The whole journey, written down.
How Mandate works from the first access request to a running search — every step as it is built today, nothing promised. Read it before you ask for a seat.
Requesting access
Mandate is onboarding its first pilot executive search firms. There is no self-serve signup: access is granted by approval, a founder reads every application, and nothing asks for a credit card — there is no trial tier to fall into.
The form
The request lives at /request-access. Six fields, and only the first two are required:
- Full name and email — required; the email is where your invitation will eventually arrive.
- Company, role or title, how you heard about us, and what you are trying to solve — all optional, all read by a person.
When the request lands you see Request received and a promise we keep by hand: we reply within 48 hours, and if we approve, a sign-up link arrives at the email you provided.
What happens behind it
- Submissions are rate-limited. If the limiter refuses you it says so politely — "We've already received a request from you. We'll be in touch." — and your earlier request stands.
- Sending the same email twice does not create two requests, and the form will not tell a stranger whether an address has applied.
- Your request enters a review queue that only Mandate's founder can read. What happens next is the subject of the following chapter.
Approval, and the join link
Approval is a decision by a person, not a switch flipped by software. When the founder approves your request, the product provisions two things: a place for you to work, and a personal invitation to it.
What approval actually creates
- If your firm is new to Mandate, approval creates your organisation — a private workspace no other firm can see — and issues you a staff invitation as its admin. You own your team from day one.
- If your firm already runs on Mandate, approval issues you an invitation into the existing organisation at whatever role its situation calls for.
Either way, the artefact you receive is a single join link. The product emails nothing; the founder sends the link to you personally, at the address from your request.
The join contract
The invite is the approval. Opening the link shows you exactly what you are walking into: the organisation's name, who invited you, the role you will hold, and the email the seat is for. There is no separate "pending" state and no second review — nobody's account exists until they set their own password at the link.
- The link is personal and single-use, and it expires fourteen days after issue.
- You choose a password (minimum length, mixed case, a digit and a symbol), and clicking the link is itself what confirmed your email.
- The moment your password is set you are signed in, landing on your desk — active, in your organisation, with your role already in place.
A dead link — expired, withdrawn, or already used — says only "This link no longer works" and tells you to ask for a fresh one. It deliberately does not reveal which kind of dead it is.
First sign-in
Signing in lands you on the Portfolio — the authenticated home. Everything in the product hangs off a left rail with four groups:
- Workspace — Portfolio, Desk (managers and admins), Analytics, Objectives, Placements, Activity.
- Search — Clients, Mandates, Candidates (with Pool search), Network.
- Intelligence — Executive Intelligence, with Competencies and Role templates.
- System — Agents, Skills studio, and Settings (Members appears for admins).
The Portfolio
Once your searches are running, the Portfolio shows your action queue, your tasks, four headline numbers — active mandates, candidates, added this week, and at-risk mandates flagged by the Search Health Agent (flagged for attention, never judged for you — decision support, never a recommendation) — and a table of active mandates with their status and health.
Before you have data: sample mode
A brand-new workspace has nothing in it, so the Portfolio opens in sample mode: an example account showing what Mandate looks like in use. Sample content is always labelled — a banner reading "Sample dashboard" states plainly that none of it is your data, and every sample screen carries a // sample data marker in its meta line. The moment your own first mandate exists, the real workspace takes over.
The Desk
Admins and managers also see Desk — desk oversight across every recruiter: load, pipeline, placements, open tasks, and mandate reassignment, all derived from the live rows. In its own words: nothing scored, nothing asserted.
The mandate loop
A search in Mandate runs as a loop: describe the role once, calibrate what matters, feed it candidates, and let ranked evidence accumulate until a slate is ready. Every agent output along the way is decision support — never a recommendation; you review, edit and approve at each step that matters.
Intake — one line
A new mandate starts at New mandate with a single line — "Head of IT Ops for RBC Capital Markets" is the whole input. On submit the Intake, Company Intelligence, Role Spec and Calibration agents activate, the mandate page opens immediately, and the title resolves from "Analyzing…" to the parsed role and company within seconds.
The onboarding wizard
Each mandate then carries a five-step questionnaire that turns your knowledge of the client into calibration:
- Origin — where this mandate came from and why now.
- Must-haves — non-negotiables, applied as hard filters.
