Pre-call / pre-meeting briefing from your vault

Before a meeting with a prospect, customer, or candidate, the user asks the agent one question — "We're meeting with [Account]. Pull everything we have on them and prep me." The agent reads everything in the shared workspace tagged to that account: every email thread, call transcript, task-tracker comment, Slack/Telegram mention, prior brief — and assembles a one-page brief in 30 seconds. People on both sides, prior agreements, open commitments, blockers, last-touch date, what was promised and whether it was delivered, three things to push, three things to avoid, a suggested opener. Folded in from two public sales-skill packs: the brief now also carries a 5-category discovery question set, a pain → cost → your-angle research schema, an objection prep table, a meeting-type × must-leave-with matrix, and an attendee influence map, so the brief tells the rep not just what happened but what to ask and who to win.

Three compounding outcomes. (1) The right person comes prepared. The canonical use case: "Have any of us ever talked to anyone at a large prospect? Pull every comm we've ever had." The agent answered in seconds and surfaced a two-year-old thread with an engineer at the prospect that ended on "let's revisit" — that became the meeting opener. (2) Account-team handoffs stop costing 30 minutes of "fill me in". Anyone covering a teammate walks in cold-prepared; prep time on a typical account drops from 30–60 minutes to ~2 minutes. (3) The discovery framework forces relevance. A public CEO prospect-research skill enforces one rule the vault-only version lacked: every insight must connect to what the user sells, not generic research. A brief full of true-but-irrelevant facts wastes the meeting; the schema forces each item to a pain, a cost, and an angle the seller can act on.

1. Foundation: shared workspace as second brain. Requires the foundation case — a single Git-backed folder tree with all company knowledge ingested, per-department repos under the umbrella so access can be restricted.
2. Ingest transcript pipeline. Cron pulls new meetings from the transcription source into Projects/Transcripts/, named YYYY-MM-DD_[client]_[meeting-type].txt, with an index file for fast lookup.
3. Account-tagging convention. Every transcript, Slack dump, task comment carries an [ACCOUNT:Name] tag. Load-bearing — without it the skill name-matches and conflates "John at one account" with "John at another".
4. The brief skill. Given an account, the agent greps the workspace, reads the latest 3 transcripts + last 6 months of email + open task items + the strategy/competitor file, and produces the structured brief (people, history, last touch, commitments, blockers, three-to-push, three-to-avoid, opener).
5. Add the discovery framework subsection. The brief skill now also emits a discovery block built from the 5-category question set and the pain/cost/angle schema (see prompts). This is the half-merge from the two sales packs: the agent doesn't just summarize, it proposes the questions to ask and the angle to land them.
6. Run it. From the workspace root: "Use sales context. Brief me for the [Account] meeting at [time]. Pull last 3 calls, all open task-tracker items, any thread from the last 6 months, and generate the discovery block."
7. Iterate the brief structure. First brief is rough; tighten the skill file (not the prompt-of-the-day) over ~3 iterations. When the brief is wrong, fix the skill or the tagging convention.
8. Tie it to the calendar. Optional cron agent reads tomorrow's calendar, runs the brief skill for every external meeting, posts a single morning Telegram digest with one collapsed brief per meeting.

Brief skill structure (~80 lines in Skills/sales-brief/SKILL.md):

1. Read sales context:./Docs/SalesAndMarketing/AGENTS.md
2. Resolve account → grep workspace for "[ACCOUNT:<name>]" and "<name>" in transcripts, emails, tasks
3. Read at most: 3 most-recent call transcripts, last 6 months of email
 threads, all open task-tracker items, the competitor file if present.
4. Output exactly this shape:

 # [Account] — brief for <time>
 ## People (us / them)
 ## History (one paragraph, dated)
 ## Last touch (date + what happened)
 ## Open commitments (ours / theirs, with owner and due-by)
 ## Blockers (theirs that affect the deal)
 ## Three things to push
 ## Three things to avoid
 ## Suggested opener (one sentence)

5. Cite every claim with [REF: <path>] so the user can verify.

5-category discovery question set (verbatim shape from a public sales-call-prep skill, repo labs21-dev/agents-stack):

For this meeting, generate 3-5 questions in EACH category:
 1. CURRENT STATE — how do they handle <problem> today?
 2. PAIN & PRIORITY — what's the cost of the status quo? is it a top-3 priority?
 3. DECISION PROCESS — who signs, what's the eval path, what's the timeline?
 4. SUCCESS CRITERIA — what does "this worked" look like in 90 days?
 5. COMMERCIAL CONTEXT — budget cycle, incumbent, switching cost.

Pain / cost / angle research schema (verbatim shape from a public CEO prospect-research skill, repo TheCraigHewitt/skills):

For every insight you surface, output exactly three lines:
 PAIN: <the specific problem they have, in their words>
 COST: <what that problem is costing them — quantify if the source allows>
 YOUR ANGLE: <how what WE sell connects to this pain — not generic research>

Rule: if an insight cannot be tied to what we sell, drop it.
Rule: be honest about what you don't know. Write "[unknown — needs ask]"
 rather than inferring. No invented quotes, no inferred attendees.

Conversation-starter test (stalker vs peer):
 BAD (stalker): "I saw you spoke at <event> and went to <school>..."
 GOOD (peer): "Most <role>s I talk to are fighting <pain> right now —
 is that on your plate or already solved?"

Meeting-type × must-leave-with matrix + attendee influence map:

Tag the meeting type, then state the ONE thing to leave with:
 Discovery → a quantified pain + the decision process
 Demo → a yes/no on whether THIS solves THEIR named pain
 Negotiation → a single open commercial term + a close date
 QBR → the next expansion or renewal trigger

Attendee influence map (label every attendee):
 decision-maker | champion | evaluator | blocker | end-user
For each: what they want, what they fear, how to address it.

Anti-hallucination rule (loaded from the root AGENTS.md / CLAUDE.md):

Every statement in the brief is either [CANONICAL] (first declaration,
sourced) or [REF: folder/file.md#section]. No invented quotes. No inferred
attendees. If unknown, say "[unknown — needs ask]".

• Brief without citations is a fabrication generator. Without [REF:...] per claim, the agent confidently invents agreements that never happened. The anti-hallucination rule is the floor.
• Tagging convention is the foundation. Without consistent [ACCOUNT:X] tags, the agent conflates same-name people across accounts. Fix tags first.
• Generic research is worse than no research. The pain/cost/angle rule exists because the most common failure of an AI brief is a wall of true-but-irrelevant facts. If an insight can't tie to what you sell, drop it.
• "Be honest about what you don't know." The prospect-research merge adds an explicit anti-hallucination clause: the agent must write [unknown — needs ask] rather than infer a decision-maker or a budget. Keep this verbatim; it's the same discipline as the [CANONICAL]/[REF] rule.
• Don't ingest into Notion; ingest into the filesystem. Notion/Confluence search is worse than grep, agents can't write back cleanly, and you lose two-layer navigation.
• Don't try to brief without the foundation case. This is a feature of the shared workspace, not a standalone install. Empty vault → invented brief.

• Calendar-driven morning digest. Cron agent runs the brief skill for every external meeting on tomorrow's calendar, posts one Telegram digest.
• Vertical-specific briefs. CS version pulls integration status + open tickets + last commercial-terms doc; recruiter version pulls candidate file + interview transcripts + role-competency match. Same skill, different Sections block.
• Cross-team intelligence query. Ad-hoc, not per-meeting: "Show me every conversation about pricing with prospects above $100K ARR in the last 90 days." Same vault, different prompt.