Format Sales Call Coaching

Sales Call Coaching

What this skill does

Given one named rep and a time window, this skill reads every real conversation that rep had in the window from the Format workspace, scores all genuine prospecting conversations (no sampling) against a behavioural rubric calibrated for short outbound prospecting calls, and produces a competency assessment with verbatim evidence and exactly one coaching priority.

It is an assessment tool, not a coaching session and not a script writer. The human runs the actual coaching conversation.

Required inputs — stop and ask, never default

1. One specific rep name. This skill assesses exactly one rep per run. If the user doesn't name one ("how are the reps doing?", "score the team"), stop and ask which single rep — never guess, never assess "the team."
2. A time window ("last 14 days", "since May 1"). If missing, ask.

Do not pull any records until both are confirmed. One sizing rule at this step: if the window turns out to hold more than ~15 real conversations, finish the count first and propose splitting the window into two runs ("[Rep] had 31 conversations in those 30 days — want me to run the two halves separately so every call still gets a full read?") rather than degrading depth. Splitting preserves the no-sampling principle; silently skimming doesn't.

Why this rubric is calibrated for short outbound calls

Standard SaaS-discovery rubrics (15–20 discovery questions, talk-ratio targets, MEDDIC multi-threading) assume long video calls with a buying committee. They unfairly fail reps who run short outbound phone calls, where the goal is almost never "close" — it is earn the next step. This rubric keeps the behaviours that matter on a short outbound call and drops the volume metrics that don't.

The skill detects the rep's actual motion before scoring (Step 4) — if the calls turn out to be scheduled discovery/demo video calls rather than outbound dials, the assessment says so up front in a read-first motion note, presents the scores as directional, and adapts the connect-rate framing to the motion that actually exists. Never fabricate a dial metric for a rep who doesn't dial.

The run

Step 1: Confirm inputs

Both mandatory inputs above, confirmed in chat. Then run quietly — the next planned chat output is the assessment.

Step 2: Resolve the rep (1 call)

list_persons({ nameSearch: "<rep name>" }) — a case-insensitive substring match that returns the person's id, email and linked company. emailSearch does the same on the address when you were given that instead. Internal people usually share the company's own email domain, which is the tell when two names collide. A name that matches nobody gets flagged and asked about, not guessed; several matches get disambiguated with the user before any records are read.

(describe_org() is worth one call here too if you want the workspace's coverage span before committing to a window — it tells you the earliest and latest conversation Format holds.)

Step 3: Pull the complete call list

list_records({ personIds: [repId], dateRange: { from, to } }). A bare YYYY-MM-DD bound is accepted and becomes the right day edge (from → start of day, to → end of day UTC); a full ISO datetime works too. Paginate to the full totalCount — that field says how many records matched, not how many came back, and limit tops out at 200 (default 50) with hasMore saying whether to page on. A missing last page is missed conversations, which this skill must never have. Note each record's insightCount and dataSource.

Step 4: Read and classify every conversation — within a context budget

Order reads by insightCount descending, but read every insight-bearing record — no cap. insightCount > 0 reliably marks a real conversation; zero-insight records are usually voicemail/switchboard — count them in the activity table, spot-check borderline short ones.

For each record, get_record({ recordId, includeInsights: true }). The content field is the transcript. Call duration comes from media[].durationSeconds — convert to M:SS for display; it is null where the integration never reported one, so show the duration column blank rather than inventing a number. Don't rely on the context field for duration or direction; its contents vary by source (on Gong records it often carries the call title). participants is not persons: persons is who Format linked to a person record, participants is the raw list the source system reported — its entries vary in shape, and a single record can mix { email, attendance } with { name, spoke }.

Work call-by-call within a context budget. Transcripts are large and a full window can hold dozens. After reading each call: classify it, score it if eligible, write its complete scorecard row and evidence cells immediately — then drop the transcript and carry only the scorecard forward. Never hold all transcripts simultaneously; never let earlier raw transcripts crowd out the later reads. (Where the environment supports delegating reads to sub-tasks, that's a fine way to parallelize — the contract is that every call gets a full read and an evidenced score, whoever reads it.)

Classify on two axes before scoring — a call is scored only if it passes both:

• Reachability: conversation (real exchange — eligible) vs non-conversation (voicemail, switchboard-only, no-answer, sub-~60s with no exchange — counted, not scored; a high non-conversation rate is a list-quality finding, not a rep-skill finding).
• Motion: prospecting (opening/advancing a new opportunity — scored) vs deal management (pricing, procurement, late-stage — flagged, not scored) vs implementation/post-sale vs support (both flagged, not scored). Never score a non-prospecting call against the prospecting rubric — it produces unfair, meaningless numbers. A rep with zero prospecting calls in the window gets exactly that finding: "[Rep] ran no prospecting calls this window (N deal-management / M support) — nothing to score against this rubric; confirm their role."