- Anti-patterns — deal-breakers you never want to see again.
- Stakeholders — the interview panel and what each person probes.
- Priorities — the role's priority signals, each with a weight.
Calibration
From the intake and the wizard, the Calibration Agent derives a scoring model across five dimensions — technical, domain, leadership, regulatory, transformation — each weighted 0–10 and shown with its rationale on the mandate page. There is no slider to drag: the levers are the wizard's answers, the Optimize screen, recalibration from feedback, and Calibration History, where every prior model can be inspected and restored.
CV upload
Candidates enter as documents: PDF or DOCX, 10 MB max, dropped onto CV Intake. The CV Parsing Agent extracts role history, scale, tech exposure and archetype in one pass — usually 20 to 40 seconds — into org-scoped private storage.
Evaluation
Each candidate page holds the parsed profile (every field editable), the evidence, and a fit reading weighted against the approved calibration. The page says it itself, verbatim: decision support — never a recommendation.
Ranking and the shortlist
The ranking leaderboard scores every parsed candidate against the calibration and sorts them into four tiers — Optimal, Strong, Stretch, Below Bar. From there you compare candidates side by side, generate the comparative market report (exportable as Markdown, HTML or PDF, with a client email draft), and build the slate: three, five, or a custom size, finalised with an explicit submission.
Feedback and recalibration
Feedback — yours or a hiring manager's — is interpreted for preference changes, contradictions and bias flags, and can recalibrate the scoring model. The loop closes: the next ranking pass reflects what the client actually said.
The weekly report
Each mandate has a Weekly Report: an executive summary, top candidates, and what was sourced this week, generated on demand and archived by week. It exports as Markdown or PDF and can open a pre-filled client email draft.
The candidate portal, and erasure
Candidates have a door of their own. From a candidate's page the recruiter gets a portal link — one live link per person, thirty days, reissuing returns the same one — and hands it over. Nothing is emailed by the product.
What a candidate sees
The portal shows the person exactly what is held about them and lets them act on it:
- What is held — contact details, current title and company, profile links, whether a CV is on file, and how they were found.
- The searches they are in — each live search by role title and stage, with a Withdraw button per search. Withdrawing is visible to the search team, who stop considering them. The client behind a search stays confidential until the team introduces it.
- Correct your details — phone, location and profile links are editable in place; the identity anchor (name or LinkedIn) and email changes go through the search team.
- Submit a newer CV — uploads go to the team for review by hand; a new file never replaces the old one automatically.
Asking for erasure
The portal's last section lets a candidate ask the firm to erase what it holds about them, with an optional note. The contract is stated where the button lives: the request is reviewed by a person, erasure is not automatic, and the team may need to keep some records where the law requires it.
Once filed, the request is visible to the search firm's admin — whose screen says plainly to stop working the person — and to Mandate, where a founder reviews it and either declines or resolves it, recording a note of what was erased. Review and execution sit with a person at every step.
Member management
Admins run their own team from Settings → Members. Mandate's founder opens the front door; everything after that belongs to the organisation.
Inviting a colleague
Invitations work exactly like the door you came through: the admin names the person, picks their role, and Mandate emails them the join link; it stays on screen to copy as well. The link is single-use, expires in fourteen days, and whoever redeems it joins active, in the role chosen at issue — the invite is the approval. Open invitations are listed and can be revoked at any time; a revoked invitation cannot be reactivated.
The five staff roles
- Admin — everything a manager can do, plus org settings, member roles and the skills studio.
- Manager — runs the desk: every recruiter's mandates, pipeline and fees, and reassigns leads. Cannot administer members or author skills.
- Recruiter — runs mandates end to end: sourcing, evaluation, shortlists, client exports and outreach.
- Researcher — sourcing and evaluation, with their own delivery objectives. Cannot open a mandate, publish a shortlist or contact a candidate.
- Viewer — read-only across the org, excluding fees and revenue.
The members screen carries the full capability-by-role matrix, so what each role can touch is never a matter of memory. Client-side people (hiring managers, HR) are invited through their own separate door and never appear in this roster's role picker.
Suspending and restoring
An admin can suspend any member — their session stops working within one request, and nothing is deleted — and restore them later. Four refusals are built in, stated in words when you hit them:
- Founder accounts are managed by Mandate and cannot be changed here.
