robomotionio Pipeline Review
Two deterministic scripts pull and crunch the pipeline; you, the agent, write the exec summary, the diagnostic, and the recommendations. Numbers stay computed (trustworthy); the model only narrates and selects recommendations.
v2.0.0 — depth upgrade.
analyze_pipeline.pynow computes per-stage age/velocity, stage-to-stage conversion, weighted pipeline (amount × stage-probability), slippage rate, per-owner win rate + avg deal size, and tiered forecast coverage. This SKILL.md carries the full metric taxonomy, recommendation-logic tables, and a complete output skeleton.
When to use
- "Review my pipeline", "deal review", "how's our pipeline looking".
- "1:1 prep", "board meeting prep", weekly/monthly/quarterly sales review.
How to run
Step 1 — pull deals (deterministic)
python3 ${SKILL_DIR}/scripts/fetch_deals.py --crm hubspot --output ${WORKSPACE}/deals.json
CRMs: hubspot, pipedrive, close, salesforce, or keyless csv:
python3 ${SKILL_DIR}/scripts/fetch_deals.py --crm csv --csv ${WORKSPACE}/export.csv \
--map "name=Deal Name,stage=Stage,amount=Amount,source=Source,created_at=Create Date,closed_at=Close Date" \
--output ${WORKSPACE}/deals.json
Emits the standard record {id, name, stage, amount, source, created_at, closed_at, owner}. For a comparison trend, pull the prior same-length window into a second file.
Step 2 — compute metrics (deterministic)
python3 ${SKILL_DIR}/scripts/analyze_pipeline.py --deals ${WORKSPACE}/deals.json \
--stage-order "Lead,Qualified,Demo,Proposal,Negotiation,Closed Won,Closed Lost" \
--qualified-stages "Qualified,Demo,Proposal,Negotiation" \
--won-stages "Closed Won" --lost-stages "Closed Lost" \
--period-start 2025-01-01 --period-end 2025-03-31 \
--expected-cycle-days 45 --target-pipeline 500000 --quota 400000 \
--stage-probabilities "Lead=0.05,Qualified=0.2,Demo=0.4,Proposal=0.6,Negotiation=0.8" \
--output ${WORKSPACE}/metrics.json
--stage-order (ordered funnel) unlocks stage-to-stage conversion; --stage-probabilities
sets weighted-pipeline factors (a sensible built-in ramp is used for any stage you omit);
--quota drives the weighted-coverage tier. Run again on the comparison-period deals for
trends.
Metric taxonomy (what every block means)
| Block | Field | Definition / how to read it |
|---|---|---|
| volume | total/open/won/lost, open_pipeline_value |
Raw funnel size. Open count without open value = amount hygiene gap. |
| qualification | qualification_rate |
Qualified+won ÷ total. <40% on inbound-heavy = leaky top; >80% = stages too loose. |
| source_attribution | per-source deals/won/won_value/win_rate |
Where revenue truly comes from. Compare win_rate, not volume — a low-volume/high-win source is the scale target. |
| per_stage | open_deals/open_value/weighted_value/avg_age_days/median_age_days |
Per-stage concentration + how long deals sit. avg_age ≫ median = a few zombies skewing it. |
| stage_conversion | per-step conversion_rate |
% that advance from stage→next. The lowest step is your bottleneck; size the fix by the value pooled there. |
| velocity | avg/median_cycle_days_won |
Won-deal cycle time. Median is the honest number; avg − median gap = long-tail deals. |
| stuck_deals | list sorted by age_days |
Open > expected cycle. These are the 1:1 action list. |
| slippage | slippage_rate |
Open deals past their expected close ÷ open deals with a close date. >30% = forecast you can't trust / sandbagged dates. |
| win_loss | win_rate, avg_won_amount |
Closed-won ÷ closed. Segment by owner/source before drawing conclusions. |
| by_owner | per-owner win_rate/avg_won_amount/open_pipeline_value |
Coaching + capacity signal. Low win_rate + high open value = at-risk number. |
| forecast | weighted_pipeline_value, coverage_ratio/tier, weighted_coverage_ratio/tier |
Σ(amount×stage-prob) is the realistic call. Coverage tiers below. |
Forecast coverage tiers (open or weighted ÷ target/quota): healthy ≥4.0 · adequate ≥3.0 · thin ≥2.0 · at-risk <2.0. Raw coverage uses open pipeline ÷ target; weighted
coverage uses weighted pipeline ÷ quota and is the one to trust for a forecast call.
