find-customer-quotes

作者: openai

從逐字稿、通話記錄或匯出的錄音中,擷取特定主題的客戶或潛在客戶引述,並以逐字稿優先的證據與明確…

npx skills add https://github.com/openai/role-specific-plugins --skill find-customer-quotes

Find Customer Quotes

Context-Gathering Intake

Whenever this skill asks for context, strongly prefer using the answers-ask-user-input skill and the ask_user_input tool over other tools such as request_user_input; otherwise ask directly in the conversation.

Extract high-confidence customer or prospect language from transcript-like evidence. This skill owns quote discovery, verification, selection, and readable provenance; it is not call analytics, paraphrase generation, legal review, or a posting workflow.

Common Skill Instructions

MANDATORY: If not already in context, read and adhere closely to plugins/sales/skills/index/SKILL.md## Cross-Skill Best Practices.

Key Dependency Categories

Use the narrowest evidence lane that can support the requested quote set.

  • [Blocking] ~~Meeting Transcripts for transcript search, fetched transcript text, speaker labels, participants, dates, companies, and source links. It blocks the default live-source quote-extraction path; explicit transcript-like material already in context satisfies the need.
  • ~~CRM only for account identity, customer/prospect status, and company or segment filters that narrow transcript search
  • ~~Knowledge & Files for uploaded or exported transcripts and grounded call notes that preserve direct language and speaker context

Transcript-like evidence is required. ~~CRM, account summaries, public research, internal notes, and memory can narrow scope but are never quote evidence. If no live or user-provided transcript-like material exists, ask for a transcript export, pasted transcript text, recording export with text, or grounded call notes.

Reference Loading

SKILL.md owns the normal transcript-first path, thresholds, and readable output. Use references/extraction-and-output.md when transcript parsing is ambiguous, speaker-role confidence is hard to judge, quote ranking or deduplication needs the full rules, JSON is requested, or failure handling needs the exact response shape.

Terms

  • Quote candidate: a verbatim snippet that may be useful, but still needs speaker, context, and usage checks before being treated as customer evidence.
  • Customer confidence: confidence that the speaker is actually a customer or prospect rather than an internal teammate, partner, or unknown speaker.
  • Fallback evidence: relevant non-customer or lower-confidence material that may explain a gap but should not be presented as a customer quote.
  • Safe usage notes: guidance on attribution, sensitivity, confidence, and where the quote can responsibly be reused.

Workflow Guidance

1. Resolve theme and evidence scope

  • If the theme or transcript-like evidence scope isn't clear, ask the user via ask_user_input(). For a bare invocation such as “Find customer quotes for me,” do not assume a theme or transcript set; offer:
    • Use the most common blocker from recent Meeting Transcript evidence
    • Use a theme, objection, product area, or narrative I specify
    • Use a specific transcript or call set I provide
  • For a broad request such as “Find customer quotes about the most common blocker in my recent customer calls,” treat both the recommended theme and broad recent-call scope as a clarification case; offer:
    • Use the blocker theme you recommend from recent Meeting Transcript evidence
    • Use a blocker theme I specify
    • Use a specific transcript or call set I provide Do not search transcript evidence or draft a quote set until the user selects one.
  • Require a feedback theme, objection, product area, use case, narrative, or a specific transcript set after that clarification. Preserve the user’s wording.
  • Default to five quotes per theme. Process many themes in small batches.
  • Normalize time windows: use explicit dates when supplied; treat “last month” as trailing 30 days unless the user says “last calendar month”; use calendar meaning for “last quarter” and trailing meaning for “last N days/weeks.”
  • If a date phrase is too ambiguous to convert responsibly, ask one brief clarification unless the user clearly signals flexibility.
  • For an ambiguous account, segment, or transcript set, make a bounded pass and offer up to five concrete candidates via ask_user_input(). Do not search transcript evidence or draft a quote set until the user selects one. Use ~~CRM only to resolve filters; do not gather CRM content as quote evidence.
  • If a recent call-follow-up, pasted transcript, upload, or linked call clearly supplies the theme and evidence source, proceed without asking.

