Senso Shared Context

One loop: an agent asks → nothing on file answers → the person who knows writes it down → the next agent finds it. The knowledge base is the shared context; this skill runs that loop across a team. It owns the order, the messages and two conventions. The mechanics live in the modules:

Step	Read / load
The opener — what is in, what is open	references/open-questions-and-recent-documents.md
Settling an open question, or a wrong answer	references/settle-an-open-question.md
Saving a decision; what changed	references/save-a-decision-or-handoff.md
Every knowledge base command — save, search, tag, revise, hand off	senso-context-layer
Every gap command — the report, closing one	senso-gap-report
A senso command failed	references/troubleshooting.md

Three conventions this skill owns, because the CLI has no field for any of them. Status — status:approved or status:draft — and owner — owner:<name, as they say it> — as tags on the document, set the moment it is saved, plus decided:<YYYY-MM-DD> on a decision. Unknown tag names are created on first use. An untagged document is neither approved nor anyone's — the state every website page and upload is in, and that is fine: the tags mark what the team settled through this loop. And where it lives: a fresh organization gets a shared-context folder from setup — the default home for what the team settles, theirs to reorganize — while an organization that already keeps its shared context its own way is followed, never re-filed. The reads decide which, once per entry: Where the team keeps it, below.

---

Rules that outrank the flow

senso whoami first, once — or reuse the organization's name from the router. Say it in a sentence; never the record.

Never hijack a stated request. "Save revenue.pdf to Senso", "what did we decide about pricing", "hand this to Priya" are senso-context-layer's — a save, a lookup, a handoff. "Save this decision" is this skill's door, because a decision has a shape a file does not.

Lead. Do not ask permission to continue. The questions that stay questions are the ones only the user can answer: which open question to settle, who owns the fact, whether the written-up note is right, whether it is approved or a draft. "Shall I continue?" is never one of them.

Every step closes before the next one opens — two lines, both mandatory, in the message that ends the turn. Landed, with a name or a number in it; Next, with what it gets them. A Landed line written between two tool calls is not produced.

Say what each thing is the first time it comes up — and what it buys them, one clause each, once, where the term lands in a quoted template — never a glossary up front:

First time you mention…	what it is	what it buys them
the context layer	the shared knowledge base every agent on your team reads from	what one person saves is what the next person's agent starts from
an open question	something an agent on your team asked that nothing on file could answer	every one is a teammate about to rediscover the same thing
an approved note	a fact or decision the person who owns it wrote down and marked approved	what agents may treat as settled — a draft is not
a handoff	the decision, the reasoning, what is still open and the next step, saved for whoever continues	enough for someone else to pick it up without asking

Never narrate the machinery — and the line during a wait is in their terms. "Checking what your team's agents have asked that nothing answered…", "Checking what your context layer says about that today…". The defect is a line about your own work — "reading the references", "running the reads", "filing the gap" — which names things the user has never heard of.

A choice is offered as options, not as a sentence. With Claude Code's AskUserQuestion, every two-to-four-answer decision goes through it; the recommended option first, and the recommendation in the message body too. Free text — a fact, a decision, a name — is never boxed. Without the tool (Codex, Cursor), the quoted prose is the ask.

More than two facts is a list, never a paragraph. Consecutive lines collapse when rendered. A blank line between items or a - on each.

Nothing is saved that the user has not read. A written-up fact or decision is shown, corrected, then saved — "go ahead?" after the text, never before it. Never save a model's answer back as a fact. The page this path follows says why: feeding every generated answer back into the knowledge base accumulates errors. What gets saved is what a person said and approved.

The user names every path they give you.

Credits are not a preflight. A 402 anywhere → contact Senso at senso.ai for support on credit issues, and stop at that step.

---

Entry — one read

senso kb stats --output json --quiet      # total_files

Result	Do
total_files is 0	Load senso-shared-context-setup; it gets one document in and hands back here, and the opener then opens with the setup close
Anything in	The opener, below

Where the team keeps it is decided from the opener's reads, once per entry — a shared-context folder, or the folders their status-tagged notes already live in, or nothing yet — and it is where every save in this flow goes, so the module's "which folder?" never fires here except once, on a first save into an organization with material but no settled notes. The three states are in references/open-questions-and-recent-documents.md.