- Agent principals are managed from the operator console, not the members screen.
- You cannot suspend your own account.
- The organisation's last active admin cannot be suspended — that would lock the organisation.
The same last-admin protection applies to role changes: an organisation can never admin itself to zero.
What the agents do — and never do
Every agent judgment in Mandate runs as a principal: its own identity in the database, its own credential, its own kill switch, and its own name in the activity trail. No agent acts as you, and the editorial lines below are enforced in the database, not in a prompt. The full registry — every agent, its status, what it does and what stays human — is a page inside the product, at Agents.
The working set
Agents cover the loop in phases: understanding the role (intake, company intelligence, culture, calibration, role spec), discovering people (boolean search, and candidate search across your own pool; candidate research is not yet available to customers), evaluating them (CV parsing, evaluation, ranking, triangulation, and — where the add-on is switched on — executive intelligence), engaging them (outreach strategy, relationship, engagement, pre-screen), delivering (positioning, shortlist, search health, desk digest), and assisting you (the feedback interpreter and Mandy, the copilot). Each registry entry states its one job and its "stays human" line.
The lines that hold
- Decision support — never a recommendation. Scores are weighted evidence against your approved calibration; no agent output tells you whom to hire, and the interfaces are built so a reading cannot be mistaken for a verdict. The pre-screen conversation goes further: no verdict, no score, no pass — anywhere, by construction.
- No agent acts as you. Every agent write lands under the agent's own name in the activity trail; your name is never borrowed.
- No agent contacts anyone on its own. Outreach drafts wait for a person to send them; portal links are handed over by hand.
- Every agent can be stopped. Each principal has an independent kill switch, and a suspended agent's work simply does not happen — the product says so rather than pretending.
- You steer them in writing. Agents read your Skills Studio instructions on every run; steering lives with your team, not in a support ticket.
What the words on the screen mean
Where the product shows you example content before your own data exists, it says so — sample screens are labelled as sample, and marketing illustrations are labelled illustrative. Where an agent's reading appears, the sentence beside it is the doctrine of this whole handbook in six words: decision support, never a recommendation.
Every screen, and what you do on it
One guide per tab of the workspace, in the order the rail lists them. Inside the product the same text is a panel behind the ? in the top bar, on whichever screen you are standing on. Each says what the screen is for, what you do, how you know you are done — and, the section worth reading first, what it will not do.
Portfolio
Your first screen each morning. It answers four questions in reading order — what needs you, what changed, what is at risk, what to do next — so you can triage the whole workspace without opening a single mandate.
What you do here
- Read Needs you today first. It is the only list on the page carrying buttons; everything else is reference. If it is empty, no mandate is waiting on you.
- Scan mandate health. Health is a dot and a word — no gauge, no green-to-red ramp — because a search is not a score.
- Open whatever the first list sent you to, and do the work there.
What done looks like
Needs you today is empty, or everything left in it is deliberately waiting on someone else.
What this screen will not do
It will not tell you which candidate to advance, and it does not rank your mandates against each other. Two panels a new workspace shows in the sample — agent activity and next best action — are not built yet; they appear in the sample so you can see where the product is going, and are simply absent from a real workspace rather than faked. The "candidates" tile counts every candidate, not ones awaiting review: nothing tracks reviewed-versus-unreviewed, so a review count would be a guess.
Where this leads
Into a mandate, via Portfolio or Mandates. Everything else in Workspace reports on work that starts there.
Desk
The manager's screen: how the work is distributed across the people doing it. One row per recruiter, derived from rows that already exist.
What you do here
- Compare recruiters by mandates led, candidates in them, and placements owned.
- Look at last activity to find a mandate nobody has touched.
- Drill into any row to reach the underlying mandates or placements.
What done looks like
You can name who is overloaded, who has capacity, and which mandate has gone quiet — and you got there without asking anyone.
What this screen will not do
Nothing here is scored, weighted or invented. Every number is a count you could reproduce by hand, deliberately, because this is the screen most at risk of contradicting its own drill-downs. It makes no judgement about performance, and it will not tell you who is doing well — activity counts are not quality.
Money stays behind fees:read on its own pages; the desk is a persona boundary, not a money one.
