Joy

Joy is your People Expert: part organizational anthropologist, part exacting talent operator. She prepares consequential conversations, builds evidence-led hiring systems, and pressure-tests people decisions without reducing anyone to a score. Her promise is simple: clearer, kinder, more defensible people moves.

Who Joy is#

Joy notices power, context and subtext, but she never pretends to read minds. She pushes back when a person is being reduced to pedigree, polish, a metric, a stereotype, or a loaded adjective. Her north star is the next useful move: insight should change how you prepare the conversation, define the bar, or close an evidence gap—not become a dossier.

Two disciplines make her insight trustworthy across both halves of the job:

• Evidence and inference are kept separate. What the material actually shows becomes a sourced observation; the plausible dynamic is labeled as a hypothesis with an alternative explanation.
• People are never the score. Candidate numbers describe document coverage; decision-audit numbers describe packet readiness; DX numbers describe a work system. Humans remain the decision-makers.

Giving Joy material#

Anything is raw material: a pasted Slack conversation, an email thread, meeting notes, a one-line description, a job description, a resume, even venting after a tough 1:1. Material often arrives as a screenshot — Joy reads it like text, names, timestamps and reactions included. Tell her your relationship to the person ("she's my peer EM", "he's a candidate for the platform role") and she records it; if that context is missing, she leaves it unknown or asks when it materially changes the work.

The relationship notebook#

Joy keeps one working relationship record per person, updated in place. New writes contain stable identity/context, sourced observations, and—only when it changes the next move—a falsifiable hypothesis with evidence and a credible alternative. Trait chips, “essence” labels, unsourced quotes, motive claims and sensitivity dossiers are legacy fields, not part of Joy's active write contract. She never saves protected or highly sensitive traits, health, family status, politics, union activity, or intent to leave.

Evidence vs inference#

This is what keeps Joy honest. A behavior in the conversation becomes an observation, with the source noted ("Slack convo, 2026-06-09"). A working hypothesis is stored only as a structured review item: the tentative hypothesis, the evidence that supports considering it, a credible alternative explanation, and confidence. Free-floating hunches and speculative mental-state stories are not surfaced as usable context; a hypothesis stays revisable and never becomes employment-decision evidence.

Tailoring your approach#

Ask Joy how to handle a specific situation with a specific person and she coaches from the sourced exchanges on file. She names what was observed, labels any interpretation, gives a credible alternative, and then proposes the smallest useful move. If the record is thin, she says so instead of manufacturing a personality story.

Learning your voice#

Whenever shared material contains your own messages, Joy can update your voice profile: tone, sentence length, greetings and sign-offs, punctuation and emoji habits, and signature phrasing. She learns only from writing identified as yours and keeps the profile at the pattern level; short representative snippets are retained only when you ask or they are necessary. If authorship is ambiguous, she asks once.

Drafting in your voice#

When you ask for a draft, Joy writes in your actual register — not polished assistant-speak — and tailors the content to the recipient's profile. So a message to your terse, data-first VP and the same message to a report who needs context first come out differently, both sounding like you. (Sage does this too, reading Joy's profiles; you can ask either.)

DX system signals#

With DX connected, Joy can use its official, intentionally read-only MCP surface to inspect team, catalog, initiative, scorecard and Data Cloud evidence. She grounds every live claim in the returned data; if the connection or query fails, she says so. The default lens is the work system: flow, friction, review load and developer experience—not individual worth.

Contribution evidence briefs#

Joy does not create “top performer” leaderboards from activity data. If you explicitly ask for an individual contribution brief, she requires a defined window, names the available evidence, and surfaces confounders such as role, tenure, KTLO, on-call load and review burden. The output is a brief for a human review—not a performance verdict.

Trends over time#

Any "over time" question—flow trend, review load by week, cycle-time drift—Joy answers with the time window, cohort, query provenance and a chart when the returned data supports one. A chart created from an MCP pull is a point-in-time analysis unless the app explicitly shows a refresh binding; Joy never promises an automatic watch she cannot execute.

Requisition & JD auditing#

Before a role goes out, hand Joy the draft job description and she audits it — because the JD is where a search quietly narrows or widens. She runs four checks and returns edits, not just complaints:

• Inclusive-language linting. She flags biased or coded phrasing — "rockstar", "aggressive", "young and hungry", gendered terms — with neutral swaps you can accept inline.
• Pay-transparency check. If there's no salary range, she flags it, since a missing range is both a candidate-experience problem and, in a growing number of jurisdictions, a legal one.
• Requirement stacking. She catches the fifteen-must-have wishlist and the seniority-vs-comp mismatch — a staff-level requirements bar pinned to a mid-level band — and says which requirements to cut or move to "nice to have".
• JD scaffold. Starting from scratch, she'll draft a clean structure — role summary, what you'll do, must-haves, nice-to-haves, band and level — for you to fill in.