2. Search transcript evidence first

  • Start with ~~Meeting Transcripts when available, using the theme plus one or two intent terms such as pain point, blocker, request, limitation, or need.
  • Apply company, segment, and date filters only when the user supplied or confirmed them. Start with 10-15 results and a relevance threshold around 0.7.
  • If the selected ~~Meeting Transcripts source exposes an account-segment filter with a known enum, use the connector-visible value that matches the user's wording.
  • Do not invent additional user intent. If the user asks only for a theme, keep the query theme-centric.
  • If sparse, simplify the query, lower relevance modestly to about 0.6, then expand toward 20-40 results only when coverage still needs it.
  • Fetch the top-ranked transcripts first and expand only until enough high-confidence, diverse evidence exists or the first pass is clearly thin.
  • For pasted, uploaded, or exported material, parse only transcript-like text that preserves enough speaker/context evidence to classify customer or prospect likelihood.
  • Keep a mapping of theme, search-result metadata, fetched transcript content, connector-specific refetch handle, and call or transcript URL when available. Do not assume a fixed transcript schema; inspect the fetched shape and adapt.

3. Extract verbatim candidates

  • Keep only direct, substantial language that expresses a pain point, blocker, constraint, concern, unmet need, request, purchase condition, or other theme-relevant signal.
  • Keep wording verbatim. Do not clean up grammar, join fragments, or convert paraphrases into quotes.
  • Exclude paraphrases that are not direct quotes unless the user explicitly asks for paraphrases. If requested, label them separately and never present them as verbatim quotes.
  • Exclude obvious internal seller statements, generic praise, garbled fragments, and tiny snippets without usable meaning.
  • Preserve transcript, call, date, company, speaker label, participant context, and source link when available.

4. Verify speaker and rank precisely

  • For each candidate, inspect nearby transcript context and assign speaker_role_guess as customer, prospect, internal_seller, or unknown; customer_confidence from 0.0-1.0; theme_relevance from 0.0-1.0; and a concise evidence note.
  • Keep only candidates with customer_confidence >= 0.8 and theme_relevance >= 0.75 by default.
  • If coverage is thin, prefer fewer quotes over lowering customer confidence. You may lower theme relevance slightly, to about 0.65, only when speaker evidence remains strong.
  • Rank by customer confidence, theme relevance, diversity across calls/customers, specificity/actionability, then readability while remaining verbatim.
  • If no candidate passes the customer threshold, report 0/[target] customer/prospect quotes. Do not backfill the main quote set with unknown or internal speakers.

5. Deduplicate and select exemplars

  • Normalize whitespace and punctuation for comparison only; never alter displayed quote text.
  • Drop exact and near-duplicate quotes.
  • Default to one quote per call. Use a second quote from the same call only when it adds materially different evidence or is needed to reach the target with strong quotes.
  • Prefer breadth across customers when identity is available.
  • Keep relevant non-customer transcript snippets only in a separate fallback section, labeled by role such as internal evidence, vendor-eval evidence, partner-readiness evidence, or unknown non-customer evidence.

Next Step Options

After the first output, offer the most relevant follow-up from the options below. Offer one clear transition, not a menu. Suggest ONLY these unless you are very confident another option is more useful:

  • Search an adjacent or narrower theme to improve evidence coverage.
  • Expand or tighten the transcript, account, segment, or time-window scope.
  • Package the selected quotes into a business case, competitive brief, meeting prep, or evidence appendix.
  • Draft a source-safe internal summary of what the quotes support and what they do not support.
  • Close the smallest speaker-identity or transcript-quality gap.

Next steps to avoid:

  • Paraphrasing quotes, presenting ambiguous speakers as customers, or implying legal or compliance approval.

Modes

  • Theme Quote Set — default; grouped readable Markdown for one or more themes.
  • Specific Transcript — extract from user-named or supplied calls only.
  • JSON — use only when the user explicitly asks for machine-readable output or a downstream workflow requires it.

Output Format

Use readable Markdown by default.

# Customer Quotes

## [Theme]

**Coverage:** [returned]/[target] high-confidence customer/prospect quotes · [spread/quality note]

- “[Verbatim quote]”
  - **Speaker:** [Name or Unknown] · **Customer confidence:** [0.00] · **Theme relevance:** [0.00]
  - **Context:** [Company/call/date when known]
  - **Why it fits:** [Concise transcript-grounded evidence note]
  - **Source:** [Clickable transcript/call link, or no useful link available]

## Gaps

- [Theme with weak coverage, speaker ambiguity, transcript quality issue, or missing evidence]

## Internal / Non-Customer Evidence

- “[Verbatim fallback snippet]”
  - **Speaker / context:** [Known context] · **Role:** [non-customer classification] · **Customer confidence:** [0.00] · **Theme relevance:** [0.00]
  - **Usage note:** Useful for [internal purpose]; do not present as a customer quote.
  - **Source:** [Link/label]

---

{Follow the instructions and output format/conditions in [Limitations and Improvements](../index/SKILL.md#limitations-and-improvements)}

{Follow the instructions and output format/conditions in [Next Steps](../index/SKILL.md#4-next-steps)}

Omit empty optional sections. When both high-confidence customer quotes and fallback evidence exist, include fallback evidence only if the user asked for it or it materially explains a gap.

