What Elevate Coach keeps, and what Claude sees

An internal OCH tool for improvement projects. Nothing in it is ever used to judge any person. Print this page if you need it on paper.

Anonymous notes and votes

On an anonymous exercise, nothing links a note or a vote to the phone that sent it, not even for the lead. A first name, if you give one, is kept only while the session runs and erased when it ends. Seats are roles (for example, Maintenance Worker), never names.

Names in the project's own records

A real project has real people in it. The Record, the documents, the evidence, the approver's own words, the parking lot, the stamps, the decision rights and the sign-offs may name staff and their positions ("Priya Raman, Director, approved the charter"). A name may stand there, with three exceptions: an owner (of an action, a control or a document) is a role, never a name; a Record line the room also hears or sees (why an option was withdrawn or left parked, the push round's question, the go/no-go call's conditions) is checked the room's way; and while a session runs, the first names its phones gave and the words its lead holds are held there too. A name the name check finds is never sent to Claude: a line with one isn't sent, and the names the tool knows (the sponsor, the process owner, the project's members) are swapped for their role. A name the check can't recognise, such as a surname on its own that isn't one of those, or a first name that isn't on its list, may go. Names never reach the room: anything the room sees (the room's own notes, a session's title, the run sheet's questions and wording, the criteria, the options, their champions' pitches and the room's question about one, a changed option's wording at the gate review, the go/no-go lines and an example shown on the TV) is checked the room's way, so a name there, or a line that points at a one-holder title, is held or refused before the TV or a phone shows it; the go/no-go call's conditions stay on the Coach and the Record. The say-back script reads out only the last session's decisions that pass the room's check, and leaves any other for the lead to say in their own words. What the project's own records still never keep: a tenant's or a household's details (a first name on the check's list in the same sentence as a tenant word, unless it is one of the project's own people, though a surname on its own, such as "The tenant Gagnon refused entry", isn't recognised; a name before a family word, such as "Maria's kids"; a name with Mr, Mrs, Ms, Dr, Mme, Mlle or M.; an email address, a phone number, a unit or an address), someone's own circumstances told beside their name or beside a title only one person at OCH holds ("Our CEO is going through a divorce"), a colleague's included, a line that judges the work of a named person, or of a title only one person at OCH holds, said of them ("Dave is the bottleneck", "The CEO is too slow", "Dave takes three times longer": it is held, and the lead says it about the process or the role instead, such as "one MA's queue takes three times longer"; a name that only reports a finding, such as "Priya found the bottleneck is the approval step", files), a tenant's own circumstances, a reference number, and the names on the project's Names to hold.

What goes to Claude

Claude is an AI model made by Anthropic. The lead or co-lead asks it for help with a button; it never acts on its own. Each time, the server sends only what that one job needs, and never who wrote anything:

for an example: the step, the room's question, the notes on the wall that never tripped the name check, and the seat roles (never who sits in them);
for an idea: the room's question, its notes that never tripped the name check, the options, the criteria and the seat roles;
for suggested groups: the notes that never tripped the name check, each by a code (n1, n2 and so on), never who wrote it, and the room's question, if it passes the check;
for "What do I do now?": the block, the counts and the clock, and which of the five taps the lead pressed, never what anyone wrote;
for tailoring a session: the run sheet (the session's plan, block by block), the seat roles, the headcount and the time available;
for a document, a "what changed" draft or a v4 answer (the certification questions each phase must answer): the project's Record and its evidence as it stands (a figure corrected or withdrawn isn't sent, and the Record's note of the change carries no figure, not even from the row's label; its correction is. The reason the lead types for a correction or a withdrawal goes on the Record and is sent as written, so it should say why without the old figure), the tool's own lists (seats, decision rights, criteria, options with their champions' pitches, the parking lot, session titles and counts; an option whose pitch can't be sent goes without it) and, for a document, the room's notes that never tripped the name check, with their seat role unless the round was anonymous, each by a short code such as R12. A "what changed" draft also gets the last version's words; a v4 answer gets only Record entries;
for an example or tailoring: the sections of the charter and the SIPOC (the map of suppliers, inputs, process, outputs and customers) that pass the name check.

The room's notes, and the lead's words in the Record, the documents and the run sheet, go only if they pass the name check; so do the project's title and the room's question. A title that doesn't goes as "(held back)"; a question that doesn't means Claude isn't asked at all. The names the tool knows are swapped for a role in everything sent: the sponsor, the process owner, a project member, the owner. The playbook's own words (a technique card's name, purpose, script, cues and warning signs) go as written: they are the tool's reviewed text, not anyone's notes.

