| name | mapping-session |
| description | Run a live mapping session — interview a team one activity at a time and write their words into ai-sdlc YAML while the page redraws beside them. |
| disable-model-invocation | true |
mapping-session
You are the editor's hand in a mapping session. A coach and a team are talking; a browser is projected beside them. Turn what the team says into YAML fast enough that the page redraws while they are still describing it.
Before the first question
-
Know which folder you are writing. It is the argument, or ask. If it does not exist yet, run ai-sdlc new <dir>. Done when team.yaml exists in that folder.
Every YAML file you create during the session opens with the schema line for its own shelf, so the coach's editor offers the fields and flags a typo as it is typed. ai-sdlc new writes it into the files it emits; a catalog file you add later needs it from you:
Swap the file name for the shelf you are writing — team, artifacts, harnesses, events, tools, or process for anything under processes/.
-
Start the renderer yourself. Run ai-sdlc serve <dir> in a background shell so it outlives the command — it runs until the session ends, and a foreground call would hang the conversation. Wait for the URL it prints, open it, and hand it over:
The page is live at the URL the server just printed — put it on the shared screen where the team can see it.
Done when the server has printed its URL. A taken port just means passing --port and reading out the new one. The redraw is the point, so nothing else starts until the page is up.
-
Read the sdlc-ontology skill, and its references/ontology-cheatsheet.md — the whole model on one page, and the reference you consult all session.
The loop
For each thing the team tells you: listen → narrate → write one small edit → redraw → point at what changed. One pass is done when ai-sdlc status <dir> names no new dangling id.
Narrate
Before every write, say what you are recording, in the model's vocabulary:
You said the spec gets signed off before anyone codes. I'm recording that as an activity Sign Off Spec, in stage Define, owned by Product Lead, consuming Spec and producing Approval. The arrow you'll see from there to Build isn't something I drew — it falls out of those two artifact names matching.
Narrating is the step that teaches. The team learns the ontology by watching their own sentences become it, and a session that ends with a correct document and an author who cannot maintain it has failed.
Keep the redraw hot
One activity per edit. Assembling a whole file in one pass buys a sixty-second silence, and that is where the room stops watching the page and starts checking phones.
What maps to what
| They say | You write |
|---|
| "then QA gets it" | a role, if new — and an activity QA performs |
| "we write a design doc" | an artifact, and produces: on the activity that writes it |
| "engineering picks it up from there" | nothing new — the arrow is derived from consumes: matching produces: |
| "that's where it always gets stuck" | constraint: on the process — the artifact it sticks on, and a note: in their words |
| "the three of us do that together" | one activity, with every one of them in roles: |
| "that box is really a whole process of its own" | activities: nested inside that activity — they inherit its stage |
| "Copilot writes most of those tests" | a tool in the catalog — plus the harness it runs in, if that runtime is new — and tooling: on that activity with a level |
| "sometimes it turns out bigger than we thought" | an event, plus a recommends: bound to it |
| "we do that by hand" | the activity, then the gap question below — their answer decides open: versus unclaimed |
When the team gives you the reason — "otherwise it bounces back at us" — that is why:. It reads in the drawer, and it is the field that makes the document worth returning to. A why: you are about to write yourself instead of quoting is the signal to read sdlc-ontology's references/worked-example.md, which shows the difference between an argument and a restated name.
Ask once about tooling — then once about the gap
Ask neutrally, once — "anything doing that work for you today?" — take the first answer and move on. A yes is tooling:.
A no leaves two different facts, and only the team can say which one it is. Ask the follow-up in their words — "would you want something here?" — because the answer decides which state the activity is in:
- Yes, we want something here. That is
open:, and the sentence they just said about what would help is need:. Write their sentence, not a product name: an open slot names the work that would be taken off them. This is the roadmap, so it is worth one extra beat to get the sentence right.
- No, this stays ours. Write neither field. The activity is unclaimed, and the document says nothing about it — which is exactly right, because the team said nothing about it.
When a level, or a fill-versus-recommendation call, will not settle, read sdlc-ontology's references/worked-example.md — it argues each of these calls against a real file.
An activity becomes open: on the team's own yes, and on nothing else. Silence leaves it unclaimed: if the team has not asked for anything, the page speaks for them only by staying quiet too.
The gap between what a team blueprints and what they have filled is the most honest thing this document produces. A session that talks a team into aspirational fills — or into open slots they never wanted — destroys the signal the product exists to show.
Record a tool the moment someone volunteers one, and open tooling questions only about activities the team has stopped arguing about.
After each write
Run ai-sdlc status <dir> and fix every id it lists as referenced but not in a catalog before the next question. That is your most likely mistake — an activity naming a role or artifact you never added to a catalog — and it is mostly silent; the cheatsheet's Verifying section says which kinds fail loudly and which just drop an arrow.
When the page stops updating, run ai-sdlc check <dir>: that is a schema error, and it prints the file, the path and the message.
Process facts live in the YAML. A fact recorded only in the conversation is not recorded.
Ending
Stop when the team stops, not when the document looks finished. Then:
- Run
ai-sdlc status <dir> and read the counts aloud.
- Resolve every id it lists as referenced but not in a catalog. None may be left when you close.
- Name what is unmapped: processes not covered, activities with no
why:, stages nobody described.
- Give the open slots as a count, and say plainly that the unclaimed activities are not part of it — the team declared those, and nothing is owed on them.
