Customer Questions Remind

Customer Questions — Remind Czar + Draft from KB (Hybrid)

Read scraped customer questions, classify whether each thread has been substantively answered, and emit reminders.json. For threads that still need follow-up (no-reply / ack-only), also draft a starter answer from the curated KB at .claude/triage/customer-faq.md — but only when drafting is enabled (see ENABLE_DRAFTS under Context Variables).

The downstream post-to-slack job posts a single internal Slack message listing the threads that still need a reply. When the JSON contains drafts, they are attached as threaded replies under each reminder.

You do NOT post to Slack. Your only outputs are the JSON file and a base64 emit of it on stdout.

Budget: ~30 tool calls for ≤10 questions, ~50 for up to the cap. Batch reads. Do NOT re-read the same file twice.

Context Variables

Injected via the workflow prompt:

• QUESTIONS_PATH — local file path to questions.json (input)
• DRAFTS_PATH — local file path where you must write the JSON output (the filename is drafts.json for compatibility with the post script)
• MAX_QUESTIONS — process at most this many questions
• ENABLE_DRAFTS — "true" or "false". Gates draft emission in Step 2b. When "false" (default), classify every question but emit null draft fields with empty flags — the KB quality bar is not yet high enough for prod. When "true", run Step 2b as documented.

Slack channel routing (test_mode workflow input) is handled by the downstream post-to-slack job and is not exposed to you. Your behavior does not depend on the target channel.

Critical Rules

• NEVER post to Slack or GitHub from this skill. The downstream post-to-slack job handles posting.
• NEVER address the customer or post on their original thread. Drafts are for internal czar review only.
• Liberal classification. When in doubt whether a reply is substantive, err toward no-reply or ack-only — over-reminding is cheap; missing a real unanswered customer question is not.
• Drafts ONLY when ENABLE_DRAFTS == "true" AND status is no-reply / ack-only. Skip drafting entirely when ENABLE_DRAFTS == "false" and, when drafting is enabled, also skip for answered (already handled) and internal-noise (not a real question).
• Drafts MUST be grounded in the KB. If no FAQ entry matches, emit draft_answer: null rather than hallucinating. A null draft + kb_miss flag is more useful than a guess.
• Drafts are short. 2–4 sentences. No preamble ("Great question!"), no padding. Front-load the actionable fix.

Step 1: Read input and the KB (~2 tool calls)

Read $QUESTIONS_PATH. The JSON shape is:

{
  "questions": [
    {
      "id": "slack-CXXXXX-1731349200.000100",
      "medium": "slack" | "github",
      "channel_or_repo": "ai-hub-models",
      "thread_url": "https://qualcomm.slack.com/archives/...",
      "submitted_by": "Customer Name",
      "submitted_at": "2026-06-22T14:30:00Z",
      "title": "",
      "question": "full question text",
      "thread_replies": [
        {"user": "name", "is_internal": true, "text": "reply text"}
      ],
      "has_internal_reply": true
    }
  ]
}

Read .claude/triage/customer-faq.md. Each entry has the shape:

### faq-<id>

**Triggers:**
- "phrase 1"
- "phrase 2"
- mentions of `code_token`

**Question shape:** <one-line description>

**Answer:**
<canonical answer body>

**Citations:**
- `path/to/file.py:NN`
- https://...

**Confidence floor:** high | medium | low
**Last updated:** YYYY-MM-DD

Hold the KB in working memory for Step 2. Do not re-read it for each question.

Step 2: Per-question classification + drafting

For each question (up to MAX_QUESTIONS):

2a. Classify (always required)

Pick exactly one status:

Substantive vs ack — read every internal reply for intent, not keywords.

Before applying any keyword shortcut, decide what each internal reply is doing. Every reply falls into exactly one bucket:

1. Handoff / routing — an internal user pings another engineer to answer. Signature: contains a Slack @mention (<@UXXX>) AND a routing verb: "chime in", "any chance you can help", "take a look", "can you look at this", "please help", "assist", "point to the right person", "who can answer this". A handoff is always ack-only contribution — even if the same message also contains a URL, a version number, or a channel name. The URL is context for the pinged engineer, not an answer to the customer.
2. Ack / info-gathering — "got it, looking", "checking with X", "let me find someone", or a reply that asks the customer for more info ("what QAIRT version?", "which device?"). ack-only contribution.
3. Redirect — "please ask in <Discord / other Slack channel / forum>" without also answering the question. ack-only unless the customer's question was literally "where do I ask?" (in which case the redirect IS the answer).
4. Substantive answer — names a fix, gives a specific number, links a doc/PR/release that directly resolves the question, gives a specific command/flag, or definitively says "not supported on X".

Per-thread rule: the thread is answered only if at least one internal reply is bucket #4. If every internal reply is #1, #2, or #3, the thread is ack-only. no-reply is reserved for zero internal replies.

