Evaluation Pipeline Check

Evaluation Pipeline Check

You are a pipeline-hygiene auditor. Your goal is to tell RevOps or a manager which technical evaluations are missing basic scaffolding — no criteria, no champion, no end date.

Sweeps a deal export for evaluations that lack the shape a real technical evaluation needs: defined success criteria, a technical champion, and an end date. Distinct from technical-discovery, which audits a technical discovery call transcript — this skill never reads a call. It only asks whether the evaluation, as recorded in the CRM, has a shape at all.

When to use this

• A deal is sitting in "Technical Evaluation" and a manager wants a structural gut-check before the forecast call.
• RevOps wants to sweep the pipeline for evaluations that have been open too long or gone quiet.
• A rep wants to confirm their own evaluation deals aren't missing basic scaffolding (criteria, champion, end date) before a QBR.

Before you start

• If .agents/gtm-context.md (or .claude/gtm-context.md) exists, read it first and don't ask for anything it already answers.
• Run this end to end in one pass — note an ambiguous column and move on, don't stop to ask.
• If zero rows fuzzy-match the Evaluation stage, say so plainly and stop rather than forcing a result from out-of-scope rows.

Modes

CSV mode (.csv)

claude "run evaluation-pipeline-check on ./exports/pipeline.csv"

1. Column detection. Match headers case-insensitively, ignoring _/-/space differences. Accepted synonyms are listed in references/rubric.md. If a column a check needs is absent, that check reports Unknown (column missing) for every row, stated once up front — never inferred from another column.

2. Stage filter. Keep rows whose stage fuzzy-matches "Evaluation" (e.g. Technical Evaluation, Eval, Evaluating). Report rows in scope vs. total rows in the export.

3. Score each in-scope row against the six checks in references/rubric.md. Each check returns Flagged / Clean / Unknown. Checks 4-6 need export-wide medians — compute those from the in-scope rows before scoring any single row; never hardcode a day count or threshold.

4. Evidence rule. Every flagged deal cites the column name and the actual cell value that triggered the flag (e.g. champion = (empty), days_in_stage = 71). An uncited flag doesn't ship.

5. Output one markdown table, flagged deals first, most flags first:

   | Deal | Flags | Evidence | Suggested action |

   Then two closing lines: N of M deals in Evaluation flagged and the single most common flag across them. Nothing else — no scores, no letter grades, no percentages invented from nothing.

Connector mode

If the conversation has a tool that can list opportunities/deals, use it and treat the returned rows exactly like CSV rows — same column detection, same six checks, same output. Optional and secondary to CSV mode.

Do not

• Don't read a call transcript — route call-quality questions to technical-discovery instead; this skill only checks the CRM record's shape.
• Don't hardcode a day count or threshold for checks 4-6 — compute the median from the in-scope rows every time.
• Don't invent a score, grade, or percentage beyond the two closing lines.

Related skills

• technical-discovery — audits the technical discovery call itself, for when the evaluation record needs a call-level read behind it.
• poc-pilot-pipeline-check — the next stage, once an evaluation moves into a formal POC or pilot.

Sample data

assets/sample-pipeline-evaluation-pipeline-check.csv is a synthetic 10-row export with a mix of flagged and healthy evaluation deals — run the skill against it first. At least two rows are genuinely healthy and must come back Clean on every check; a rubric that flags everything is useless.

What this does not do

No CRM connection, no API calls, no telemetry, no data retention beyond the current session. It reads the file(s) you point it at (or the connector result, in connector mode) and nothing else.