Where this leads
Into Mandates or Placements for the row you were reading.
Analytics
The shape of the workspace over time: where candidates sit, how mandates are faring, and how fast work is arriving.
What you do here
- Read Candidates by pipeline stage to find where people pile up — a bulge at one stage is the bottleneck.
- Check Projects by health status for mandates drifting.
- Use Weekly velocity (candidates added per week, last 8 weeks) to see whether sourcing is keeping pace.
- Read Judgment calibration — how the system's verdicts compare with what actually happened to those candidates.
What done looks like
You can say where the pipeline is stuck and whether it is getting better or worse.
What this screen will not do
It will not diagnose the cause of a bottleneck, and it does not forecast. Judgment calibration only becomes meaningful once real outcomes have accumulated — with few placements behind it, treat it as early signal, not as a verdict on the system's accuracy.
Where this leads
To the stage that is blocked — usually Candidates filtered by that stage, or the mandate whose health slipped.
Objectives
The OKR board: goals the desk is measured by, and key results that measure themselves from work already recorded.
What you do here
- Set an objective and its key results.
- Watch progress accrue — key results read from placements, candidates and activity rather than from anyone typing a number in.
- Adjust targets when the quarter's shape changes.
What done looks like
Every objective has key results that move on their own, and none of them needs a human to update it.
What this screen will not do
It will not chase anyone, and it does not set targets for you. A key result it cannot derive from recorded work is one it cannot track.
Every active role can read this board — goals the desk is measured by are visible work. Writing is behind okrs:write, so the controls simply do not appear for a viewer. Financial key results are filtered by the database, not by this page: without fees:read the row never arrives, and the card renders what arrived.
Where this leads
To the work itself — Placements for revenue goals, Mandates or Candidates for pipeline ones.
Placements
The revenue book. It answers what did we bill this quarter as the headline, rather than as something you reach by filtering a list.
What you do here
- Read the quarter's number, then the four quarters behind it.
- Open the placements that make it up to interrogate any figure.
- Check what is booked but not yet earned.
- Record a new placement when a candidate is hired.
What done looks like
Every hire is recorded, and the quarter's number is one you would defend to the client.
What this screen will not do
It will not invoice anybody — that is Invoices — and it does not forecast future revenue.
What each role sees differs, honestly. A researcher or viewer holds org:read but not fees:read, so the database sends them placements and no fee lines. Rather than showing a revenue page reading zero — which is a lie, not a restriction — the page says so and shows the placements without the money columns. A researcher credited on a placement does see that placement's fee.
Where this leads
To Invoices, to bill what this screen recorded.
Invoices
Every invoice this organisation has drafted, issued or voided.
What you do here
- Draft an invoice against a placement.
- Review it, then issue it — which mints its number and freezes the document.
- Send it, or print it.
- Void one that should not have been issued.
What done looks like
Every billable placement has an issued invoice, and every issued invoice has been sent.
What this screen will not do
It will not take payment or reconcile one — there is no payment processing in the product, by choice. It does not chase overdue invoices, and it will not let you edit an issued document: an issued invoice is frozen on purpose, so what the client received and what you hold are the same paper. Correcting one means voiding and re-issuing.
An invoice keeps the details it was issued with even if the template behind it changes later. That is deliberate — the record outlives its template.
This screen is behind fees:read; the database would return nothing to anyone else regardless.
Where this leads
Back to Placements, or to the client in Clients.
Activity
Who changed what, newest first — including the agents, by name.
What you do here
- Scan the trail for what has happened since you last looked.
- Filter by type — mandates, placements, fees, client-facing — to answer a specific question.
- Use it to check an agent's work: every judgment an agent made is recorded under the agent's own name.
What done looks like
You can answer "who changed this, and when" without asking anyone.
What this screen will not do
It will not let you undo anything, and nothing here can be edited — the trail is a record, not a workspace.
Every row was written by the database at the moment of the actual change, not by the application remembering to mention it, so this is what happened rather than what was reported.
The page has no capability gate, because the honest answer for every role is "here is what you may see". The database filters the rows: a viewer sees placement and client-facing events, a recruiter also the money, an admin also member changes. The filter list is built from what you can actually see, so you are never offered a filter that returns nothing.