Common false-positive traps (all → ack-only, NOT answered):

• Handoff that also cites a URL for the pinged engineer's reference (e.g. "<@Eng> any chance you can help — see this thread"). The URL isn't answering the customer.
• Redirect to Discord / Qualcomm Developer forum / another Slack channel when the customer's technical question is still open.
• A reply that only lists next-step questions to ask the customer — that's info-gathering, not an answer.
• Multiple handoffs stacked in one message (e.g. <@A> <@B> can one of you chime in). Still ack-only regardless of URL count.

Common false-negative traps (all → answered, NOT ack-only):

• A short internal reply that names a tool + quickstart URL (or a released app URL) as the answer, even when the customer's question was multi-part. The reply can be higher-altitude than the ask — "use tool X, or the shipped app" supersedes sub-questions about which flag / preprocessing step of tool Y.
• A one-line reply that redefines the recommended path ("this is now the supported way") + URL. If the URL directly resolves the question, it's bucket #4 regardless of reply length.

Only after ruling out #1–#3 for every internal reply, apply keyword shortcuts to confirm bucket #4:

• "use v0.56.0", "set --target-runtime onnx", "fixed in #1234", "that model isn't supported on X chip" → substantive answer.
• A reply containing a specific number, file path, URL, or command without handoff / redirect signals → substantive answer.

Age calculation: compute age_hours as the integer hours between submitted_at and the current UTC time. Treat any parse failure as 0.

Topic tag (1–2 words): for every classified question — including answered and internal-noise — emit a topic string of 1–2 words that captures what the thread is about. The czar uses this to scan the reminder list at a glance without having to click each thread or read the verbatim question.

Good topics (concrete, scannable):

• qwen3 export, genie crash, sa8295 quant, tps regression, release assets, context length, qairt version, byom, melotts, npu dual, dlc adapter, windows arm, ubuntu setup, gemma roadmap, aimet config

Avoid topics that are too generic to be useful:

• bug, help, error, question, support (these don't narrow anything)

If the question is internal-noise (a bump, channel chatter), use noise or bump as the topic — the czar will skip those.

2b. Draft (drafting-enabled only; only for no-reply and ack-only)

Gate: if ENABLE_DRAFTS == "false" (the default), skip this section entirely — for every question (regardless of status), emit draft_answer: null, kb_citation: null, confidence: null, flags: []. Do not read the KB. The reminder-only path is the default behavior until the KB quality bar is met.

Otherwise (ENABLE_DRAFTS == "true"), for each question classified no-reply or ack-only, attempt to draft a starter answer from the KB:

1. Match against FAQ triggers. Scan each FAQ entry's Triggers block for substring / keyword overlap with the customer's question (and the thread context). A match means the question is the same shape — not necessarily verbatim. If 2+ entries match, prefer the one with the most specific triggers.
2. If a match is found:
   • draft_answer: 2–4 sentences. Paraphrase the FAQ's Answer body, adapted to the customer's specific wording (their chipset, their model, their error). Front-load the fix. Drop preamble.
   • kb_citation: the FAQ id (e.g. faq-qwen3-x2-elite-assets).
   • confidence: copy from FAQ's Confidence floor.
   • flags: []
3. If no FAQ matches:
   • draft_answer: null
   • kb_citation: null
   • confidence: null
   • flags: ["kb_miss"]

Drafts NEVER:

• Address the customer ("Hi ")
• Mention internal-only details (memory, conversation context, agent name)
• Include code blocks longer than 3 lines (link to docs instead)
• Speculate beyond what the FAQ says

Drafts SHOULD:

• Lead with the actionable fix or the recommended command/version
• Cite the FAQ's Citations inline if relevant
• Stay under ~80 words

For answered and internal-noise questions: do not draft. Leave draft_answer, kb_citation, confidence as null and flags as [].

Step 3: Output schema

Write to $DRAFTS_PATH:

{
  "drafted_at": "<UTC ISO8601>",
  "drafts": [
    {
      "question_id": "<from input id>",
      "thread_url": "<from input>",
      "channel_or_repo": "<from input>",
      "submitted_by": "<from input>",
      "question_excerpt": "<first 200 chars of question text, single-line>",
      "topic": "<1-2 word category, e.g. 'qwen3 export'>",
      "age_hours": 27,
      "status": "no-reply" | "ack-only" | "answered" | "internal-noise",
      "draft_answer": "<2-4 sentence draft, or null>",
      "kb_citation": "<faq-id, or null>",
      "confidence": "high" | "medium" | "low" | null,
      "flags": []
    }
  ]
}

Field notes:

• Include every classified question, including answered and internal-noise. The post script filters to remind-worthy statuses; keeping the full list in the artifact makes review easier.
• question_excerpt: replace internal newlines with a single space, then take the first 200 chars. No mrkdwn formatting. Retained in the JSON artifact for forensic review even though the parent Slack message no longer prints it.
• topic: 1–2 words, see the topic tag guidance above. This is what the parent Slack bullet shows in lieu of the question excerpt.
• flags may include kb_miss (no FAQ matched) or kb_partial (matched but the FAQ only partially addresses the question). Leave [] otherwise.

