Got some goodies in here for you today! If you have not already, go over to the Substack and subscribe, we give away the goods every week and today is no different with some sweet skill.md files inspired by and used from the podcast guest.
Lets get into it!
Our guest is Eli Portnoy who is the founder of BackEngine and takes us through his going down in agents and up in effectiveness.
You can go to Youtube, Apple, Spotify as well as a whole other host of locations to hear the podcast or see the video interview.
Every “I have 60 agents running” post on LinkedIn is describing a problem, not a win. Eli Portnoy found that out the expensive way. At one point BackEngine.ai had roughly 200 scheduled AI jobs firing across the company. Nobody was reading any of them. Today they run about 10. Pipeline is up, deals close faster, and every one of the 10 has a human owner and a meeting attached to it.
Eli has founded three AI companies (two acquired, by Telenav and Medallia) and runs 90% of BackEngine’s go-to-market inside Claude. On this week’s episode he shared his screen and walked through the actual setup. Here is what a revenue leader can take from it, and at the bottom there is a skill you can paste into Claude today that rebuilds his highest-value job for your own team.
1) The stack lives inside the chat window now, and the tools did not go away
The first thing on Eli’s screen was the connector list: BackEngine, Fireflies, Gmail, Calendar, Granola, HubSpot, Notion, Slack, Superhuman, Zoom, plus two MCP servers he vibe-coded himself (one just returns website stats). His framing is worth stealing for any exec who asks “are we replacing our tools with AI?” The tools stay. The 15 browser tabs go. HubSpot is still the system of record; he just never opens it, because every read and write happens from inside Claude.
The practical implication for a team: connector coverage is the new adoption metric. A rep with a Claude license and zero connectors is doing email polish. A rep with calendar, CRM, inbox, and call recordings connected can run the jobs in the next section. Audit your team on that axis before you audit them on prompt quality.
2) The scheduled job is the unit of value, not the chat
Most people still use AI in what Eli calls single-player mode. One person, one prompt, one answer nobody else sees. Scheduling flips it to multiplayer: the same output hits the whole team on a cadence, so the meeting starts from shared facts.
The four jobs that survived the cut at BackEngine:
Founder Sales Daily Pulse. Runs at 4pm. Scans all outreach activity against the company knowledge base (Eli calls it “my brain”) and emails what happened, what did not, and what needs a reply.
Follow-up audit. End of day. Reads the calendar, finds every meeting with no follow-up sent, pulls the transcript, drafts the follow-up.
PMF from the last 20 prospect calls. Fires every time 20 new prospect calls land. Reports what is resonating, what is not, which discovery questions are working, and whether a new story he tried moved anything. Lands once or twice a month. Eli: “probably the highest value email I get.”
Health change alert. Any deal or customer that moves a grade (B to A, A to C) triggers an email with the reason, to the whole team.
Notice the pattern. Three of the four are triggered by events in the business (a call count, a health change, an unanswered meeting), not by a clock. The one clock-based job is a summary. If your scheduled jobs are all “every Monday, summarize X,” you are building a newsletter for yourself, not a system.
What makes the PMF job good is the context under it, not the prompt. Eli gave it three things: an expansive prompt that explains what the product does and where it is going, access to all call recordings, and the BackEngine MCP with the sales deck, customer conversations, and prospect history. That is why it can tell the difference between a story he has told 50 times and one he tried yesterday.
3) Your CRM connector is sampling, and it will not tell you
Anyone who plugged Claude or ChatGPT straight into Salesforce and called it done should read this twice. Ask “which of my deals is going sideways?” and the model has no search path for that question. Answering it properly means reading every email and every transcript on every deal, which will not fit in a context window, and answer quality degrades well before you hit the limit. So the model samples.
Eli’s number from watching this across customers: a direct connection to a CRM or call tool pulls roughly 30% of what it needs. The answer reads fine, because 30% of real data still sounds real. The missed deals live in the other 70%.
