lingyu9495-source/feasibility-report-engine

Feasibility Study Report Engine (Industrial Project Feasibility Study Report · Complete Battle Manual for External Delivery)

Skill Type Note: This skill = the "Word Full Version + Excel Calculation Model" delivery version of feasibility study report writing. Deliverables: ① Complete Word report (cover → main text → appendices, per NDRC 2023 outline) ② Excel investment calculation model (all formulas pre-built). Related skills in the same business family (distinguished by delivery format): PPT version feasibility presentation, PDF version feasibility report — to be developed separately (planned, see README skill roadmap). Quick Usage: Client/project team provides materials → This engine executes six steps → Deliver the three-piece set (report .docx + calculation model .xlsx + preparation notes .docx) to the user chat. This engine is the full-process practical crystallization from a real large order in September 2026 (300 billion ocean industry cluster, nine volumes, 200,000 Chinese characters), including all pitfalls encountered.

v1.6 Capability Upgrade (This Update): ① Added report_qa.py for long-document QA (caching, assertion check, regex-based search and replace) and pipeline_check.py for pipeline self-check, pure Python standard library, zero dependencies, ready to run out of the box; ② Added "Three Gates of Report Quality" and "Pre-Delivery Verification Checklist," transforming "delivery by feel" into "delivery by ticking off each item"; ③ Supplemented "Hard Rules for Pagination and Images (Word rendering ≠ LibreOffice rendering)" and PDF delivery; ④ Added a chapter on "Revision and Data Alignment" to prevent disasters when clients return drafts with changes or updated numbers; ⑤ Added financial calculation data consistency check table, four-statement cross-validation check, and sensitivity analysis requirements.

I. Trigger Conditions and Positioning

Trigger: Client/entrusting party requests a "feasibility study report," "feasibility study," "project initiation report," or "financing due diligence report"; the client provides scattered materials (project name/location/scale/a few images) and wants a formal report suitable for investor review, regulatory approval, or investment promotion.

Division of Labor with Related Skills:

Skill	Purpose
This Engine	For external delivery to clients (compliance gold standard = NDRC 2023 outline)
report-production-pipeline	General report production pipeline (five stages + self-check scripts); steps 2–4 of this engine are isomorphic with it
feasibility-study-writing	For internal decision-making/business case argumentation
mh-calculation	Financial calculation engine (17 tables; this engine calls it for calculations)

Commercial Delivery Red Lines (Invisible OEM) :

• The client signs their name on the deliverable; we only provide the content—the preparing unit/date/seal are left blank for the client to fill in.
• Absolutely no AI traces (file names must not contain "final version/AI/dual-model review," metadata is cleared, no AI symbols in the main text).
• Qualification seals: Advise clients to collaborate with qualified engineering consulting units for stamping (compliance red line + business acquisition opportunity).
• Pricing Anchor: Sell "speed + professionalism." Custom standard tier: 8,000 to 15,000 RMB. Do not take orders below 3,000 RMB.

II. Industry Gold Standard: Two Outlines (Determine the Type Before Writing)

The National Development and Reform Commission's Document No. 304 [2023] "Outline and Instructions for Writing Feasibility Study Reports for Investment Projects" — the only compliance benchmark for writing feasibility studies.

Type Determination	Government Investment Projects (General Outline, 11 Chapters)	Enterprise Investment Projects (Reference Outline, 10 Chapters)
Funding Source	Government investment/capital injection/special bonds	Self-raised capital/market-based financing
Typical Clients	Municipal investment/platform companies/government projects	Private enterprises/industrial investment/mixed ownership
Emphasis	Construction necessity/social benefits/fund balance	Market demand/profitability/business model

For the complete chapter-by-chapter requirements, see references/ (gov-dagang-2023.md for government version, ent-dagang-2023.md for enterprise version). Consult the corresponding outline chapter by chapter when writing.