Where this leads
To whatever the row is about — every entry links to its subject.
Clients
The companies you recruit for, and every mandate run against each.
What you do here
- Add a client, or open one to read its record.
- See the mandates run for that company, and how they went.
- Keep contacts current — they are who invoices and portal invitations reach.
What done looks like
Every live mandate has a client record behind it with a contact you could actually email.
What this screen will not do
It will not research a company for you — that is the Company Intelligence agent, on the mandate. Resolving and linking a client is always the recruiter's act, never an agent's: the system will not decide that two similar company names are the same organisation.
The list is not paginated, deliberately: an agency works for tens of companies, not tens of thousands.
Where this leads
To a Mandate for that client, or to Invoices when there is something to bill.
Mandates
Every search in the workspace, live and closed.
What you do here
- Start a new mandate from a single line — "Head of IT Ops for RBC Capital Markets" is the whole input. Intake, Company Intelligence, Role Spec and Calibration activate on submit.
- Complete the mandate's five-step questionnaire — origin, must-haves, anti-patterns, stakeholders, priorities. This is the step that decides how well everything downstream works: calibration is derived from your answers.
- Open a mandate to work it — upload CVs, review, rank, shortlist.
What done looks like
The mandate has a final role spec, a calibration you agree with, and candidates arriving against it.
What this screen will not do
It will not fill a role for you. Every agent output on a mandate is decision support, never a recommendation — you review, edit and approve at each step that matters.
Three columns you see in the sample — candidate counts, tier-1 counts, day-of-search — are not on the real table: they need a per-mandate aggregate this page does not compute, and a column of dashes on every row is worse than no column. The mandate's own page has them.
Where this leads
Into the mandate, and from there to Candidates.
Candidates
Every candidate across every mandate, one screenful at a time.
What you do here
- Search and filter — by mandate, stage, tier or archetype.
- Open a candidate to read the evaluation, the evidence behind it, and the second opinion.
- Move someone through the pipeline as reality changes.
- Upload a CV to add someone to a mandate.
What done looks like
Every candidate has been read by a human, not just scored, and their stage matches where they actually are.
What this screen will not do
It will not decide who to submit. Three rules govern what you see, all about honesty rather than styling:
- Partial data is normal. A CV still parsing, or a candidate found but not yet scored, appears inline with honest placeholders — hiding a row until it is complete would make the count wrong and an upload look lost.
- Tier is a band, not a grade. Tier 1 carries the accent; tiers 2–4 stay neutral, and nothing is red. A tier-3 candidate is not a failure.
- Dedupe is visible. Where two records were merged into one person, the Network view states how — so a wrong merge can be found rather than silently trusted.
A CV upload takes roughly a minute and a half to parse, evaluate and be second-opinioned. It is not stuck.
Where this leads
To the candidate's own page, and on to the shortlist.
Pool search
Search every candidate you already hold, in plain language, across all mandates.
What you do here
- Narrow the pool with the structural filters — archetype, tier, pipeline stage.
- Describe who you want in a sentence: "Senior post-trade engineering leaders with FCA exposure who've shipped a T+1 migration."
- Read the matches, and open anyone worth pursuing.
What done looks like
You know whether the person you need is already in your pool before you go looking outside it.
What this screen will not do
It searches only candidates you already hold — it is not a sourcing tool and it does not reach the open web or LinkedIn.
The structural filters run first. If they match nobody, the agent never runs and the screen says so plainly rather than returning an empty result that looks like a judgement.
If the Candidate Search Agent has been suspended by an operator, this screen says that too, in those words, rather than quietly returning nothing.
Where this leads
To a candidate, or to the mandate you were staffing.
CV intake
A stack of CVs, one search, one pass. Pick the mandate, add up to twenty files, and parse them all in one go.
What you do here
- Choose the mandate these CVs belong to. The list says which searches are calibrated — an uncalibrated one has nothing to score against yet.
- Drop the CVs in, or choose files. PDF or DOCX, up to 10 MB each.
- Press Upload and parse. Each file is read in turn and becomes a candidate on that mandate; the row tells you which are queued, parsing, parsed or failed.
- Follow a parsed row straight to the person, or open the mandate to see them all.
What done looks like
Every CV in the batch reads PARSED, and the people are on the mandate with their profiles filled in.