Sourcing#

Give Joy the role and she builds you the tools to find people: boolean and Google X-ray search strings, an ideal-candidate persona distilled from the JD, and title expansion so you're not missing the ten adjacent titles that do the same job under a different name. She'll tune the string with you — loosen it when it's too narrow, add a must-have when it's too noisy.

Candidate evidence matrix#

Give Joy a job description, a resume, and the explicit job-related criteria you confirmed. She never silently derives requirements from JD keywords. Her deterministic matrix reports each criterion as found in this document or not found in this document, shows the source excerpt, distinguishes a stated skill from an action/outcome example, and gives you the exact probe to close the gap.

The 0–100 headline is simply document coverage: the percentage of stated criteria with exact textual evidence in that resume. It is not candidate quality, fit, potential, seniority, or a recommendation. Exact token matching also prevents “Go” from matching “ongoing” and “R” from matching arbitrary prose.

Decision evidence audit#

Before a hiring debrief, promotion committee, performance review or compensation discussion, give Joy the packet. Her new code-driven audit scores evidence readiness across five dimensions: specific examples, outcomes and impact, stated-criteria coverage, counterevidence and unknowns, and bias hygiene. It flags “culture fit,” “executive presence,” gut-feel language, absolutes and explicit protected-trait references for human review.

The audit deliberately caps a packet that lacks concrete evidence or contains sensitive-trait language. The card then asks the smallest set of questions needed to make the packet defensible. Again, the score is the packet—not the person and not the decision.

Structured interviews#

Joy designs the loop, not just a single interview. Give her the role and she lays out a competency-based interview loop — which interviewer covers which competency, with no overlap and no gaps — backed by a question bank per competency and anchored scorecards so "3 out of 4" means the same thing to every panelist. Afterward she produces a debrief evidence summary: recommendation distribution, coverage, unproven competencies and dissent. She never converts panel votes or average ratings into a hire/no-hire decision.

Interview plans & verdicts#

For a single interview, Joy compiles a structured, ready-to-run interview plan and hands it over as an interactive card with a "Conduct interview" button — you rate and take notes per question as you go. Afterward she reads back the captured evidence and helps formalize it. A verdict card is saved only when a named human decision owner explicitly confirms the inclination; Joy supplies the evidence write-up, never the hiring decision.

Hiring pipeline#

Joy does the funnel math so a hiring plan is a plan, not a hope. She reads pass-through at every stage and finds the leak — the step where good candidates fall out — and she works the yield math backwards: given your stage conversion rates, how many candidates you need to source to land the hire. She tracks time-to-fill, offer-accept rate, and interview-panel capacity — because a pipeline that needs forty onsites a month from a team that can run twenty is the real bottleneck.

Greenhouse (live ATS)#

Connect Greenhouse from Connections with its official OAuth MCP endpoint—no custom bridge URL or pasted Harvest token. Greenhouse MCP is currently an open beta for Core, Plus and Pro organizations; a Site Admin enables the available scopes in Dev Center → MCP Access, and the user's own Greenhouse permissions remain in force.

Joy can then work from the live recruiting data and capabilities your organization actually enabled: pipeline visibility, stage movement, scorecards, interview kits, job completeness and supported actions. Because Greenhouse's beta catalog varies by scope and account, Joy uses a code-driven search → inspect → run facade to reach the complete live catalog without dumping every remote schema into the prompt. Exact names are checked live and every operation remains fail-closed behind approval; Greenhouse's scopes and audit controls are an additional boundary, not a replacement for Zimac's.

Joy also appears on the official Calendly connection for interview availability, event types, invitees and scheduling links. Zimac carries Calendly's complete published MCP tool catalog, while the same lazy facade keeps the prompt compact. Calendly uses OAuth 2.1 with PKCE and dynamic client registration; all remote operations still cross the exact-input approval boundary.

Watches, studios & lessons#

Joy can open People, Interviews, Calendar, Mail and Connections with a one-click chip, query user-shared datasets, and record durable corrections such as “always show the confounders” or “never screen without explicit criteria.” Night Shift can still watch the native pull types it lists; DX and Greenhouse MCP pulls are point-in-time until a typed refresh/watch recipe is actually available.