Type Determination Self-Check (Getting It Wrong = Rewriting the Entire Report. Answer These Four Questions First)

1. Where does the money come from: Fiscal funds/special bonds/government platform investment → Government version; Self-raised capital/commercial loans/market-based financing → Enterprise version.
2. Who is the audience: NDRC approval/authorization, fiscal fund application → Government version; Bank loan review, investor decision-making, investment promotion → Enterprise version.
3. Whose need does it address: Public need (regional shortcomings/livelihood/mandatory superior planning) → Government version; Market demand and profitability → Enterprise version.
4. Does it require these chapters: Performance management plan, carbon peak and carbon neutrality analysis → exclusive to the government version; Higher weight for business model/marketing strategy/competition analysis → Enterprise version.

For mixed funding (government capital + self-raised funds): Use the approval/filing outline as the main framework. Incorporate enterprise financial analysis into the "Investment, Financing, and Financial Plan" chapter, and clarify the dual approach in the overview.

III. Six-Step Workflow (Full Process)

Step 0: Material Inventory + Type Determination + Three Pre-Engagement Questions (0.5h)

Ask these three things before writing:

1. Purpose (Project approval/bank financing/investment roadshow/internal decision-making) → Determines layout depth and outline.
2. Owner Nature (Government platform/private enterprise) → Determines whether to use the government or enterprise outline.
3. Deadline → Determines Mode A (complete, 7-15 days) or Mode B (rapid first draft delivery, 1-3 days). ⭐

Material Inventory: Identify "what we have vs. what's missing." Ask for all missing items at once, not back and forth. Missing investment estimates → Use industry benchmarks and note "per [industry] convention." Missing land/location → Use public planning due diligence. Client provides little information → Use Mode B.

Completion Criteria: One sentence documenting the type determination conclusion (government/enterprise version + basis); a written list of material gaps; Mode A/B clearly specified.

Step 0.5: Data Request List (Three Batches, Ask for Everything at Once, Don't Write and Ask Alternately)

Batch	Content	Consequences of Missing
Batch 1: Quantity, Price, and Investment Basics	Construction content and scale, equipment list and quotations, land price/rent, utility costs (water/electricity/gas), labor costs	Can only produce a "preliminary justification version"; investment estimates are based on industry conventions
Batch 2: Entity and Compliance Confirmation	Business license/main entity documents, land pre-approval and site selection, env/energy/water/stability assessments, sea area use approvals and processing status	Compliance sections can only state "in process"; the review meeting will criticize this
Batch 3: Market and Order Evidence	Letters of intent/framework agreements, upstream/downstream contracts, operational data from similar projects, policy support documents	Market analysis can only use public data; depth is significantly reduced

Template list: references/data-request-list.md. When sending the list, simultaneously explain: "If Batch 2 is missing, only a preliminary justification version can be provided" — this is both professional and an opportunity for secondary billing.

Step 1: Online Due Diligence (Mandatory! 1-3h, Four Parallel Streams)

Never write without searching. The difference between a "presentable" feasibility study and "template copy-paste" is that every assertion has data/policy support.

1. Policy Stream: Provincial/city industrial policies/special plans/industry access requirements → Search for "[Industry] 15th Five-Year Plan" "[Province/City] Industrial Policy [Industry]"
2. Market Stream: Industry scale/growth rate/competitive landscape → Brokerage reports/industry associations/statistics bureaus; cross-reference multiple sources
3. Regional Stream: Site location park/land/infrastructure/similar projects → Search for "[City] [Industry] Industrial Park"
4. Benchmark Stream: Investment intensity/return rate of similar projects → Refer to mh-calculation industry benchmarks

Record the source for every key piece of data (institution + year + link). Cite sources in the text. For missing data, do not fabricate; write "to be confirmed via due diligence" and provide the industry range.

Data Extraction Channels (Escalate by Page Difficulty; Don't Give Up at the First Level) :

• Level 1: Standard scraping/search engine summaries — Original policy documents, news, public reports.
• Level 2: Scraping with anti-bot bypass (rendering JS, handling Cloudflare challenges) — Government websites, industry association sites, research report sites.
• Level 3: Browser control (clicking, pagination, forms) — Data requiring interaction (query systems, public notice databases).
• Save raw materials to disk (under <project name>/due-diligence/ named by source) for traceability during writing and to avoid re-searching during revisions.

