cawplan-internal-qa-coding-humaninputs

CawPlan Internal QA — Classify One Human Input

Task

Given one piece of content (required) and, optionally, its paired assistant_message (this turn's AI reply) and prev_message (the last paragraph of the immediately previous assistant reply — not the previous human turn), classify using references/CATEGORY_TAXONOMY.md and references/TOPIC_TAXONOMY.md. This makes no CawPlan API calls and needs no cawplan auth — it is a pure reasoning check against uid.core-product's current classify rules, not a lookup.

Workflow

1. Read references/CATEGORY_TAXONOMY.md, references/TOPIC_TAXONOMY.md, and references/PRODUCTION_CLASSIFY_PROMPT.md (full production snapshot) if you haven't already this session. When the curated docs and the snapshot disagree, follow the snapshot.
2. new_feature is the last resort topic. When more than one topic plausibly fits, do not pick it — prefer the specific one (design_ui for anything about how something is displayed, including a NEW node/panel/screen whose ask is what it shows; integration_api for wiring or for a field an external system must carry; security, infra, refactor, bug, investigation). A new screen or a new field does not by itself make it new_feature.
3. Prepare assistant_message the same way production does before reading it — don't reason over raw untruncated text:
   • Normalize literal \n / \t escapes and line endings to real whitespace.
   • Strip [REDACTED] placeholders (upload-time redaction — not real bracketed content like [Done] / [QA Testing], which must stay).
   • Strip markdown noise within each paragraph (table rules, --- horizontal rules, heading/bold markers, emoji).
   • Split on blank lines into paragraphs; keep only the first 3 paragraphs.
   • Join those paragraphs with \n\n, then cap at ~1200 runes (head cut with ... if longer) — mirrors ClassifyAssistantSnippet / AISessionClassifyAssistantMaxRunes.
4. Prepare prev_message when provided (same paragraph split + per-paragraph sanitize as step 2, but take only the last paragraph; cap at ~500 runes) — mirrors ClassifyPrevAssistantTail / AISessionClassifyPrevMaxRunes. When omitted, treat prev as empty; bare follow-ups ("commit & push", "直接改") may be ambiguous without it.
5. Route categories exactly like production (see CATEGORY_TAXONOMY.md):
   • Primary signal: content. Always read this first.
   • prev: use ONLY on short bare follow-ups (typically ≤12 Han chars or ≤6 English words) to distinguish decision (human names/repeats the concrete action the assistant offered) vs approval (generic go-ahead: "直接改", "可以", "do it"). Never use prev for any other category; never treat it as a prior human requirement.
   • assistant_message: use for category ONLY when content is a bare hand-off (URL, file path, screenshot path, log path) with no intent words. Do not infer category from what the assistant already did or verified.
6. Determine every category from the taxonomy that clearly applies, then sort by the priority order in the reference doc (highest → lowest). Use the exact snake_case leaf names (direction_constraint, rejection_rollback, question_clarification, other_meta, … — not shortened forms like direction / rejection / question / other). Don't invent a secondary category just to fill the list. When only one applies, the list has one element.
7. The first element of that sorted list is the primary category.
8. Pick exactly one topic from TOPIC_TAXONOMY.md using content + prepared assistant_message only — never use prev_message for topic. Apply ASKING vs CHANGING, bare commit & push → git_ops, and Slack-heavy hints from the reference doc.
9. Write one short topic_reason sentence and topic_confidence (0.0–1.0 per TOPIC_TAXONOMY.md) mirroring production.

Output

Report:

• category — the primary category (single value, first element of categories).
• categories — the full priority-ordered list (may be a single value).
• topic — exactly one value from the 17-topic list in TOPIC_TAXONOMY.md.
• topic_reason — one short sentence explaining the topic pick.
• topic_confidence — float 0.0–1.0 (how sure you are about the topic pick).
• One short sentence explaining the primary category pick, citing the specific phrase in content, prev_message, or prepared assistant_message that drove the decision — this is what makes a caller's downstream disagreement (e.g. vs. an already-persisted cloud category) reviewable rather than opaque.

References

• references/PRODUCTION_CLASSIFY_PROMPT.md — full prompt exported from uid.core-product
• references/CATEGORY_TAXONOMY.md — curated category routing + examples
• references/TOPIC_TAXONOMY.md — curated topic routing + examples

Sync snapshot after prompt ships:

node skills/cawplan-internal-qa-coding-humaninputs/scripts/sync-classify-prompt-from-core-product.mjs