How the old work quiz works

The interactive version administers this instrument and generates the output. This document specifies it: the question bank with the purpose of every item, the scoring constants, the phase assignment rules, the eight divergence checks, the thirteen prerequisite triggers, and the structure of the brief. Read it if you need to defend the output to somebody, adapt the instrument to a domain it does not currently handle, or satisfy yourself that the ordering is derived rather than asserted.

1.1What it is for

Grok Bot sets itself up badly from a description of what somebody wants and well from a description of how they work. The gap between those two is where almost all first-week failure lives. This instrument exists to produce the second thing.

Its output is a single document a person pastes into their first Bot before creating any others. That document carries four things a Bot cannot infer and will otherwise invent: the operator's context and vocabulary, the boundaries in the operator's own words, the order in which capability should be handed over, and an instruction set telling the Bot to propose rather than act.

The order is the part that justifies the instrument. Everything else could be captured in a conversation. Sequencing cannot, because the correct sequence depends on properties of each job that people systematically misjudge about their own work.

1.2What it refuses to do

It will not rank by importance. The respondent's stated priority is collected in two places and used only to detect disagreement. It never influences the build order. A job that would transform the business but cannot be verified quickly still sorts below a trivial job that can, because the purpose of the first build is not value. It is evidence.

It will not raise the ceiling on request. Every generated brief starts at draft-only for four weeks regardless of the autonomy answer, and every irreversible action stays gated permanently. When a respondent asks for sending or transacting, the instrument records the request, fires a divergence explaining the refusal, and overrides it in the brief. This is a deliberate design position rather than an oversight: approvals in this product gate proposed actions and do not reverse completed ones, and this earlier assessment used a fixed review policy. See the current security records for available Enterprise audit controls.

2.1Why order beats preference

People abandon agent tools in week two, and almost never because the tool was incapable. They abandon them because they handed over something important on day two, it went eighty percent right unsupervised, and they had no way to tell whether that was competence or luck. Having no way to tell is the actual failure. The work was never the problem.

That failure is entirely predictable from properties of the task chosen, and those properties are knowable in advance. A job you can check in thirty seconds produces evidence every time it runs. A job that takes an hour to verify produces evidence roughly never, because nobody spends the hour. Trust accumulates from the first kind and cannot accumulate from the second, whatever the second is worth.

So the instrument sorts by evidence yield rather than by value, and tells the respondent plainly that it is doing so. The most valuable job in the set frequently sorts fourth. That is not an error to apologise for; it is the finding.

2.2The three dimensions that decide a phase

Three further dimensions (frequency, duration and whether an example of good output exists) decide rank within a phase but never move a job between phases. This separation is the core of the design: value competes only against comparable risk.

3.1Constants

Six dimensions are collected per job. Four of them carry numeric weights:

The two remaining dimensions are duration (minutes per run, entered as a number and floored at one) and readiness (whether an example of the job done well exists, worth one or zero).

3.2The formula

The readiness multiplier is deliberately small. Its job is to break ties in favour of work the operator can already demonstrate, not to reorder the list. A job worth two hundred hours a year with no example still outranks a job worth forty with one.

Frequency is expressed as runs per month, so an ad-hoc job is scored at 0.5 rather than zero. Ad-hoc work is real work; it simply does not compound, and the low weight reflects that a job which happens unpredictably cannot anchor a routine.

3.3Phase assignment

Evaluated in order. The first matching rule wins.

Phase 0 is not produced by this rule set. It is assembled separately from the prerequisite triggers in section 5.2 and contains no jobs at all, only conditions that must be true before any Bot is created.

3.4A worked example

Six jobs from a two-person trade business, scored as the respondent entered them.

Quote drafting is the largest single cost in the set at 154 hours a year, roughly forty percent more than anything else. It sorts last, because a customer sees the output. Morning inbox triage, worth less, sorts first because it is daily, checkable at a glance and harmless when wrong.

The respondent listed chasing supplier invoices first, which is a reliable signal that it is the most irritating item rather than the most valuable one. It scores into Phase 3. The instrument reports all three of these observations as disagreements rather than quietly reordering, because the respondent needs to understand the reasoning to accept the sequence.

4.1The question bank

Sections 00 to 08 are asked of everyone. Sections M1 to M7 open conditionally, and the gates are listed in section 5.3. Seventy-two of the 86 items are required; the optional ones are those where a blank answer is itself informative.

00 · Before anything else 5 questions

Gate. Any of these can stop the exercise or add a Phase 0 prerequisite. Asked first so nobody spends half an hour planning something they cannot run.

01 · You 10 questions

Operator assessment. Predicts review behaviour and supplies the raw material for voice. Four of these are behavioural rather than opinion questions.

02 · Your actual week 8 questions

Evidence collection. Surfaces work the respondent would not think to list, and produces the reported-hours figure used in the under-listing check.