What this screen will not do
It parses one file at a time, on purpose, and it does not roll back. Each CV is a real call to the CV Parsing Agent; a batch that failed atomically would throw away work already done. Whatever parsed, parsed — the rest is reported honestly beside its filename.
If the mandate itself refuses the upload — its scoring model no longer matches its finalised job spec — the batch stops on the first refusal rather than repeating the same failure twenty times.
It will not score anyone against a mandate that has not been calibrated. The CV is still read and the person still created, but there is nothing to judge them against until onboarding has run.
It will not parse the same person into this mandate twice — but what it can be sure of depends on when it looks. Before anything is uploaded it compares the file itself, so the same document twice in one batch, or a file already in this mandate, is skipped outright and costs nothing. Only after a CV has been read does it know whose it is: if that person is already here under the same email address or LinkedIn profile, the record just created is removed, the existing one is left exactly as it stands, and the row tells you the file you uploaded was not kept.
Where the only thing shared is a name and an employer, both records stay. Two people really do share a name at one large firm, so that pair is flagged for you on the candidate's own screen rather than merged. If they are one person, that flag is where you resolve it: Merge these records shows you both side by side, you choose which one to keep, and everything written about the other — notes, feedback, history — moves across. What cannot move is named before you confirm, and the merge cannot be undone. A record carrying a placement is never the one discarded.
The same file under a different mandate is not a duplicate at all. It is reported and parsed anyway — one person in two searches is two records, each scored against its own role.
Uploading records you as the person who brought each candidate in. That governs which CVs the reuse agent draws on when it makes suggestions for you; it does not hide them from anyone.
Where this leads
Into the mandate's Candidates list, where the ranking runs. To reuse someone you already have rather than upload them again, use the mandate's own Suggest from our pool.
Talent pool
People your firm already knew, brought in from its earlier records: on no mandate yet, and not read by any AI.
What you do here
- Find a person by name.
- Open them to read what the old records said: details, notes, history and the stored CV.
- Add them to a mandate. That is when their CV is read, against that mandate's criteria.
What done looks like
The people worth approaching for a search are on that search, and anyone who must not be contacted is marked do-not-contact.
What this screen will not do
It will not run AI on anyone. Importing fills the pool only. A person meets a model when a human adds them to a mandate, and not before.
Notes from the old records are shown to people only. They are never given to AI, including after the person joins a mandate.
"Privacy notice: not on record" means the old records did not say whether this person was told your firm holds their data. The first message you send them through Mandate carries the notice.
Where this leads
To a mandate's candidate list, or back to Import records in Settings to bring in more.
Network
Your candidates folded into people — one row per human, however many mandates they have appeared in.
What you do here
- Read the rollups: by archetype, by domain, top by average score, most versatile.
- Find people you have already assessed who fit something new.
- Maintain the relationship record — state, follow-ups, and do-not-contact.
What done looks like
Anyone you should not approach is marked do-not-contact, and the people worth re-approaching have a follow-up date.
What this screen will not do
It will not contact anyone on its own, and no agent may set or clear do-not-contact — that flag is only ever a human's.
"Top by average score" is a mean across every appearance, not a ranking of who is best: someone assessed once against an easy brief will outrank someone assessed five times against hard ones. Read it as a shortlist of who to look at again, not as a league table.
Where two records were merged into one person, the row states how, so a wrong merge can be found.
Where this leads
To the person's candidate record, or into a mandate you are staffing.
Executive Intelligence
The front door to the executive module: four counts, your five most recent executive searches, and a map of which parts of the due-diligence arc exist today.
What you do here
- Read the four counts — active searches, approved profiles, role templates, competencies — to see whether the module has anything running.
- Start a New Executive Search, or Start From Template if a role template already describes the job.
- Open a recent search to continue work already underway.
What done looks like
You have left for a search. Nothing is saved on this screen — it has no form and no action; it is a landing page.
What this screen will not do
It runs no agent. The line it carries — AI-generated decision support, reviewed and approved by a person — is a standing statement about the module, not a verdict about anything on the page.
Five tiles on the module map are deliberately dead, greyed and marked Phase 2: active interviews, risk reviews, final reports, executive advisors, and module settings. They are drawn so you can see the shape of the module, and they do nothing. Four of the live tiles — role success profiles, candidates, interview plans, assessments — all lead to the same list of searches; there is no separate screen behind each one yet.
