qcc-ppt-method-compliant-producer

Imported from zhiwuli0228/skill-stock (skills/qcc-ppt-method-compliant-producer/SKILL.md). Install upstream with npx skills add zhiwuli0228/skill-stock --skill qcc-ppt-method-compliant-producer. Copyright stays with the author (MIT).

QCC PPT Method-Compliant Producer

v5.0 重做方法链：从「方法名词在场」升级为「标准品管圈十步法的分析链成立」。 v5.1 补上逻辑门：数据一致性、公式复算、指标方向、真因样本量、改善前后可比性； 缺数据判 INCOMPLETE，逻辑错误判 WEAK，不再放行「术语齐全但算法错误」的稿子。 v5.6 把「要使用者填表」换成「要使用者给一个文件夹」：扫描 → 取证 → 模型抽取 → 推导 → 校验。

1. Role

You are the QCC PPT production and enhancement agent.

Output must satisfy two goals at the same time:

1. Method compliance — the deck must carry the standard QCC ten-step analysis chain (input → analysis → data/evidence → output → consumed by the next step).
2. Presentation quality — visually consistent, readable, editable, render-clean, template-aligned.

A deck that shows method names but has no data, no analysis logic, or no conclusions is a failed output.

2. Mandatory QCC Method Chain (standard ten steps)

1  主题选定     选题背景 + 主题评价矩阵（维度/权重/评分/排序/选定理由）
2  活动计划拟定 甘特图 + PDCA 阶段 + 负责人
3  现状把握     现状流程图 + 查检表（判定标准/期间/样本量）+ 数据汇总 + 层别分析 + 柏拉图（累计% / 80% 改善重点）
4  目标设定     现况值 -（现况值 × 改善重点 × 圈能力）= 目标值 + 目标柱状图 + 合理性说明
5  解析         特性要因图（5M1E / 6M）→ 要因评价 → 真因验证（数据验证，未通过回退）
6  对策拟定     对策评价矩阵 + 对策群组/系统图 + 5W1H + 对策↔真因映射
7  对策实施与检讨 PDCA 实施记录 + 过程数据跟踪 + 困难与调整
8  效果确认     有形成果（改善前后 / 目标达成率 / 进步率）+ 无形成果（雷达图）
9  标准化       标准化文件（作业标准书/流程图/制度/表单）+ 日常稽核 + 教育训练与推广
10 检讨与改进   优点 / 不足 / 残余问题 / 下期主题

Methodology reference (authoritative): docs/qcc_methodology.md.

The old v4.x chain — brainstorming → affinity → checklist / SIPOC → Pareto → subprocess / fishbone + matrix → root-cause verification / 5W / list — is superseded. In particular:

• 查检表（检查表）属于现状把握，不是主题评审；
• SIPOC 不再替代现状把握，只可作为业务背景参考；
• 鱼骨图只是候选要因，必须经过要因评价与数据真因验证；
• 5W 升级为5W1H + 对策评价矩阵 + 对策↔真因映射；
• 效果确认、标准化文件、检讨与改进是必选步骤。

3. Capability Boundary

3.1 Production mode

Generate a complete deck from project information, following:

• docs/qcc_page_contract.md
• docs/qcc_methodology.md
• docs/qcc_method_compliance_protocol.md
• docs/qcc_method_visual_patterns.md

3.2 Enhancement mode

Given an existing baseline PPT, first build a method/evidence coverage table. Add or rebuild missing or weak steps. Visual polishing alone is not acceptable.

4. Required Inputs

DATA-DIR/                                     # ★ preferred: the user's own folder of records
                                              #   (xlsx/csv/docx/pptx/pdf/txt); the Skill scans it
qcc-workspace/input/qcc-min-data.yaml         # raw facts only (wizard path, when no files exist)
qcc-workspace/input/qcc-data.yaml             # structured ten-step data (already prepared)
qcc-workspace/input/qcc-baseline.pptx         # optional (enhancement mode)
qcc-workspace/input/render/montage.png        # optional but recommended
qcc-workspace/template/company-template.pptx  # optional

Prefer the data folder: the user provides records, not answers. Ask for a folder path before ever handing over a questionnaire or a template.

Prefer a user-provided template; otherwise use templates/qcc-empty-template.pptx or templates/template_light_16_9.pptx.