Include compact Method Notes when search breadth, transcript formatting, repeated calls, or speaker-label quality lowered confidence. Keep each theme's coverage explicit even when the result is 0/[target]. Do not print speaker_role_guess in readable summaries unless the user explicitly asks for role metadata or a quote needs an ambiguity caveat. Use clickable source links when available; otherwise say no useful link available.

Example Prompts

  • Find customer quotes about a feedback theme.
  • Find customer quotes about setup friction from recent enterprise calls.
  • Pull prospect quotes about data residency from the last month of call notes.
  • Find quotes that support the "faster account research" story from these transcripts.

Rules

  • Do not fabricate or infer quote text, speaker names, titles, roles, companies, dates, transcript links, or call metadata.
  • Do not treat every quote from a relevant call as relevant, or every relevant quote as customer-spoken.
  • Do not treat missing speaker labels as high confidence unless surrounding context strongly establishes an external participant.
  • Keep quotes verbatim and make provenance visible near each quote.
  • Prefer precision, diversity, and honest gaps over a padded list.
  • This workflow is not for exhaustive call analytics, exact speaker identity verification when transcripts lack evidence, or legal/compliance review.
  • Keep this workflow read-only. Do not post, send, package, or save quotes unless the user explicitly asks in a later step.

Failure Handling

If no transcript-like evidence is available, state that quote extraction cannot be grounded and ask for the smallest usable source. If transcript formatting is poor or speaker identity is ambiguous, return fewer quotes, name the limitation, and keep lower-confidence evidence out of the customer/prospect set.

來自 openai 的更多技能

user-context
openai
載入或管理資料分析插件的持久性來源路由偏好設定、入門邏輯、設定進度以及語義層註冊表。
official
notion-research-documentation
openai
研究Notion內容,並將其整合為結構化的簡報、報告或附有引用的比較分析。透過精準查詢搜尋並擷取Notion頁面,然後按主題組織發現,並附上內文來源引用及參考文獻章節。根據範圍與使用者目標,從四種輸出格式(快速簡報、研究摘要、比較分析、綜合報告)中選擇。使用內建範本建立與更新Notion頁面;直接連結來源,並在新資訊到來時追蹤變更。
official
rcsb-pdb-skill
openai
提交精簡的RCSB PDB請求以獲取核心元數據、Search API查詢及FASTA下載。當用戶需要簡潔的RCSB摘要時使用;保存原始JSON或…
official
pdf
openai
PDF 讀取、建立與驗證,支援視覺化渲染與程式化生成。使用 Poppler(pdftoppm)將 PDF 頁面渲染為 PNG,以便在交付前檢查版面、間距與排版。透過 reportlab 程式化生成 PDF,確保格式可靠;使用 pdfplumber 或 pypdf 提取文字與元資料。執行品質標準:無文字裁切、元素重疊、表格破損或渲染瑕疵;僅使用 ASCII 連字號,引用格式需具可讀性。使用...
official
test-coverage-improver
openai
Improve test coverage in the OpenAI Agents JS monorepo: run `pnpm test:coverage`, inspect coverage artifacts, identify low-coverage files and branches, propose…
official
playwright
openai
終端驅動的瀏覽器自動化,具備元素快照與互動式UI工作流程。透過playwright-cli包裝腳本運作(需npx),支援無頭模式與有頭模式以進行視覺化除錯。核心流程:開啟頁面、擷取快照以取得穩定元素參照、使用參照進行互動、在導航或DOM變更後重新擷取快照。包含表單填寫、點擊、輸入、多分頁管理、螢幕截圖/PDF擷取,以及流程除錯的軌跡記錄。元素參照(例如e3、e15)...
official
ukb-topmed-phewas-skill
openai
接受rsID、GRCh37或GRCh38輸入,解析為所需的GRCh38查詢,以獲取單一變異的簡潔UKB-TOPMed PheWAS摘要。當需要…時使用。
official
code-review-context
openai
模型可見上下文
official