Completion Criteria: For every key number, you can answer "Who said it, which year, what measurement." Raw materials are saved on disk and traceable.

Step 2: Distributed Writing (Model Matrix — This Is the Key to Producing a 200,000-Character Epic)

Do not let one model consume the entire document in one pass. Assign roles:

• Master Brain (the model in the current conversation): Build the structural skeleton, write the overview/introduction/strategic argument, perform final assembly and review.
• Expansion Model (a more capable model API, e.g., DeepSeek V4 Pro / Qwen Max): Deep-dive expansion of heavy chapters like financial evaluation and market analysis — feed each chapter with "current status of this chapter + expansion requirements + example." Set max_tokens high (12000+, otherwise reasoning tokens eat up all output).
• Calculation Model: Invoke mh-calculation to compute IRR/NPV/payback period (do not calculate manually).
• Review Model (a smart model reads the full text to find flaws): Using Qwen/Doubao, have it assume the role of a senior practitioner and read the entire text (not a summary) to find logical gaps, inconsistent data, and distorted numbers.
• Visual Model (for layout verification, see Step 4): Cross-validate using Mimo 2.5 / DeepSeek Vision / GLM-4V / Qwen-VL.

Tip: The same API key can be used for two tiers (Flash for daily tasks, Pro for deep work). The reviewer must be given the full real text, not a summary, or the review will be worthless.

Before writing, produce a chapter outline (Each chapter = what to say | why this evidence supports it | where the evidence comes from). Send chapters with blank evidence columns back to Step 1 to fill in. Do not start writing with empty evidence. Template for the outline: templates/report-plan.md.

Completion Criteria: The chapter outline has all three columns filled with no blanks; each chapter's first draft has data, sources, and a closed logical chain.

Step 3: Three Gates of Quality (Between the AI Draft and a Draft You Dare to Deliver — Don't Skip)

Gate	What to Do	What to Do If It Fails
① Reflection	Review the process: Is the data complete? Are the sources reliable (primary/secondary)? Are there logical leaps?	Add data, change sources, fill in arguments
② Review	Check against the checklist item by item: Do conclusions match evidence? Is there any overreach? Are opposing views addressed? Are numbers/names/dates accurate? Is the terminology consistent?	Correct each item; cannot be "approximately correct"
③ Correction	After fixing, must re-scan the corrected areas (key focus: check for new errors introduced) and re-scan the entire text for residual old terminology	If the re-scan fails, repeat the entire process

Checklist version: templates/review-checklist.md. Machine-executable bottom-line check:

python3 scripts/pipeline_check.py report.md          # Source annotation / uncertainty disclosure / opposing perspective / conclusion / structure / number terminology
python3 scripts/pipeline_check.py report.md --strict --json   # Non-zero exit code = hard errors; can be directly integrated into automation

Drafting and review cannot be done in the same pass — it's a cognitive principle, not an attitude problem, that you can't see errors in your own writing. Change the perspective/change the model/change at least one pass.

Step 4: Document Engineering (From AI Draft to Professional Delivery Version, via scripts/)

Delivering a docx directly generated by AI is a hallmark of an amateur. Use the scripts pipeline:

# 1) Build the skeleton from scratch: cover page + signature page + auto table of contents + header/page numbers (style system built-in)
python3 scripts/gen_report_docx.py "Full Project Name" "" output.docx
#    (middle parameter = preparing unit; pass empty string for external delivery to leave it blank)

# 2) For an existing draft/client docx: add the front matter (cover/signature/TOC fields) → see scripts/face_docx.py --help
python3 scripts/face_docx.py input.docx output.docx "Full Project Name" "" "KY-2026-XXX"

# 3) Polish the document (AI trace removal + style reconstruction + table beautification) ⭐ Most commonly used
python3 scripts/docx_polish.py input.docx output.docx "Full Project Name" --title "Report Title"