Separate disqualifications from caves. When a prospecting call ends quickly with no further probing, decide why before scoring:

• ✅ Correct disqualification — the prospect is a genuine, structural non-fit (a setup the product can't serve, a committed equivalent solution, a hard ICP miss) and the rep recognised it and exited cleanly. Good judgement, not failure: label it ✅, mark Objection handling and Next step N/A (excluded from averages, never scored 0), and keep it out of the coaching priority entirely.
• 🔴 Cave — the rep abandoned a viable opportunity. Timing/budget/political stalls ("budget frozen", "call me next quarter", "new leadership settling in") are not disqualifiers — a viable opportunity should be probed and a low-commitment next step attempted. Caving on a stall scores accordingly.

The test: could a competent rep realistically have advanced it? Yes → cave. No → disqualification.

Step 5: Score each prospecting conversation

Score on the seven dimensions using the behavioural anchors in references/rubric.md — read references/scoring-example.md first; it is the gold standard for evidenced scoring.

Evidence cells legible at a glance: plain point → Rep:/Prospect: tagged verbatim quote (gloss only if not self-evident; quotes stay in the call's original language) → for any score below 2/2, Missing: [≤5 words]. Dimension scores always X/2; per-call weighted totals always X.X/2.0. A dimension with genuinely no opportunity (pure scheduling call) is N/A, excluded from that call's total.

Step 6: Roll up — summary first

Aggregate across scored prospecting conversations (disqualifications and N/A dimensions excluded). Per-dimension averages, headline weighted average, and the one coaching priority: the earliest weak rung on the foundational ladder — openings → discovery → value → objections → next step. Don't coach advanced objection reframing on a rep who isn't getting past the gatekeeper (see references/coaching-frames.md).

Output structure

Lead with the headline; tables and short paragraphs throughout; light emoji signposting (📊 🟢 🔴 🎯 ✅ ⚠️).

# 🎯 Sales Call Coaching Assessment — [Rep]
**Window:** [range] · **Org:** [name] · **Rep:** [email]

[Motion note — read first, when the rubric's calibration and the rep's
actual motion differ: what the rep actually runs, what that means for
reading the scores.]

## 📊 1. Activity & headline
| Metric | Count |
[Total records · non-conversations · real conversations · scored ·
✅ correct disqualifications · not-prospecting (by motion) ·
connect rate — labelled for the motion that actually exists.]

**Overall weighted score: X.X/2.0** across [N] scored prospecting
conversations[ · ⚠️ low-confidence — small sample, if N < 8].

[Strengths/weaknesses table: dimension · avg X.X/2.0 · one-line read.]

**🎯 Coaching priority (one thing):** [the earliest weak rung, one short paragraph].

## 🔍 2. Deep dive
[Full per-call scoring sheets for the most instructive calls — strongest,
weakest, clearest pattern examples. Header: company/contact, date, M:SS,
X.X/2.0, source link.]

## 🗂️ 3. Every other conversation (condensed — nothing skipped)
[One row per remaining conversation: contact, date, M:SS, type, score or
N/A, one line, link. Deep-dive set + this table = every conversation.]

## ✅ 4. Disqualifications (for the record, not coached)
[Contact, the verbatim disqualifying line, why it's a genuine non-fit.]

## 🎯 5. Coaching priority, framed SBI
- **Situation:** [call, date, moment] → [link]
- **Behaviour:** "[verbatim quote]" (observable, no adjective)
- **Impact:** [consequence]

## ⚠️ What this misses
[The standing caveat — recorded calls only; no emails, in-person meetings,
or off-call progress; transcription quality varies — a low score on one
call is a prompt to listen, not a verdict; scores reflect call behaviour,
not pipeline outcomes.]

Principles

These are the defaults that make the assessment fair. They're guidance, not law — depart when the situation genuinely calls for it, and say so when you do. Three exceptions stay firm: quotes are never fabricated (and links never invented), the two required inputs are never defaulted, and a correct disqualification is never coached against.

• Every conversation gets read. No cap, no sampling, page to the full count — sampling the "best" calls biases the assessment toward the rep's best work. When the window is too big to read honestly, split the window (proposed at Step 1), don't thin the read.
• Classification decides what's scored. Both axes, before any scoring; the prospecting rubric never touches a deal-management, implementation, or support call.
• Every score is substantiated — evidence cell plus a source link (the record's sourceUrl, or an insight's shareUrl) per call. A score with no evidence and no link doesn't ship; a link that can't be verified gets dropped, not invented. sourceUrl is nullable, so fall back to the insight link when it is absent.
• Quotes are verbatim, in the call's language. Explanation lives outside the quote marks; a bracketed gloss when the call isn't in the user's language.
• One coaching priority, framed SBI — situation, observable behaviour, consequence. No adjectives about the person; overwhelm kills coaching. If the same weakness recurs across reps in separate runs, name it as a team enablement gap rather than repeating it per rep.
• Assess, don't script. No call scripts or role-play dialogue unless asked separately.

How to prompt this skill

Using the Format MCP and the format-sales-call-coaching skill, assess [one rep name] over [time window].

• "How are our reps doing on calls?" → "Which single rep should I assess, and over what window?" — asks for both, never defaults.
• "Score the outbound team since May 1." → "This skill assesses one rep at a time. Who first?" Then the single-rep flow, re-run per rep.