silvercontext · Field notes

Three scenarios.
Real product. Synthetic data.

These are walkthroughs of silvercontext working through a situation. None of the clients or deals are real, and the companies are composites.

Scenario/01 · Territory review · 08:14

Twelve accounts. One hour before standup.

A rep opens the territory view at 8am. Thirty accounts. The scoring diff from the last seven days surfaces three that moved: Apex Logistics (+18), Calloway Partners (-12), Vertex Systems (+22, second time this month).

The rep zooms to Vertex. Call prep pulls two signals: two RevOps hires posted in the last 14 days and a job description that mentions "revenue intelligence tooling." The angle writes itself.

Open to brief took four minutes. No CRM fields got updated and no notes got written; the workspace did the reading.

┌─ northeast · diff 7d ────────────────────────┐
│ Apex Logistics       score:82  +18  hiring  │
│ Calloway Partners    score:34  -12  press   │
│ Vertex Systems       score:91  +22  hiring  │
│ Meridian Health      score:61   -3                │
│ Cedar Freight        score:58   ±0                │
└───────────────────────── 3 moved · 27 stable ─────┘

> focus: Vertex Systems
  last contact    11d ago (email, replied)
  signal a        +2 RevOps hires · 14d
  signal b        "revenue intelligence tooling" · JD · 6d
  angle           expansion motion confirmed, budget cycle Q2
Scenario/02 · Sequence composition · cold start

A territory, a thesis, a sequence.

The rep has a hypothesis: mid-market logistics firms expanding RevOps headcount are underserved by their current tooling. They score the territory by that signal. Fourteen accounts surface.

They pick the top six by score, drop them into sequence composition. The tool drafts three touches per account: first email framed on the specific hiring signal, follow-up with a relevant case angle, LinkedIn note for the principal contact.

Each draft is reviewable before sending. Outcome attribution is built in: when a reply comes, it traces back to which touch, which angle, and which account. It's per-account signal, not an aggregate stat.

thesis: midmarket logistics + RevOps hires 30d
  scored: 14 accounts · top 6 → sequence

Apex Logistics: draft/1
  t1 email    "saw the RevOps hires, what stack are you
               standing up?"                    signal: JD
  t2 email    case: Meridian pattern            angle: QBR
  t3 li       VP Revenue, warm intro path       ·

  on reply → attribute: t?, angle: ?, account: Apex
Scenario/03 · Model Context Protocol · query from chat

Asking the workspace a question.

A sales manager opens their preferred model interface. They connect silvercontext's MCP server.

"Which accounts in the northeast territory have had a signal increase in the last 14 days but haven't had outbound contact in 30 or more days?"

The model calls query_accounts with those parameters. It returns four accounts. The manager asks for call prep on the top one. The model calls generate_call_prep. The output lands in the conversation.

There was no dashboard, no export, and no pivot table. The workspace just answered the question.

> tool_call query_accounts({
  territory: "northeast",
  signal_delta: "+, 14d",
  last_contact: "> 30d"
})
< returns
  Vertex Systems      +22   last: 31d
  Apex Logistics      +18   last: 42d
  Beacon Analytics    +11   last: 36d
  Harbor Freight      +08   last: 47d

> tool_call generate_call_prep({ account: "Vertex Systems" })
< returns
  last contact   31d ago (email, no reply)
  signal         +2 RevOps hires, JD mentions tooling
  angle          budget cycle Q2, expansion pattern

Archived notes

Note/04 · 2026.03 Scoring diffs are more useful than snapshots. The change is the signal, not the state.
Note/03 · 2026.02 Call prep quality is bounded by public signal density. For known accounts it's high. For cold ones it surfaces what's there and stops, rather than inventing the rest.
Note/02 · 2026.01 ASCII as interface was a constraint that became a feature. When I needed the model to read the workspace, plain text was the only format that worked. I stopped apologising for it when the model started answering questions I could not.
Note/01 · 2025.11 The first version had a normal UI. I rebuilt it in ASCII because the AI kept getting the spatial relationships wrong when reading a DOM. Moving to a single plain-text format fixed it.