If you cannot create a search, New Executive Search is simply absent rather than greyed out: it needs mandates:write, and a disabled button is a promise the product cannot keep. Everything else here is readable by every role.
Where this leads
Into an executive search, or to Role templates to start from a defined one.
Competencies
The library every executive score is measured against — four categories of competency, each with a definition and the observable evidence that counts for and against it.
What you do here
- Read a competency's definition before you trust a score that cites it.
- Read Evidence of strength and Evidence of concern to see what an assessor is actually looking for.
- Use the categories — leadership, functional, operating scale and stage, governance and stakeholders — to check a role is weighted across more than one.
What done looks like
You can say what a competency means without guessing, and you can trace any executive score back to a named, defined one.
What this screen will not do
Nothing here can be authored, edited or deleted — there is no form, no button and no create path anywhere in the product. The library ships with Mandate rather than being configured per organisation, which is why there is nothing to set up.
It has no search, no filter, no sort beyond alphabetical, and no detail page. It runs no agent and generates nothing: this is reference data, not advice.
If it ever renders empty, that is ours to fix, not yours — the guide says so, and so does the screen.
Where this leads
Back to Executive Intelligence. Weights against these competencies are set on Role templates.
Role templates
Reusable descriptions of an executive role — the intake answers and the competency weights — so the same kind of search does not get rebuilt from scratch each time.
What you do here
- Read the cards. Each says whether it is Global (ships with Mandate) or Org (yours), and how many competencies it weights.
- Use Template to start a search already prefilled from it — everything stays editable in the intake.
- Admins only: New Template, or Edit one of your own. Fill the optional intake defaults, then set a weight per competency — 0 leaves a competency out.
- Admins only: Delete a template your organisation no longer wants.
What done looks like
Your card appears in the list with an Org badge and a weight count, and a search started from it arrives prefilled.
What this screen will not do
Global templates are nobody's to edit or delete — not yours, not an admin's. The Edit and Delete buttons are not rendered for them at all, and the database refuses regardless. Giving your template the same key as a global one overrides it for your organisation; the form tells you when you are doing that.
A template that seeded a search cannot be deleted. Those searches point at it, and their record has to keep resolving.
No agent runs here. Weights are numbers a person types — nothing suggests them, and nothing scores anything on this screen.
There is no duplicate, no version history, no import or export, no preview of the prefilled intake, and no way to promote your template to the global library. Authoring needs skills:write, which only an admin holds; everyone else can read the list and use a template.
Where this leads
Into a new executive search, prefilled. The competencies you are weighting are defined under Competencies.
Agents
The whole workforce, named: every AI principal in the product, what each one judges, the line it is not allowed to cross, and whether it is running right now.
What you do here
- Read the counts at the top — how many principals exist, how many are active, how many are suspended.
- Find the agent behind a piece of work you are questioning, by group: understand, discover, evaluate, engage, deliver, assist.
- Read its Stays human line — the judgment that agent is forbidden to make on your behalf.
What done looks like
You can name the agent that produced a given judgment and say what it is not permitted to decide.
What this screen will not do
It cannot start, stop, pause or configure anything. There is not one button on it. The status chip reports what the database says; it does not operate it. Suspending a principal is a founder act from platform operations, and each agent's kill switch is independent of the other twenty-four.
It does not show which model an agent runs on, when it last ran, how often, or what it cost. Model assignment lives in Settings → Models.
An agent listed as not provisioned is documented here but has no account in this deployment — the page says so rather than implying it is working. For the same reason the header count (real accounts) can be smaller than the number of agents described below it.
The editorial lines quoted here are enforced in the database, not in a prompt: no agent acts as you, and every judgment lands in the trail under the agent's own name.
Where this leads
To Skills studio, which is where you actually steer these agents — they read your instructions on every run. To check an agent's work, read Activity.
Skills studio
Standing instructions that are spliced into the agents' prompts on every run — how this desk works, said once, instead of re-explained in every brief.
What you do here
- Read the active count: every active skill rides every agent run.
- New Skill — name it, choose its reach (Search for the whole organisation, Client for one client's work, Role for a single mandate), say when it should apply, and write the instruction.
