discovery-questions

---
name: discovery-questions
description: Builds the five questions that decide whether a deal is real, so you stop losing weeks to opportunities that were never going to close. Use before any discovery call.
---

# Ask the five questions that decide whether this deal is real

Five questions in the buyer's language, in order, plus the answers that mean walk away.

## Claude does

1. Asks for four things and guesses none: the company and what it does; what you sell and the deal size; what has been said; who you have met, by job title.
2. Drafts one question per MEDDICC decision: the problem and its cost today (Metrics); who signs (Economic Buyer); what they judge on (Decision Criteria); how they buy, legal and procurement included (Decision Process); when the money must be spent (Compelling Event).
3. Writes each in the buyer's words, opening the two hardest with "Can you help me understand..." or "Can you walk me through..." Gong found these pull long answers, not a yes or no.
4. Plans the call on Gong's 519,000-call study: 11 to 14 questions, spread evenly not frontloaded, uncovering three or four problems. The five are the spine, rest follow-ups.
5. Cuts the count for C-suite meetings, where Gong found heavy questioning harms the call.
6. Puts budget and price on the first call. Gong, 11,331 opportunities: price raised on call one wins 42%, call three 15%, never 5%. Budget on call one, 49% versus 7%.
7. Writes, per question, the answer meaning the deal is dead, and the follow-up testing it once.
8. Ends with the next-step ask. No next steps on a first call correlates with close rates falling 71%.

## Then Codex does

Codex mass-produces the layer around the finished five: a call sheet, a scoring grid, a CRM-paste block, and the five rewritten for every other title you sell. Formatting and repetition, not judgement.

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

Claude fills first: {COMPANY}, {WHAT_YOU_SELL}, {DEAL_SIZE}, {STAKEHOLDER_TITLES}, {FIVE_QUESTIONS}, {DISQUALIFIERS}.

## Claude checks

1. Every question is open. Any answerable yes or no gets rewritten.
2. No question carries a fact Codex was not given. Invented headcount, budget or tech stack means rerun.
3. The five map one-to-one onto problem, metric, buyer, process and timing.
4. Budget or price sits in the first-call set.
5. The sheet totals 11 to 14 questions, spread out, not stacked at the start.
6. Each disqualifier is an answer you could hear, not a label like "no budget".

Fail any check, rerun Codex naming the failure. Never hand-patch its output.

## Rules

- Ask, then stop talking. Discovery is a tennis match, not a football game.
- Three or four problems, not one. One is a feature request, four a business case.
- Never ask what public record already answers.
- Fewer questions for C-suite meetings.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude produces the sheet, grid and variants itself, and says so.

## Built from

- Gong Labs, 519,000 discovery calls: https://www.gong.io/resources/labs/deal-closing-discovery-call/
- Gong Labs, 11,331 opportunities: https://www.gong.io/blog/data-reveals-the-best-time-to-talk-price-and-budget
- MEDDICC, Andy Whyte: https://meddicc.com/resources/5-signs-you-dont-have-a-champion
- louisblythe/Sales-Skills, 143 stars (api.github.com): https://github.com/louisblythe/Sales-Skills

# Ask the five questions that decide whether this deal is real - Codex task

Format a sales discovery call sheet. Work only from what is below.

## You are given

- {COMPANY} - buyer's company name and one line on what it does.
- {WHAT_YOU_SELL} - the seller's product, one line.
- {DEAL_SIZE} - typical contract value.
- {STAKEHOLDER_TITLES} - the job titles this seller normally meets.
- {FIVE_QUESTIONS} - five finished questions, each tagged with one of: problem, metric, economic buyer, decision process, timing.
- {DISQUALIFIERS} - per question, the answer meaning the deal is not real.

## Produce

Four files in the working folder.

1. `call-sheet.md` - one page to hold during the call. The five questions in the given order, each with three follow-up prompts underneath, spread as five blocks across the call rather than grouped at the front. Total question count including follow-ups must land between 11 and 14. State the total at the top. End with one next-step question asking for a specific date and named attendees.
2. `scoring-grid.md` - a table, one row per question, columns: what was asked, what they said, real or not real, evidence quoted from the buyer. Leave the answer columns blank. Footer line: three or four distinct business problems named means proceed; one or none means requalify.
3. `crm-notes.txt` - plain text, no markdown, under 900 characters, labelled fields Problem, Metric, Economic buyer, Decision process, Timing, Next step. Blank values, ready to paste.
4. `by-role.md` - the same five rewritten once per title in {STAKEHOLDER_TITLES}, keeping the same five decisions. For any title containing CEO, CFO, COO, CTO, Chief or Founder, cut that block to three questions and mark it "C-suite: fewer questions".

## Rules

- Every question must be open. If it can be answered yes or no, rewrite it.
- Use only the facts in the placeholders. Do not add headcount, budget, tools, competitors, locations or dates that were not given.
- Do not invent statistics, customer names or quotes.
- Budget or price must appear in the first-call sheet, not deferred.
- British English. No em dash characters anywhere.
- Buyer's plain language, not sales jargon.
- If a placeholder is empty, write `MISSING: <name>` in the file and continue. Do not fill the gap yourself.

## Return

Print to stdout, in order:

1. The four filenames written, with byte sizes.
2. The total question count on `call-sheet.md`.
3. The five questions as they appear in the sheet, numbered, each with its tag.
4. `OPEN-CHECK: pass` only if no question in any file can be answered yes or no; otherwise `OPEN-CHECK: fail` and list the offending lines.
5. Any `MISSING:` markers you wrote.