pipeline-review

---
name: pipeline-review
description: Weekly pipeline review from a CRM export: health score, a dated verdict on every stale deal, a triangulated forecast, three owned actions. Use weekly, or when someone asks what will really close.
---

# Know what will really close this month

One page: a health score, a verdict on every stale deal, a number you can defend, three owned actions.

## Claude does

1. Ask for this week's export, last week's, each rep's commit, the win rate by segment. Map headers to: deal, account, amount, stage, close date, created date, last activity, owner, primary contact; write {FIELD_MAP}.
2. {PROBABILITIES} per stage, from buyer evidence: Commit 85-95%, Best Case 40-60%, Upside 15-30%, Pipeline 5-15%. Commit needs all five: economic buyer met within 14 days, decision criteria in writing, terms agreed, paperwork started, close date inside the period. Fail one and it is Best Case, whatever the rep says.
3. {VERDICTS}: each deal stale (14+ days; 21 above {BIG_DEAL}), stuck 30+ days in stage, past its close date, pushed 3+ times or silent 60+ days gets re-qualify, downgrade or kill, dated. Silent 60+ days is kill without written evidence. Expect 15-30% kills first time.
4. {CLOSE_DATE_RULINGS}: each deal due within 30 days with no prospect-agreed next step, date and owner gets a new close date.
5. {COVERAGE_TARGET}: 1 over the win rate per segment, 3x floor; no win rate, 3x labelled "floor, not calibrated".
6. {MOMENTUM}: one most-pressing deal per rep: what changed, next verifiable step, economic buyer, what blocks signature. Count each rep's deals in the top two stages.
7. {THREE_ACTIONS}: the three that most change the number this week, owner and date each, after reading back last week's.

Claude never guesses a probability, win rate, amount or date the evidence lacks.

## Then Codex does

Codex gets the exports and Claude's decisions and does the arithmetic and writing: days-since counts, severity flags, per-rep roll-up, health score, week-on-week waterfall, three forecast views against rep commit, coverage, the one-page file. Fixed formulas over every row.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Fill before running: {REVIEW_DATE}, {PIPELINE}, {LAST_WEEK}, {REP_COMMIT}, {BIG_DEAL}, {FIELD_MAP}, {PROBABILITIES}, {VERDICTS}, {CLOSE_DATE_RULINGS}, {COVERAGE_TARGET}, {MOMENTUM}, {THREE_ACTIONS}.

## Claude checks

1. Recompute the expected number from the per-deal table; it matches the page to the pound.
2. Every deal over a threshold carries Claude's verdict; none added, missing or doubled.
3. Nothing normal is flagged: a discovery deal closing in 90 days is fine; a verbal deal at 65% is not.
4. Each health dimension matches its flag count; high list capped at 15, medium at 20.
5. The change in the number is split six ways, positive re-rating labelled "optimism".
6. Every action has owner and date; no "could consider" anywhere.

Any failure: name the wrong line, rerun Codex; never patch by hand.

## Rules

- Probability from stage and buyer evidence, never the seller's mood; a verbal yes is 50%, not 90%.
- Under three deals per rep in the top two stages: "not enough at the top"; actions are prospecting, not closing.
- A missing input field is the first finding; over 40% high flags, the number is labelled unreliable.
- Stalled is 14+ days silent; each gets re-qualify, downgrade or kill, never "keep watching".
- A close date within 30 days needs an agreed, dated next step or it moves out, and the page says so. "Continue discussions" is not one.
- Hygiene is prep; the meeting validates data, never enters it. Data, not stories; momentum, not history; one deal per rep.
- Coverage is 1 over the win rate per segment, never flat 3x, never blended.
- "Fix this", never "could maybe consider"; never flag what is normal; never trust one model, field or rep.
- Close dates are expected signature, not hope; stages move backwards when evidence does.
- A repeated action is broken ownership. Audit the data; do not redesign the CRM.
- One page, headings as in CODEX.md, a table for flagged deals, no adjectives.

## If Codex is not installed

Claude does the "Then Codex does" section itself, says so, and still runs the checks.