The fix, whether you buy it or build it, has three parts:
Join the systems (CRM, call recordings, inbox, docs) into one layer instead of connecting them one at a time
Build an index and catalog so a question retrieves exactly the records it needs and nothing else
Permission at the data layer, so a rep’s query on a deal can legally pull the VP’s email thread on that same deal
A quick test for this week: ask your connector “which deals went quiet in the last 14 days,” then hand-check three deals it did not mention. If one of them is quiet, you have your answer on coverage.
4) Two change-management rules that made the AI pay for itself
Eli’s third mistake was treating all of this as a technology problem. His words: “getting real value out of it is not a technology problem, it’s an enablement problem.” Two rules fixed it.
Fewer, owned jobs. Cap the count. Treat each scheduled agent like a hire: it only performs if someone has time to manage it. Ten with attention beats 200 without.
No workflow, no job. Every output must trigger a pre-engineered action: a standing agenda item, a required response, a decision with a date. If an email arrives and nothing is scheduled to happen because of it, kill the job. “Information for the sake of information is not helpful.”
The PMF report is the example. The email is the small part. The team meeting that reads it and changes the pitch is the part that moves revenue. That is also why Eli says the ROI is hard to isolate: the mechanism is a team that spends pipeline reviews on “how do we move this” instead of “what happened,” and catches a product miss in weeks instead of a quarter.
5) A 20-second taxonomy for prompt vs. project vs. skill vs. plugin
Teams burn hours arguing about this. Eli’s version:
Prompt: a one-time task
Project: the same task repeated, with saved context and documents attached
Skill: a how. The method you want applied every time (”analyze it this way, in this order”)
Plugin: packaging. A connector plus a skill plus context, bundled so you can hand it to someone else
Put together a skill for you inspired by all of this and you can grab it at the www.gtmaipodcast.com
The tactical shift
AI maturity is the number of decisions that changed because an agent ran. The agent count tells you nothing about that.
What to do this week:
List every scheduled AI job your team runs. Pause anything without a named owner and a meeting or response tied to it.
Run the 14-day quiet-deal test on your CRM connector and hand-check three deals it skipped.
Install PMF Pulse, fill in the owner and meeting lines, and run it once manually on your last 20 calls before you schedule it. The first run tells you whether your context is good enough to automate.
Eli’s full prompt library (29 sales use cases, each tagged prompt, scheduled, or live artifact) is free at prompts.backengine.org.
Ten agents with owners beat 200 with none.
Top 5 quotes from the episode
“Just because we could meant that we did, and so we had, like, 200 of these firing at any given [time]... and then the second that happens, no one’s reading them.”
“When you connect directly to a CRM or any of these tools, it’s really only grabbing about 30% of what it needs to be grabbing. And so you’re getting answers that look good... it’s just missing a huge amount of actual important information.”
“Getting real value out of it is not a technology problem, it’s an enablement problem.”
“Information for the sake of information is not helpful. It has to be a change that we’re actually implementing.”
“AI is typically pretty single-player mode. It’s not very good at multiplayer mode. What I like about what we’re talking about here is that by scheduling it, it becomes multiplayer mode.”
Skill for you!
---
name: pmf-pulse
description: Reviews your last N prospect calls (default 20) against your company context and reports what is resonating, what is falling flat, which discovery questions work, and what to change in the pitch or product. Run on demand ("run PMF pulse") or as a scheduled job every time 20 new prospect calls land.
---
# PMF Pulse
## What this skill does
Every 20 prospect calls, produce one report that tells a founder or sales leader whether the pitch is landing, where it is breaking, and what to change next week. The output is a decision memo, never a call summary. If a section does not force a decision, cut it.
## Before you run: context you need
Ask for, or locate via connectors, the following. Do not run without at least items 1 and 2.
1. Call transcripts. The last N prospect calls (default N = 20). Pull from Gong, Fireflies, Granola, Zoom, or whatever is connected. Prospect calls only. Exclude customer success, internal, and partner calls.
2. Company context. The current sales deck, a one-paragraph product description, the current ICP definition, and the three to five discovery questions the team is supposed to ask. If a knowledge base is connected (Notion, Google Drive, BackEngine, or similar), read it. If nothing exists, ask the user to paste the product description and current pitch in two paragraphs.