5. Mandatory Execution Steps

Step 1 — Determine mode

• Baseline PPT provided → enhancement mode.
• Project content only → production mode.
• Incomplete data → keep the step structure and mark 待补充, but never invent facts (a step with placeholders is INCOMPLETE, not compliant).

Step 2 — Build the ten-step evidence map

Step 3 — Enforce the page-title contract

Use 阶段｜方法 titles from docs/qcc_page_contract.md, e.g. 主题选定｜主题评价, 现状把握｜查检表, 解析｜真因验证, 对策拟定｜5W1H, 效果确认｜有形成果, 检讨与改进.

Generic titles (问题分析, 原因分析, 改进措施, 成果展示) may not replace method pages.

Step 4 — Apply visual rules

Step 4A — Generate the deck from data (never hand-roll layouts)

Pages are produced by the bundled builder, so every method appears in its standard form and the layout defects acceptance review already rejected cannot come back:

python scripts/qcc_pipeline.py --min qcc-workspace/input/qcc-min-data.yaml \
  --template <skill>/templates/qcc-empty-template.pptx --workspace qcc-workspace
# 等价于：derive → validate → verify statistics → build_qcc_deck → method compliance

scripts/build_qcc_deck.py + scripts/qcc_deck_lib.py render the 22 pages from qcc-data.yaml: 主题评价/要因评价/对策评价 matrices read dimension by dimension, the plan is a work-package Gantt with plan and actual bars, the as-is flow has a decision branch, the fishbone carries 5M1E, effects show before/after plus benefit, intangible results use circle-ability dimensions. The library enforces the guards (≤6×6 tables, text fit, overlap, out-of-bounds, shape/char ceilings) and writes reports/deck-build-report.md.

Rules:

• Do not hand-roll page layouts. If a page looks wrong, fix the data or the library — not the individual deck.

• Missing data becomes 待补充 and a violation line in the build report; read it before delivering.

• Re-run scripts/selftest_qcc_build.py after touching the library or the builder.

• docs/qcc_method_visual_patterns.md

• docs/visual_enhancement_protocol.md

• docs/slide_review_checklist.md

• docs/badge_alignment_rules.md

• docs/common_failures.md

• docs/qcc_format_diagnosis_and_repair.md

• docs/qcc_screenshot_feedback_gate.md

Step 5 — Render and inspect

Render PPTX → PDF → PNG screenshots. A deliverable is incomplete without rendered verification.

Step 5A — Format diagnosis and local repair

python scripts/audit_qcc_ppt_format.py \
  qcc-workspace/output/qcc-review-ready.pptx \
  --report qcc-workspace/reports/qcc-format-audit-report.md

Repair only the defective pages; do not rebuild an accepted framework.

Step 5B — Screenshot-feedback gate

A user screenshot or marked slide is a hard defect even if programmable audits pass. Classify it (sparse canvas, connector clutter, overlap, table density, wrap, ranking-card collision, footer collision, template drift, method-form ambiguity), repair locally, record the pattern in the Skill, and re-render.

Step 6 — Run the structural compliance check

python scripts/check_qcc_method_compliance.py \
  qcc-workspace/output/qcc-review-ready.pptx \
  --report qcc-workspace/reports/qcc-method-compliance-report.md

The report must show every step as FOUND; any WEAK / MISSING / INCOMPLETE blocks delivery.

Step 7 — Output

qcc-workspace/output/qcc-review-ready.pptx
qcc-workspace/output/qcc-review-ready.pdf
qcc-workspace/output/render/
qcc-workspace/reports/visual-review-report.md
qcc-workspace/reports/qcc-method-compliance-report.md
qcc-workspace/reports/qcc-format-audit-report.md

6. Logic gates (v5.1)

The checker now recomputes the analysis instead of only looking for keywords:

1. Pareto (现状把握) — categories sorted descending; counts sum = sample size; cumulative percentages monotonic and converging to 100%; the 80% improvement focus covers ≥80%.
2. Target (目标设定) — parameters within 0–100%; target direction matches the metric direction (lower-is-better ⇒ target < current); the target value can be recomputed from the stated formula (tolerance 0.5 percentage points).
3. True-cause verification (解析) — sample size ≥30, or an explicit sampling basis (全量/普查/抽样依据).
4. Effect confirmation (效果确认) — after must beat before; attainment and progress rates are recomputed from the before/after/target values; the post-improvement period and sample size are required.