After writing, print one summary line: Classified N questions: <a> no-reply (<d1> drafted, <m1> kb_miss), <b> ack-only (<d2> drafted, <m2> kb_miss), <c> answered, <d> internal-noise.

Step 4: Emit the JSON for the downstream post-to-slack job

The reusable agent workflow does NOT propagate SLACK_NOTIFIER_TOKEN to your shell steps. The downstream post-to-slack job recovers the JSON from this log via known markers.

Run this as ONE Bash call (the Breeze allowlist permits python3 but not echo / base64 / compound shell — the helper wraps both):

python3 scripts/breeze_customer_questions/emit_drafts_b64.py "$DRAFTS_PATH"

Step 5: Done

Do not open a PR. Do not post on the customer's original thread. Do not post drafts directly to the customer. Your output is: $DRAFTS_PATH (the file) plus the base64 emit on stdout. The downstream post-to-slack job posts the reminder message + thread replies and uploads the JSON as a workflow artifact.

Examples (mental model)

Question: "We can't load Qwen2-7B on Snapdragon X2 Elite — release-assets.yaml missing." Thread replies: [{user: "Eng A", is_internal: true, text: "Looking at this — will follow up."}] KB match: faq-qwen3-x2-elite-assets → status: ack-only, topic: "qwen3 export", draft_answer: "Qwen2-7B is old and we're not adding precompiled assets — recommend migrating to Qwen3-4B-Instruct-2507. Assets for X2 Elite are at https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/qwen3_4b_instruct_2507/releases/v0.56.0/qwen3_4b_instruct_2507-genie-w4a16-qualcomm_snapdragon_x2_elite.zip", kb_citation: "faq-qwen3-x2-elite-assets", confidence: "high".

Question: "Where is the prebuilt UNet QNN bundle for v0.48.0?" Thread replies: [{user: "Eng B", is_internal: true, text: "https://qaihub-public-assets.s3... — that's the v0.48.0 UNet bundle."}] → status: answered (specific URL given). topic: "release assets", draft_answer: null, kb_citation: null, confidence: null.

Question: "@here just bumping this — any update?" Thread replies: any. → status: internal-noise. topic: "bump". draft_answer: null.

Question: "When will you support Llama 4?" Thread replies: [], has_internal_reply: false. KB match: none (roadmap question). → status: no-reply, topic: "llama roadmap", draft_answer: null, kb_citation: null, confidence: null, flags: ["kb_miss"].

Question: "Why does my export fail with BQ is not supported on SA8295?" Thread replies: []. KB match: faq-sa8295-bq-not-supported → status: no-reply, topic: "sa8295 export", draft_answer: "Upgrade to QAIRT 2.45 or later. The 'BQ is not supported' error on SA8295 (v68) was resolved in 2.45; not a model issue, a QAIRT version gap.", kb_citation: "faq-sa8295-bq-not-supported", confidence: "high".

Question: "How do I derive the 100 Dense INT8 TOPS spec for the IQ-9075 from profiler metrics?" Thread replies: [{user: "Czar", is_internal: true, text: "<@Eng> any chance you can help here? <@customer> since this is device-specific, you'll likely have better luck asking in Qualcomm Developers Discord (https://discord.com/invite/qualcommdevelopernetwork)."}] → status: ack-only (handoff + redirect, no substantive answer). The Discord URL is a redirect, not an answer to the TOPS math. topic: "tops derivation", draft_answer: null, kb_citation: null.

Question: "What's the right workflow for running a VLM on-device via GenieX (QAIRT)? I've tried genie-t2t-run.exe (text-only) and see genie-app-script.txt in the bundle. Is genie-app.exe the intended VLM path? What's the preprocessing pipeline for a custom image (.raw files)?" Thread replies: [{user: "Eng", is_internal: true, text: "Now you can run LLM/VLM directly via GenieX (https://github.com/qualcomm/GenieX#quickstart), or try the released app GenieX Chat Windows (https://aihub.qualcomm.com/apps/geniex_chat_windows)."}] → status: answered. Bucket #4: fresh URLs, no @mention handoff, no info-gathering, no forum redirect. The one-line reply is higher-altitude than the multi-part ask — "use GenieX quickstart or the shipped app" supersedes the sub-questions about which .exe and which preprocessing step. A short URL-directed reply can answer a long question. topic: "geniex vlm", draft_answer: null, kb_citation: null.