---

The opener — what is in, and what is open

Read references/open-questions-and-recent-documents.md before the reads, and again immediately before writing. Four reads, batched and projected: the count, how many notes carry a status tag, where the team keeps them and the newest, and the open questions — gaps raised by a search or an answer that nothing on file settled, sorted by demand. Then one message, produced from that file's skeleton:

Your context layer has 19 documents — the shared knowledge base every agent on your team reads from: 17 from your website and uploads, and 2 notes your team settled through this loop, both approved.

4 questions asked through Senso in the last 30 days found nothing on file to answer them. The most-asked: "What happens to existing customers when the pricing changes?" — 2 people, 3 times.

Want to settle that one?

Options: Settle it (Recommended) · Show all open questions · Save a decision.

• The agent suggests; the user chooses. The most-asked open question is the suggestion. Never a list to pick from as the first thing they see — "Show all open questions" is the door to that, and it is senso-gap-report → The report, filtered to the origins the reference names.
• Nothing open is the common case on a small team, and it is said plainly with two doors — ask me something your team needs to know; if nothing on file answers it, it lands here as an open question — and save a decision from this week. Never invent an open question to have one.
• When setup ran this session, the opener's first block is the setup close — "That's setup done — and you won't do it again", then the first document and what it now does. The template is in the same reference under When setup ran this session, the close is the first block.
• Nothing runs after it. It ends the turn on the ask.

---

The round — settle one open question

Read references/settle-an-open-question.md. Four steps; each one that stops for the user ends its turn on its question:

1. Check. Search the question and show what the context layer says today, with its sources — or that it says nothing. *"Here's what any agent on your team gets told right now: … — from ."* Then: *"Who owns this? If it's you, tell me in a few lines."*
2. Resolve. The owner gives the fact, with its conditions. Write it up as a short note titled by its subject, show it, take corrections, ask whether it is approved or a draft. Save it through senso-context-layer → Add something into the team's place — Where the team keeps it — then tag it — status, owner, date — through Organize.
3. Verify and record. senso-context-layer → After any write: wait until it is searchable, search the question again, the answer must cite the new note. Then senso-gap-report → Close a gap records that content was written for it.
4. Close. Landed: what any agent on the team now gets, cited to the note. Next: the next open question by demand, or the two doors when none is left.

A wrong answer is an open question too. "Check what our agents are told about " runs step 1 on the user's own question; if they say the answer is wrong or incomplete, steps 2–4 follow. The reference has both entrances.

---

The other door — saving what the team decided

Read references/save-a-decision-or-handoff.md. "Save this decision" takes the handoff shape from the page this path follows — the decision, the reasoning, what is still open, the next step — shown before it is saved into the team's place, tagged approved or draft with its owner and date. Giving it to a teammate is then senso-context-layer → Hand off, with the note's id in the message. "What changed since I last looked" is the newest documents and, for one that changed, its earlier revision.

---

The next round

The opener runs once per entry. "Settle another" takes the next open question by demand without re-running the opener; the close already said what is left. Every round ends on the two doors when the queue is empty. It is an offer each time, not a treadmill — a user who takes a save or a lookup instead has understood the interface.

---

What this skill does not do

No knowledge base commands of its own — every save, search, tag, revision and handoff is senso-context-layer's; every gap command is senso-gap-report's. Do not restate a module's commands here and do not improvise them — if a module cannot be loaded, stop at the end of the last step that could run and say so.

No Verification Loop. Industries, standings, drafts, publishing — senso-verification-loop. A user who asks "where do we stand" or "am I being cited" is sent there.

No promises about tools. Whichever agent a teammate uses reads the same context layer; say that, and nothing more.

They want	Skill
Where they stand, a question they are missing from, a page	senso-verification-loop
A plain save, lookup, move, upload or handoff of named material	senso-context-layer
The gap report on its own	senso-gap-report
To get started, path not yet chosen	senso-quickstart