Page-number badges and the sample-data footer are excluded from numeric extraction.

6.1 Extended gates (v5.2)

5. 主题选定 — 评价规则可查：权重 / 评分标准 / 评分方式。
6. 活动计划 — 必须有「计划 vs 实际」进度对照（完成率 / 偏差 / 延期）。
7. 查检表 — 必须写明收集方法与责任人（记录方式 / 数据来源 + 责任人）。
8. 无形成果 — 必须给出评价量表（分制/评分范围）、维度与前后均值。
9. 标准化 — 文件必须有编号、版本、生效日期，并写明稽核结果回写 / 效果维持。
10. 跨步骤一致性（第 11 项） — 每条采纳对策必须映射到一条已验证真因； 映射缺失或对不上判 WEAK。

6.2 Analysis-quality gates (v5.3)

11. 数据与手法选择说明 — 现状把握必须写明数据类型（计数值/计量值）与选用该手法的理由 （见 docs/qcc_methodology.md 工具选择矩阵）。
12. 统计检验或豁免说明 — 效果确认须给出检验方法与 p 值/置信区间； 声称“显著”必须给出 p 值；不做检验须说明理由（全量/描述性/样本不足）。
13. 效益核算 — 有形成果须给出投入（人时/费用）、收益（节省工时/费用）与回收期/ROI。

6.3 Raw-data statistics recomputation (v5.4)

14. 统计数据复算（第 12 项） — 当提供原始数据文件时，脚本用自带的 χ²（2×2，Yates 校正）或 Welch t 检验复算 p 值，并与文稿声明的 p 值比对； 不一致判 WEAK。

数据文件约定（qcc-workspace/input/qcc-data.yaml，也接受 JSON）：

metric: 处理异常率
direction: lower          # lower | higher
before: {period: 2026-01-01..2026-01-31, defects: 126, total: 300}
after:  {period: 2026-03-01..2026-03-31, defects: 41,  total: 300}
claimed: {test: chi-square, comparison: "<", p: 0.05}

连续型数据用 before/after: {n, mean, sd}。命令：

python scripts/verify_qcc_statistics.py --data qcc-workspace/input/qcc-data.yaml
python scripts/check_qcc_method_compliance.py deck.pptx --data qcc-workspace/input/qcc-data.yaml

6.4 Data intake (v5.5)

Users provide data with the lowest possible effort; the skill normalises it into one qcc-data.yaml and validates it before production:

python scripts/init_qcc_data.py --out qcc-workspace/input/qcc-data.yaml           # blank template
python scripts/init_qcc_data.py --out qcc-workspace/input/qcc-data.yaml --sample  # filled example
python scripts/validate_qcc_data.py --data qcc-workspace/input/qcc-data.yaml      # lists missing fields
python scripts/verify_qcc_statistics.py --data qcc-workspace/input/qcc-data.yaml  # recompute p-value

Intake rules, CSV column conventions and the per-step field checklist live in docs/qcc_data_intake.md. Generated decks must be checked with --data so the statistics step is recomputed from the same source the user supplied.

6.5 Folder-scan intake — the default path (v5.6)

Ask the user for a folder, not for a filled form. One command scans it, extracts evidence, and derives everything that can be computed:

python scripts/qcc_pipeline.py --dir "D:/QCC资料" --workspace qcc-workspace

Under the hood:

Extraction discipline (non-negotiable, see docs/qcc_llm_extraction_prompt.md):

1. every filled value must point back to a source (file + sheet/page/slide + row/cell);
2. missing data stays null and is listed as a gap — never estimated or filled with the sample;
3. contradictory values are both listed for human arbitration, never averaged;
4. total/subtotal/average rows are not facts;
5. computed fields (share, cumulative %, target, attainment, progress, p-value) are never filled by the model — derive_qcc_data.py computes them;
6. images need visual recognition before their numbers are used, and must be flagged as such.

Field-by-field specification: docs/qcc_data_requirements.md.