03 · The business 8 questions

Context and routing. Sets the vocabulary the brief writes in, seeds the task suggestions, and captures the stated priority used in the divergence check.

04 · Your systems 6 questions

Feasibility and cost. Determines where each job sits on the connection ladder, which is the single largest driver of both reliability and spend.

05 · The work itself 3 questions

The engine. Everything else contextualises; this section produces the ordering.

06 · Voice and standards 5 questions

Output quality. Decides whether drafts read like the respondent or like a model.

07 · Boundaries 6 questions

Boundary capture, in their own words rather than a template. Produces the Red list in the brief.

08 · Constraints and appetite 8 questions

Constraints and module gates. Six of these eight questions open or suppress a module.

M1 · Regulated or confidential data 5 questions

Opens when regulated data is in scope. Changes the data boundary in the brief and can add a Phase 0 item requiring qualified review.

M2 · Multiple clients or accounts 4 questions

Opens for many-client or agency shapes. Addresses the fact that Grok Bot cannot enforce separation between clients.

M3 · Other people 4 questions

Opens above one person. Determines whether the brief must survive being handed to somebody else.

M4 · What you already run 3 questions

Opens when something already runs. Prevents the plan duplicating or contradicting existing automation.

M5 · Software and code 4 questions

Opens when software work is in scope. Sets the outer-loop configuration and checks for a staging environment.

M6 · Physical operations 3 questions

Opens when the work touches the physical world, where failures cost journeys and appointments rather than tokens.

M7 · Cost control 4 questions

Opens when cost is a stated concern. Ten more questions on the product mechanic that surprises people most.

4.2The behavioural questions

Ten items do not ask what the respondent thinks. They ask what the respondent did, because stated intent is a poor predictor of behaviour and people are least accurate about exactly the things this instrument needs to know. Each is written to defeat a specific reporting bias.

5.1Divergence rules

These compare stated preference against scored evidence and report every conflict. They are the reason the output is persuasive rather than merely correct: a sequence somebody does not understand is a sequence they abandon.

Rules 6 and 8 are the two that most often change what somebody does. Rule 6 because under-listing is near universal and invisible without the arithmetic; rule 8 because it is the only place the instrument openly overrides a direct request.

5.2Prerequisite rules

Thirteen conditions produce Phase 0 items. None involves a Bot. Any of them left unresolved either blocks the product or removes a control the rest of the plan assumes.

5.3Module gates

Gates are re-evaluated on every relevant answer, so modules appear and disappear live rather than at a submit step. Progress is calculated against active sections only, which keeps the completion figure honest for a respondent who opens four modules.

6.1Anatomy of the brief

The generated document is written to be read by a Bot rather than by a person. That changes the register: it is imperative, it front-loads constraints, and it repeats the things a model is most likely to drift from.

7.1Adapting it

Adding a question. Append an object to the question bank with a section id, a unique id, a type, and whether it is required. Rendering, progress and validation pick it up with no further work. If the answer should appear in the brief, add the line to the brief builder; if it should change the order, it belongs in the scoring instead, which is a much bigger decision.

Adding a module. Add the section, add its questions, and add one gate function. Gates receive the whole answer set and return a boolean, so a module can depend on any combination of earlier answers.

Changing the weights. The four constant tables are the only place scoring lives. Raising a frequency weight changes rank inside phases; changing a verifiability or reversibility value changes which phase jobs land in, which is a change to the instrument's position rather than its calibration. Do that deliberately.

Adapting to a domain. The task seed suggestions are keyed to business shape and are the cheapest thing to localise. Adding a shape means adding its seed list and, usually, one module.

7.2Known limits

1. It trusts the respondent's scoring. Somebody who marks everything as checkable in thirty seconds gets a fast, wrong sequence. The behavioural questions catch some of this and the divergence rules catch more, but the instrument has no independent evidence and does not pretend to.
2. It has no view on whether a job should exist. A task that is pure waste scores the same as one that is essential. Automating the first is worse than not automating anything, and the instrument will not tell you which is which.
3. Duration is self-reported and usually understated. People estimate from the good runs. The hours figures should be read as a lower bound.
4. It cannot see interdependence. Jobs are scored independently, so two tasks that only make sense together may be split across phases. A human reading the sequence should look for that.
5. Frequency weights assume a stable week. Strongly seasonal work is scored at its average, which understates the peak and overstates the trough. The seasonality question is collected but does not currently enter the formula.
6. It is a snapshot of a product that changes weekly. Every constraint the brief cites was verified on 2 September 2026. The four product facts in the closing section are the ones most likely to move, and the ones whose movement would most change the output.

COMPANION TO THE GROK BOT INTAKE · Built on The Grok Bot Operator's Manual, verified against SpaceXAI documentation and Cursor billing pages on 2 September 2026. The question bank and scoring tables in this document are generated from the tool's own source, so they cannot drift apart. Nothing here is financial, legal or security advice.