## Built from

- Anthropic knowledge-work-plugins, pipeline-review: https://github.com/anthropics/knowledge-work-plugins/blob/main/sales/skills/pipeline-review/SKILL.md (23,877 stars, api.github.com, 5 Sep 2026)
- NEON-Rutger/B2B-revops-skills, revops-forecasting: https://github.com/NEON-Rutger/B2B-revops-skills/blob/main/revops-forecasting/SKILL.md (46 stars; sourced benchmarks, v1.6.0)
- elijeangilles/revops-skills, pipeline-hygiene-audit: https://github.com/elijeangilles/revops-skills/blob/main/skills/pipeline-hygiene-audit/SKILL.md (8 stars; 251-deal test dataset)
- gtmagents/gtm-agents, forecast-discipline: https://github.com/gtmagents/gtm-agents/blob/main/plugins/sales-pipeline/skills/forecast-discipline/SKILL.md (397 stars)
- Gong, pipeline reviews for growth: https://www.gong.io/blog/pipeline-review (updated 4 Mar 2026)
- Clari, forecast call guide: https://www.clari.com/blog/the-ultimate-guide-to-your-forecast-call/

# Know what will really close this month - Codex task

You have no memory of the conversation; everything is below. Do not ask questions; report empty inputs in Return and produce what you can.

## You are given

- {REVIEW_DATE}: review date.
- {PIPELINE}: this week's export, one row per deal.
- {LAST_WEEK}: last week's export; may be empty.
- {REP_COMMIT}: each rep's committed number; may be empty.
- {BIG_DEAL}: amount above which stale means 21 days, not 14.
- {FIELD_MAP}: export header to field: deal, account, amount, stage, close_date, created, last_activity, owner, contact, plus next_step, pushes, contacts_count where present.
- {PROBABILITIES}: probability and category (Commit, Best Case, Upside, Pipeline) per stage.
- {VERDICTS}: re-qualify, downgrade or kill per flagged deal, dated, with reason.
- {CLOSE_DATE_RULINGS}: deals whose close date moves, with the new date.
- {COVERAGE_TARGET}: multiple per segment.
- {MOMENTUM}: one deal per rep, four answers.
- {THREE_ACTIONS}: three actions, owner and date each.

## Produce

Apply {FIELD_MAP}; compute per deal days since activity, days in stage, days to close (negative if past). Write `pipeline-review-{REVIEW_DATE}.md`, one page, seven headings in order:

1. **Health score** X/100, four dimensions of 25: stuck 30+ days in stage; stale 14+ days (21 above {BIG_DEAL}); close date past; one contact. Each = 25 x (1 - flagged / total), rounded.
2. **Flagged deals** table: deal, owner, amount, days since activity, days in stage, flag, severity, verdict, date. High: stale on Commit or Best Case or above {BIG_DEAL}; past close date; no next step at Negotiation or later; silent 60+ days. Medium: stuck 30+ days; pushed 3+ times; one contact; discovery closing inside 30 days. Low: export probability 25+ points off {PROBABILITIES}. One row per deal, worst flag first; cap high at 15, medium at 20, state the overflow. Verdict from {VERDICTS}, else NO VERDICT.
3. **Hygiene**: missing close date, amount, next step, contact: count, deals, action. Per rep: high x 10 + medium x 3 + low, highest first.
4. **Close dates** due within 30 days: old date, next step, new date per {CLOSE_DATE_RULINGS}.
5. **The number**, after close-date moves, this period and next: stage-weighted (amount x probability); conservative = Commit x 0.9; expected = conservative + Best Case x 0.5; optimistic = expected + Upside x 0.2; {REP_COMMIT} beside, with percentage gap. Coverage = open pipeline / target, beside {COVERAGE_TARGET}. Change since {LAST_WEEK}: new, won, lost, pushed out, pulled in, re-rated (positive re-rating labelled "optimism"); no {LAST_WEEK}, write "movement not available".
6. **Momentum**: {MOMENTUM} as given; each rep's count in the top two stages.
7. **Three actions**: {THREE_ACTIONS} as given.

Dates DD Mon YYYY. Currency as in the export.

## Rules

- Only {PROBABILITIES}; never a percentage from the export or notes.
- Never change, add or drop a verdict, ruling or action; flag a gap, do not fill it.
- Never invent a missing field; blank the cell and flag it.
- Never flag a deal twice, or a discovery deal closing in 90 days.
- "Fix this", never "could maybe consider". No adjectives.

## Return

Print, in order:

1. Path of the file.
2. Per-deal table: deal, stage, category, amount, probability, weighted amount, period, flags.
3. Health score by dimension, flagged/total.
4. Both periods: stage-weighted, conservative, expected, optimistic, rep commit, coverage; last week's expected; the six-part waterfall.
5. Counts: deals, flags by severity, NO VERDICT rows, close dates moved, over-cap deals.
6. Missing inputs, unmapped headers, unclassifiable rows, one line each.