What never goes to Claude

A line that names someone. When the name check finds a name in it (a first name, "Mr" or "Mrs" and a name, an email address, or a first name a phone gave in this session), the line isn't sent at all, not even with the name blanked out. "Claude" on its own is read as this tool's name, not a person's ("Claude Tremblay" and "Jean-Claude" are people), and a phone's first name or a name to hold is held whatever it is. The names the tool already knows (the sponsor, the process owner, project members and document owners) are swapped for their role wherever they appear, a surname on its own included. A sponsor or process owner typed as a title, such as "Director, Tenant Experience Operations", is swapped as the whole title, so its words ("tenant", "operations") stay everywhere else.
A reference number (a work order or account number, seven digits or more) or a tenant's own circumstances (their health, a death or a divorce in the family, an arrest, their benefits or immigration status, for example). A line with one isn't sent at all. How OCH handles tenancies as a category isn't a circumstance: "Tenants in arrears get the notice late" can go, and so can non-access, not home, refused entry, turnover, a transfer or a subsidy, unless the line also points at one household (a name, a unit, an address, a phone number, an email address or a reference number). In a line with a tenant word, a circumstance is the tenant's in whatever sentence it stands. OCH's own department, Tenant Experience, isn't a tenant, and Lean words such as "effort-blind", "a blind vote" or "families of options" aren't a circumstance. A circumstance said of a staff member right beside the staff word ("a sick worker's jobs", "the worker's surgery", "the MW was off sick") isn't the tenant's; with anything else between them, a comma or "she", it is.
Tenant details: an address, a unit number or a phone number. A line with one isn't sent at all.
A role only one person at OCH holds (the CEO, the rest of the C-suite such as the CFO or any Chief … Officer, a Director, or a seat the lead marks "Only one at OCH"), written so that it points at them ("our CEO"), even when the lead kept it for the room with one tap, or named it as who does the job in a risk round, a go/no-go check, Who stops what or a document field that asks who does it, where the tool doesn't hold it. A job title many people hold, such as the TEM or an MA, isn't one. Nor are the project's own roles (the sponsor, the process owner, the approver, the lead, a co-lead), even when a seat is named for one: they are roles in the project, not a person's job at OCH, so "the sponsor chose option 3" goes as it is. The names of the people who hold them never go: they are swapped for the role, as above.
A word the lead holds, while the session runs (a note with one is held at the reveal and never goes after). At End the lead is asked about each one, ticked to keep: a ticked word goes onto the project's Names to hold, below, and stays held (a phrase such as "the new rc" as the whole phrase, never word by word); a ticked word that doesn't fit the list's rule (2 to 60 letters) is skipped, and the end screen names it; an unticked one is erased with the session.
A name on the project's Names to hold list, and the list itself. The lead or co-lead types these names on Project home, or keeps a session's held words there at End, so a name stays held after the session ends and the phones' first names are erased. A line with one isn't sent at all, and the name check holds it everywhere on the project. Only the lead and co-lead see the list; it never reaches the TV or a phone.
A note the name check held: not even once the lead has shown it to the room, and never the original of a note the lead rewrote.
Who wrote anything, and the first names phones give.

How you can tell what Claude wrote

AI-written example, not from this room: An example Claude wrote for the room's question, shown to the lead first ("AI-written, from this project's own words. Only you see it first." and "Written for this step by the Coach's AI."). The lead chooses: show it on the TV, where this label goes with it; say it out loud; or ask for another angle. Project home counts each session's examples ("Examples: 3 (2 AI-written)"), and the Record line says "AI-written".
Example from the playbook, not from this room.: An example from the technique card, used when Claude isn't available ("From the card, another setting: the AI isn't available right now."). It is shown and logged the same way.
Waiting for the lead: What Claude has written but the lead hasn't taken yet carries its own banner, and nothing is saved or changed until the lead accepts it: "AI-drafted by Claude, not saved: check every line" on a document; "AI-written by Claude, not applied yet: check each change" in Prepare, with each change beside what it replaces ("Question for the room, Claude's", "Order, Claude's"); and "AI-written. Draft: rewrite in your own words" on a v4 draft.
AI-drafted, accepted by the lead: A document section Claude drafted, accepted word for word.
AI-drafted, lightly edited: A section Claude drafted that the lead has changed, with less than half of Claude's words gone. Reordering Claude's sentences keeps the mark, and adding words never clears it: it goes only once at least half of Claude's words have been removed or replaced. Viewers see the mark on a released document, never Claude's original wording.
AI-drafted lines added: A document section whose words were already saved, with lines Claude drafted in "Draft what changed" added below them. The section keeps where it began: one the lead wrote stays the lead's, and its version history never says it began as an AI draft; one Claude drafted still does. The mark goes once at least half of Claude's added words have been removed or replaced; the section's own words then carry their own mark, if any.
Began as an AI draft: In a document's version history, on every section that started as Claude's, whatever became of it since.
Draft: rewrite in your own words: A v4 certification answer that is still Claude's draft word for word. Claude drafts only from the Record entries that answer the question. The mark stays until the lead has rewritten at least half of Claude's words, and an answer that keeps half or more of Claude's own words is marked even when it was copied in by hand.
Draft, lightly edited: rewrite in your own words: That draft, changed, with less than half of Claude's words gone.
From the Coach's AI, for you only: Claude's answer to "What do I do now?": at most three sentences, on the lead's screen only. A sentence that repeats four or more of the room's words in a row, or that would name someone, is set aside and the rest of the answer is kept; nothing of it goes on the Record.
Suggested by the Coach's AI, for you only: the room names each group: Suggested groups for a board of notes, on the lead's screen only. The room names every group it keeps; Claude's suggested name is a hint for the lead and never reaches the TV.
Suggested twice: the Coach's AI suggested this merge too: On the lead's screen, under a board of notes, the tool's own check offers pairs of notes worded nearly the same, near-duplicates and misspellings ("These may be the same point (checked by the tool, not Claude): the room decides"). That check runs on the server and sends nothing to Claude. When Claude's suggested groups name the same pair, it is shown once, with this label. The room decides every merge and group either way.
Tailored by Claude, accepted by you: In Prepare, a block whose question, participant wording, speaking notes or cues Claude reworded and the lead accepted. Tailoring changes words and order, and may suggest adding or dropping a card, never how a block runs, and never the room habits: the setup or opening stays first, the close stays last, and no vote or ranking comes before the ideas. If the lead rewrites Claude's question, the mark goes. A reworded question needs trying on two people again.
Added by Claude: In Prepare, a technique card Claude suggested adding and the lead accepted. The mark stays whatever the lead rewrites.
AI-written idea: you're stepping out of role: An idea the lead asked Claude for once the room's own ideas are up, shown to the lead only. The lead says it in their own words, and the room keeps it or drops it.
AI idea: An option the room kept from Claude's idea. It is never counted as the lead's own. The mark shows on the lead's screens, on Project home ("(AI idea)") and in the gate pack; never on the TV or a phone.