Recommendation logic (metric → diagnosis → action)
| If the metric says… | Likely diagnosis | Recommended action |
|---|---|---|
One stage_conversion step ≪ the others |
Bottleneck stage (e.g. Demo→Proposal) | Inspect those deals; fix the stage exit criteria / enablement for that motion. |
avg_age_days high in a mid-stage |
Deals stalling, not dying | Force a next-step or disqualify; add a stage-age SLA. |
slippage_rate > 0.3 |
Dates are fiction / deals slipping | Re-date every slipped deal with a real reason; tighten close-date discipline. |
weighted_coverage_tier = thin/at-risk |
Not enough realistic pipeline to hit quota | Trigger pipeline-gen now; don't wait for stage math to fix itself. |
A by_owner rep: low win_rate + high open value |
Rep needs deal help or has junk pipeline | Deal coaching + scrub their open deals for stage accuracy. |
qualification_rate very high but win_rate low |
Stages entered too early (sandbagging the top) | Tighten qualification gate; recount what "Qualified" requires. |
A source with high win_rate, low volume |
Under-invested winning channel | Shift spend/effort toward it; it's the scale lever. |
median_cycle_days_won ≫ expected cycle |
Slow motion / multi-threading gap | Map the slow step; add mutual action plans. |
Step 3 — write the report (you, the agent)
Read metrics.json and write an executive summary + a detailed diagnostic, every
claim citing the metric that drove it. Use the data_quality block to caveat any section
its gaps undermine (mostly-blank source → caveat attribution; stages_without_probability
→ note the weighted figure under-counts those stages; high open_deals_missing_expected_close
→ slippage is partial).
Output skeleton (fill, don't invent numbers):
# Pipeline Review — [Period]
## Executive Summary
| Metric | Value | vs prior / benchmark |
|---|---|---|
| Open pipeline | $X (N deals) | … |
| Weighted pipeline | $X | … |
| Coverage (weighted ÷ quota) | X.X× — [tier] | … |
| Win rate | X% | … |
| Avg won deal | $X | … |
| Median cycle (won) | N days | … |
| Slippage rate | X% | … |
**🔴 Red flags:** [worst conversion step · at-risk coverage · slippage · stuck value]
**🟢 Green lights:** [healthy stages · winning sources · strong owners]
**Top 3 actions:** 1) … 2) … 3) … (each tied to a metric above)
## Diagnostic
### Volume & Qualification — [commentary + rec]
### Source Attribution — [win_rate by source, where to invest + rec]
### Stage Conversion & Velocity — [the bottleneck step + the slow stage + rec]
### Stuck Deals — table (name · stage · amount · age · owner) + per-deal next step
### Slippage & Forecast Integrity — [slippage_rate read + date-discipline rec]
### Win/Loss & Owner Performance — [by_owner table + coaching/capacity rec]
### Forecast & Coverage — [weighted call + coverage tier + pipeline-gen rec]
## Data-Quality Caveats
[which sections are weakened by which blank fields]
Outputs
pipeline-review-[YYYY-MM-DD].md— exec summary + diagnostic + stuck-deal action list + structured recommendations. Workspace file + Agent Teams channel attachment; optional export to Sheets/Notion/Slack/email.
Credentials / env
- Required: none. The default
--crm csvpath is keyless (CSV/paste export of deals). No LLM key — the report prose is your job as the agent. - Optional: if a CRM key is set (
HUBSPOT_API_KEY/PIPEDRIVE_API_TOKEN/CLOSE_API_KEY/SALESFORCE_ACCESS_TOKEN+SALESFORCE_INSTANCE_URL) →fetch_deals.pypulls deals live from that CRM; if not → the keyless--crm csvpath (default). Export-target creds (Google Sheets, Notion MCP,SLACK_*, an email API) used only if you export the finished report.
Notes & edge cases
- Degrades gracefully: minimum viable = deal name + stage + created date. Missing
amountdrops revenue/forecast; missingsourcedrops attribution — the script still emits every other block and counts the gaps indata_quality. - Stage roles are passed as flags so the skill is CRM-agnostic; map your funnel's exact
stage names into
--qualified/won/lost-stages. - Report depth auto-scales with period length (weekly → volume/stuck; quarterly → trends/forecast). Skip trends if no comparison window was pulled.