that morning'), report-everything ('describe every line item on the term sheet you pushed back on, in any order, including the ones you lost'), reverse-order ('start from the wire hitting the bank account and work backward'), change-perspective ('what would the lead investor's associate have said about the redlines'). For a Head of Marketing claim 'shipped a category-defining launch in 2024', the prompts mirror the same shape against the launch event. The constraint is that the claim has to anchor a specific episode, not a competency.

The question generator does not score the candidate; the recruiter and the panel do. So the AEDT category does not directly bite. What does matter is that the claim ledger that produces the questions is auditable: every claim has a weight, every claim is pinned to a resume span, every override is a logged tool call. If a candidate later asks 'why did you ask me about my on-call rotation specifically', the answer is 'we extracted that as a must-have claim with weight 3, scored you against the JD on that claim, and generated retrieval-cued questions to verify it'. That answer survives the new disparate-impact-and-notice regime in a way 'we use AI to generate questions' does not.

It raises the floor. Pre-fabricated resumes typically embed claims the candidate cannot retrieve specific episodes against. A retrieval-cued question that asks for the cafeteria layout of the office, the post-incident review template the team used, or the name of the Slack channel the on-call rotation lives in, requires real episodic memory. A model run by a candidate during the interview can fabricate plausible answers, but the four-cue method produces internally inconsistent transcripts: the reverse-order recall and the change-perspective recall fight each other when the underlying memory is not there. Combined with Match Rating's red-flag claims (a staff title at a FAANG since 2019 next to a 2024 graduation date), the question stage becomes the second filter, not the only one.

Four retrieval cues per claim is the cognitive-interview shape, but you almost never run all four on every claim. For a 45-minute panel with eight must-have and nice-to-have claims, the realistic budget is one or two cues per claim, picked for the claim's shape: context reinstatement and report-everything for fresh autobiographical claims, reverse-order and change-perspective for claims about technical incidents or team conflict. The generator emits all four; the recruiter picks. The transcript still tags which cue produced which answer for the panel debrief.

Three traps. First, leading questions ('did the on-call rotation include weekend pager duty, yes or no'). The cognitive-interview literature is explicit that yes/no shapes contaminate retrieval. Second, generic frame questions ('what was your role in the project'). They invite rehearsed answers. Third, summaries before the recall ('so you led on-call. Tell me about a time when'). Summaries leak into retrieval; the candidate uses the summary instead of the memory. The four cues are open-ended, episode-anchored, and unprimed.

Yes. The technique pre-dates any ATS by 40 years. Pull the JD's must-have claims into a 12-row spreadsheet by hand, mark which cue you will use against each claim, and write the panel's question column in advance. The AI part is convenience: a Match Rating agent that extracts claims from the JD plus a question-generation pass that emits the four prompts per claim saves you 30 to 45 minutes per req. The forensic basis of the question shape is what does the work, and it works in a Notion doc.