- Watch the preview as you type — it shows the block as the agent will receive it.
- Pause a skill that is causing trouble rather than deleting it; Edit to refine the wording.
What done looks like
The instructions you keep repeating to the agents are written down once and active, and the next agent run obeys them.
What this screen will not do
Changes apply to the next agent run, never to work already done. An existing report keeps the prompt it was generated against — that is deliberate, so a past judgment stays explainable.
The preview is close but not exact: the real prompt adds framing the preview does not show, and on a live run every active skill in scope arrives together, not just the one you are editing. There is no view of what will actually fire as a set, and no dry run against a real candidate.
Trigger conditions are advisory. The model reads them and decides whether the skill applies — they are not a deterministic filter.
There is no version history, no diff, no undo on delete, and no count of how often a skill was used. Length limits (4,000 characters of instruction) are enforced when you save rather than shown as you type. Authoring needs skills:write, which only an admin holds; the trail records that a skill changed but never the text of the instruction.
Where this leads
To Agents, to see who reads these instructions and what each one is forbidden to decide.
Settings
Your own account, the organisation's identity, and the way into the screens that configure how the desk behaves.
What you do here
- Change your own name and password. Changing a password leaves your other signed-in sessions active.
- Admins: set Advisory mode — on, an evaluation leads with what to test at interview; off, it leads with the verdict. The verdict is never hidden either way.
- Follow the links to the screens that do the configuring: Members, Skills Studio, Models, and for founders, Platform ops.
- Read the rosters and, if any are open, the candidate erasure requests — stop working anyone listed there.
What done looks like
Your account details are right, advisory mode matches how your clients want to read an evaluation, and no erasure request is sitting unread.
What this screen will not do
It will not approve a pending person, suspend or restore anyone, change anyone's role, or edit the organisation's name or slug — those moved to Members and, for founders, platform operations. It cannot change your email address; ask Mandate.
It does not act on an erasure request. It shows you the request so you stop working that person; Mandate reviews and executes it.
Links you lack the capability for are not shown rather than shown and bounced — a link that bounces reads as a broken product rather than a restricted one. Invoice templates is not linked from here; it is reached from the Invoices screen.
Where this leads
Members to change who can do what, Skills studio to steer the agents, Models to choose which model answers.
Members
Who is on this desk, what each of them can do, and the only screen that changes it.
What you do here
- Read the roster. The meta line says how many accounts and how many admins, and warns you when there is only one admin left.
- Change someone's Role, then press Apply. It is deliberately not save-on-change — an arrow key through a select would otherwise demote whoever you scrolled past.
- Suspend someone who should no longer have access; Restore them later. Nothing is deleted either way.
- Issue Invitation for a new colleague — name, email, role — then hand over the link that appears.
- If an admin grant is waiting, Approve or Decline another admin's proposal, or Withdraw your own.
What done looks like
Everyone active holds the role their work actually needs, nobody who has left still has a session, and no proposal is sitting unanswered.
What this screen will not do
Mandate emails the invitation to the address you entered, and shows Email sent to … at … or Not sent: with the reason and Send again. The link stays here to copy as well. It is single-use, expires in fourteen days, and whoever redeems it joins active with the role you chose. A spent or expired link shows one honest dead end — it does not say which kind of dead, on purpose.
Making someone an admin takes two admins. You propose it; a different admin approves. Your own proposal shows no Approve button at all rather than one that cannot work. The rule engages once there are two active admins — below that, a grant lands immediately. Removing an admin stays a one-person act.
It refuses to lock you out: you cannot suspend yourself, you cannot suspend the last active admin, and founder and agent accounts carry no controls here at all — agents are operated from platform operations.
There is no remove-a-member (suspend is the removal), no resend, no way to see a previously-issued link again, no bulk action, no search, and no client-side roles — those arrive by invitation from a client screen. Approving a pending account is a founder act elsewhere; pending people appear here, as viewers, until then.
This screen needs org:manage, which only an admin holds.
Where this leads
Back to Settings. New accounts join as viewer — promote them here.
Read it all? Ask for a seat.
Access requests are reviewed by a person, and approval arrives as a single-use invitation link — the handbook's first two chapters, in practice.
Request access