Coaching
What was agreed, and whether it happened.
Maddox reads the coaching record: the commitments made at a one on one and their current status, the items on a development plan, and the evaluations and meeting grades a manager has already recorded. It reports what a human filed. It does not form its own opinion of anyone.
How it works
Reading the coaching record
Three kinds of question, all answered from what was already written down.
Commitments
What was promised at the last one on one, what is still open, and what is overdue. This is self-scoped, so a rep can ask about their own commitments without any management permission.
Development plans
The items on a plan, each carrying whether it has been confirmed by a manager. An unconfirmed suggestion is marked as a proposal rather than presented as a plan.
Recorded evaluations
The latest stored evaluation of a rep, verbatim, and their meeting grades. Where nobody has evaluated someone, that absence is stated plainly rather than softened.
In the product
The record, not a re-interpretation of it.
The mechanism
The coaching tools, and the distinctions they will not collapse
Coaching is the largest area in the assistant and the one where a fluent summary does the most damage, because every record in it is about a person. Each tool below is built around a distinction the schema went to the trouble of keeping.
A proposal and a ruling are two separate columns
Every commitment from a one-on-one carries two verdicts: what a model proposed, and what a manager actually ruled, with the name of the person who ruled it and when. Both use the same vocabulary, so an assistant that merged them would present a proposal as a decision with nothing downstream able to tell the difference. Maddox is instructed never to merge them, and where they disagree, to report the disagreement rather than pick the tidier one.
An unconfirmed item is a proposal, not a plan
Development-plan items carry a confirmation state saying whether a human has agreed to the item, and an item nobody has agreed to must not be described as a plan. Each item also carries where its baseline came from, and when there is no baseline it says why there is none. That absence is never reported as a baseline of zero, because “you have not improved” and “we could not measure you” lead to opposite conversations and only one of them is a fair thing to say to somebody.
An evaluation has four states and only one of them has a narrative
A stored evaluation is quoted verbatim. Where none exists the tool returns that nobody has ever run one, and offers to start a real evaluation through the real pipeline — and the assistant is told plainly not to assess the rep itself, not to infer anything from their numbers, and not to soften the absence. The other two states are an evaluation currently running and one that failed, which carries its error and is a reason to try again. A generated opinion that looks on screen like a recorded one is the exact artefact this design exists to prevent.
Who is slipping reports findings, and says how many people it looked at
It returns reps with a recorded development-plan finding on a dimension — somebody wrote it down. It does not look at a metric, notice a dip and decide something about a person, so a rep with a weak number and no recorded finding does not appear. The result carries how many people were in scope, which stops three findings being read as three out of three, and the assistant is told to report that denominator rather than smooth it over.
Playbook adherence counts gradings, not silences
The process record shows which playbook steps were graded done, partial or missed across deal stages and meetings, with the grader’s own evidence quote attached. It counts recorded gradings only. A step with no row against it was not graded, which is not the same as skipped, and the tool says so rather than letting the gap read as a failure.
Who it is for
For both sides of the conversation
A rep asking what they owe, and a manager asking who is slipping, are reading the same ledger from two directions.
- Reps, before a one on one, checking what they said they would do
- Managers, checking which reps have a recorded finding on a dimension
- Anyone who wants the coaching conversation to start from the record
Related
Questions
The things people actually ask.
Will Maddox tell my manager what I asked?
No. Conversations are private to the person who had them.
Can it tell me who on my team is struggling?
It can tell you which reps have a recorded development-plan finding on a dimension, and it reports how many people it considered so you can see the denominator. It will not analyse your team and volunteer an opinion.
Does it see coaching notes I have not shared?
No. Sharing state is respected, and the manager-only projections of a one on one are not readable through the assistant.
Why does it refuse to assess a rep when I ask it to?
Because an assessment that nobody accountable has signed is not evidence, and a coaching record that mixes the two stops being useful. It will show you what the manager recorded, including when the answer is that nobody has.
Start the one on one from the record.
The conversation gets shorter and better when both people already agree on what was said last time.