What goes on the Record by itself

When the lead steps out of role, the tool writes the Record line itself, with the time and the step: an example shown on the TV or said out loud, with its label; an option of the lead's own; and Claude's idea offered to the room, kept or dropped. A wrong line is corrected by a note beside it, never erased.

A signature (the transfer of responsibility, the charter's approvals, the kickoff and the Control Plan's handover) is made by the person themselves, signed in, with their password, and dated by the tool. The lead chooses, from the project's members, who signs as the process owner and who as the sponsor; only that person's own sign-in can sign for them, and nobody can type a signature into the document. Neither can be the lead or a co-lead: the people who hand the work over never approve their own handover. A sign-off is on one version: a new version means each signs again, and a sign-off stops counting once someone else is chosen to sign for that part, or its signer leaves the project or starts leading it; someone chosen again signs again. The password is checked, never kept, logged or shown; five wrong ones stop that person signing for 15 minutes. It confirms that the person signed in means to sign; it can't tell who is at an unlocked screen. Once the project is closed, its sign-offs stand as they did at close: nothing more is signed, a version saved after close voids none, and who signs, and the place on the project of anyone who signed, no longer change. The Record says who signed, which version and when.

When a project touches a tool built with Claude's help, its kickoff run sheet carries a line the lead reads aloud and tailoring can't remove, and the phones show it before anyone joins:

Before we start, one thing to declare: this project touches [the tool's name], which was built with help from Claude, the AI this tool uses to help me prepare. If that worries anyone, say so now and I'll note it on the Record.

What reaches the TV

Only what the room made or the lead chose: an example the lead put up, with its label; a question in Claude's wording that the lead accepted, which is then the lead's question and carries no AI label; a group the room kept, named by the room; and an idea the room kept, as an option on the board without its AI mark. Everything else Claude writes stays on the lead's screens. A phone shows the same question the TV does, in the lead's accepted wording, and an option the room kept, never with an AI mark; nothing else of Claude's reaches a phone.

Who sees Claude's calls

Every call to Claude is kept with its project for as long as the project is: what was sent (already stripped as above), what came back, and what it cost. The lead and co-lead see a list on Project home of what each call was for, whether Claude answered or was set aside (and, when part of an answer was set aside, why, never its words), and its cost, never what was sent; viewers don't see the list. What was sent and what came back can be read only in the tool's database, by whoever runs the server.

Nothing Claude writes goes on the Record, into a document, a run sheet or the options board until the lead accepts it or uses it in the room. A change to a document's scope or decision rights, Claude's or the lead's, also waits for the approver: it is released only once an approval on the Record covers it: one that cites that exact version (for example "charter, version 3"), or the Record entry an unchanged Claude-drafted change rests on. A first version waits too, except a charter released while the project is still in Define.

If Claude isn't there

With no key, no answer in time, a refusal, an answer that doesn't check out, or a usage limit reached, every feature says so and carries on without it: the playbook's examples, the run sheet as planned, the rules' answer to "What do I do now?", the room's own grouping, the first layer of each document (the lines the tool draws straight from the Record, with their sources), and the Record entries under each v4 question. Asking for an idea says Claude isn't available. When part of an answer doesn't check out (a suggested group, a change to the run sheet, a sentence of "What do I do now?"), only that part is set aside and the rest is offered.

What happens when the project closes

A closed project stays in the app, as it is today. The plan, not built yet: at close, the whole project, including the Record, the documents and the list of Claude's calls, is archived to OCH's SharePoint as the permanent record.

Who develops and runs this tool: About Elevate Coach