3. The last PMF Pulse report, if one exists. You are measuring change, so you need the baseline.
4. The change log. Anything the team deliberately tried since the last report: a new story, a new discovery question, a new pricing frame, a new demo flow. If untracked, ask: "What did you try differently on calls in this batch?"
## How to analyze
Work call by call first, then across calls. Do not skim a sample. If the transcripts will not fit in one pass, process them in batches of five and keep a running tally; never conclude from fewer than the full set.
For each call, capture:
- Persona and company type (map to the ICP definition, flag if outside it)
- The pain the prospect stated in their own words (quote it)
- Which parts of the pitch got a reaction (a question, a "say more," a specific objection) and which got silence
- Objections raised, with exact phrasing
- Discovery questions the rep asked, and whether each produced a usable answer or a dead end
- Competitor or alternative mentioned (including "we'll build it" and "we do this in spreadsheets")
- Outcome: next step booked, vague follow-up, or no next step
- Anything new the rep tried (compare against the change log)
Then across all calls, look for:
- Pain statements that repeat across three or more calls (your real positioning)
- Pitch segments that consistently produce silence (dead weight)
- Objections that cluster by persona or segment
- Discovery questions with a hit rate above 70 percent and below 30 percent
- ICP drift: share of calls outside the stated ICP, and whether those went better or worse
- Change-log experiments: did the new story, question, or frame move outcomes versus calls that did not use it
## Output format
Keep the whole report under 700 words. Use this exact structure.
PMF Pulse: calls [date range], N = [count]
1. The one-line verdict. Is the pitch landing better, the same, or worse than last report? One sentence with the evidence.
2. What is resonating. Three items max. For each: the theme, how many of N calls it appeared in, one verbatim prospect quote.
3. What is falling flat. Three items max. Same format. Include any pitch segment that produced no reaction in the majority of calls.
4. Discovery question scorecard. Table: question, times asked, times it produced a usable answer, keep / rewrite / kill.
5. Objection clusters. Objection, count, persona or segment, and whether the rep's response worked (next step booked after the objection) or not.
6. ICP check. Share of calls inside the stated ICP. Whether outside-ICP calls performed better or worse. One sentence on whether the ICP definition should move.
7. Experiments. For each change-log item: used in how many calls, outcome versus calls without it, verdict (keep, extend, drop).
8. Three changes for next week. Specific and assignable. "Cut the integrations slide" beats "tighten the deck." Each change names an owner.
9. Confidence and gaps. What you could not determine from the transcripts, and what context would fix it next run.
## Rules
- Quote prospects verbatim. Paraphrase loses the signal.
- Count everything. "Several prospects" is banned. "7 of 20" is required.
- Never invent a metric. If the transcripts do not contain the answer, say so in section 9.
- Prefer the uncomfortable finding. If the founder's favorite story is producing silence, say it in section 3.
- Be concise. This is read on a phone, before the meeting starts.
- No em dashes.
## Owner and workflow (required before scheduling)
Do not schedule this skill until these three lines are filled in. A report with no owner and no meeting is noise.
- Owner: [name]. Reads the report the day it lands and brings it to the meeting below.
- Meeting: [weekly pipeline review / founder sync / Monday stand-up]. Section 8 becomes a standing agenda item.
- Kill criteria: If two consecutive reports produce no changes the team actually ships, pause the job and revisit the inputs.
## Scheduling prompt
Paste into Claude's scheduled tasks. Use a fixed cadence (every two weeks is a reasonable default for a team doing 10 calls a week) or a count-based trigger if your call platform exposes one.
Run the pmf-pulse skill on all prospect calls since [last run date]. Pull transcripts from [connector]. Load company context from [knowledge base or file]. Compare against the previous PMF Pulse report at [location]. Email the report to [owner] and [team list] with the subject "PMF Pulse: [date range], N = [count]". Keep it under 700 words.
## Adapting this skill
- Customer success: swap prospect calls for renewal and QBR calls, replace "pitch" with "value story," replace the ICP check with a churn-risk check.
- Enablement: add a section scoring each rep on discovery question usage, route the report to the enablement lead.
- Marketing: run on inbound discovery calls only, add a section on which campaign or content the prospect mentioned.