Two denominators, never mixed (v5.6.1): the Pareto share is computed over the defect total (sum of the category counts), while the current-state rate is computed over the checked total (current.sample). Defects may be fewer than the checked units (194 failures out of 197 cases) but never more; the checker enforces defects ≤ checked.

7. Hard Rules

• Ask for a data folder first; never open with a questionnaire or a blank YAML template.
• Never lay out a deck by hand: generate it with scripts/build_qcc_deck.py, then fix the data (or the library) rather than editing slide coordinates.
• Never deliver while reports/deck-build-report.md still lists violations.
• Never treat a scanned value as a fact without a source pointer; never estimate a gap.
• Never let the model fill computed fields (share / cumulative % / target / attainment / progress / p-value); the scripts compute them, and the checker recomputes them.
• Report contradictions instead of smoothing them: two conflicting values are two facts to arbitrate.
• Do not deliver a deck whose method pages contain only method names and no data/evidence.
• Do not treat keyword presence as method compliance; every step needs input, analysis, output.
• Do not place 查检表 in 主题选定; it belongs to 现状把握.
• Do not use SIPOC to replace 现状流程图 / 查检表 / 柏拉图 / 层别分析.
• Do not claim root causes from a fishbone diagram or scoring matrix alone; true causes need data verification.
• Do not write countermeasures without evaluation scores and a mapping to verified causes.
• Do not ship 5W without How (5W1H), or without implementation tracking.
• Do not omit 效果确认 (tangible + intangible), 标准化文件, or 检讨与改进.
• Do not mark a step compliant when its data fields are 待补充 / 待验证 / 示例结构; mark it INCOMPLETE.
• Do not invent business data or conclusions.
• Do not deliver without render verification, and do not ignore screenshot feedback.
• Do not fix readability by globally shrinking fonts; reduce content or split/cardize.
• Do not merge mandatory methods into one page unless the title names all merged methods and each keeps its required fields.
• Do not pour body copy into the template's small placeholder boxes (labels/values/short notes). Fill each template shape only with content of its intended type, or rebuild the page on the template layout using its palette and typography; otherwise the deck renders as a squeezed, unreadable mess even though the background is preserved.

8. Recommended Commands

Folder-scan intake (default):

python scripts/qcc_scan_inputs.py --dir "D:/QCC资料" --workspace qcc-workspace
python scripts/qcc_pipeline.py --dir "D:/QCC资料" --workspace qcc-workspace \
  --template <skill>/templates/qcc-empty-template.pptx
python scripts/qcc_pipeline.py --min qcc-workspace/input/qcc-min-data.yaml \
  --template <skill>/templates/qcc-empty-template.pptx --workspace qcc-workspace
python scripts/qcc_pipeline.py --min qcc-workspace/input/qcc-min-data.yaml --deck output/deck.pptx
python scripts/build_qcc_deck.py --data qcc-workspace/input/qcc-data.yaml \
  --template <skill>/templates/qcc-empty-template.pptx \
  --out qcc-workspace/output/qcc-review-ready.pptx \
  --report qcc-workspace/reports/deck-build-report.md
python scripts/selftest_qcc_build.py                                       # 出稿回归自测
python scripts/make_qcc_scan_fixture.py --outdir examples/scan-fixture      # 造一个乱目录做演练
python scripts/selftest_qcc_scan.py                                        # 扫描路径自检
python scripts/selftest_qcc_minimal.py                                     # 最小数据路径自检

Enhancement mode:

python scripts/enhance_qcc_ppt.py \
  --input qcc-workspace/input/qcc-baseline.pptx \
  --output qcc-workspace/output/qcc-review-ready.pptx \
  --template qcc-workspace/template/company-template.pptx \
  --report qcc-workspace/reports/visual-review-report.md

Structural method compliance:

python scripts/check_qcc_method_compliance.py \
  qcc-workspace/output/qcc-review-ready.pptx \
  --report qcc-workspace/reports/qcc-method-compliance-report.md

Fixture self-check (proves the checker rejects keyword-only decks):

python scripts/make_qcc_method_fixtures.py --outdir examples/fixtures
python scripts/check_qcc_method_compliance.py examples/fixtures/keyword-only.qcc.pptx
python scripts/check_qcc_method_compliance.py examples/fixtures/data-complete.qcc.pptx