# 4) Calculation model (formulas pre-built, blue cells = inputs) ⭐ Professional version v2 (three-line tables + five-color semantics + cover instructions)
python3 scripts/gen_model_xlsx_v2.py output.xlsx "Project Name" total_investment_ten_thousand_revenue_per_year_reaching_capacity construction_period operational_period
#    Example: python3 scripts/gen_model_xlsx_v2.py model.xlsx "XX Project" 3000000 600000 4 16

# 5) Preparation notes and data request list (1-2 pages, one of the three delivery pieces)
python3 scripts/gen_shuoming.py output.docx "Project Name" "Approximately 30 billion yuan"

4.1 Hard Rules for Pagination and Images (Word Rendering ≠ LibreOffice Rendering — Lessons Learned the Hard Way)

Core Insight: Using LibreOffice to convert to PDF for layout self-check ≠ the effect in Word/WPS. 90% of clients use Word/WPS. Issues like pagination and cropped images only appear in the Word engine. You must first check the XML before drawing conclusions.

Symptom	Root Cause	Fix
Image shows only a thin strip or a corner	The Normal style uses lineRule="exact" fixed line height, which crops inline images (Word-specific; LO does not crop)	Change the body text style line height to atLeast (preserves visual line height but allows images to expand); for paragraphs containing images, explicitly set spacing line=240 lineRule=auto
Only one line of text on a page / large blank area	Every row in a table has cantSplit (cannot break across pages) → if a large table doesn't fit, the entire table jumps to the next page	Only lock header rows with cantSplit; allow all data rows to break across pages; add tblHeader to header rows (automatically repeats header on the next page)
Table causes abnormal Word pagination	Total table width exceeds the text area (A4 text area = 8730 twips, 3cm left margin)	Proportionally scale gridCol values to ≤ 8700 twips
Orphaned line at the bottom of a page (only 2 lines of a sentence split)	A paragraph breaks across a page with only 2 lines remaining at the bottom	Add keepLines to that paragraph (keeps the paragraph together on one page); move the entire paragraph to the next page
Orphaned line on the last page ("(End of Report)" alone on a page)	After a large appendix table, the closing paragraph can't fit on the previous page	Directly delete the closing mark or compress the preceding table
Cover page "Preparing Unit: ______" line	Underline space placeholder run (<w:u> + space)	When the unit name is filled in, delete the underline space run; do not leave a blank placeholder line

After fixing, re-render and re-check: Use fitz to scan every page for "almost empty pages" (max_y < 350) and confirm image widths > 200pt are complete.

4.2 PDF Delivery (For Investor Reading Formats)

• Use soffice --headless --convert-to pdf (LibreOffice). Heading outlines automatically become bookmarks (766 = professional level).
• Page number fields / TOC fields rendering as placeholders is normal (they update only after pressing F9 in Word) — PDF rendering via LO is unaffected; for the docx, remind the client to press F9 in Word.
• For large docx files (>10 MB with multiple images), preview in mobile/IM apps may not show images → attach a PDF version when delivering + remind the client to open the docx in Word.

Step 5: Report QA (Automated Health Check for Numbers and Terminology)

python3 scripts/report_qa.py stats report.docx                 # Scale check: page count / paragraph count / character count / embedded images
python3 scripts/report_qa.py grep  report.docx "old terminology|old number"    # Regex search (includes table cells; skipping table cells = skipping half the numbers)
python3 scripts/report_qa.py check report.docx baseline.json    # Assertion check: banned words must have zero hits; required items must appear
python3 scripts/report_qa.py cache report.docx                  # Build plain text cache; repeated checks during revision are ~14x faster

• Before delivering a revised version, write each client request as an assertion (banned words / required items) and run a single command to prove all pass — more reliable than scanning three times by eye.
• If the assertion set is empty, an explicit warning will be issued to avoid the false positive of "0/0 all green."

Step 6: Visual Inspection + Delivery (Three-Piece Set + Dual Output)

Visual Inspection (The Final Gate Before Delivery) : Issues in AI-generated layouts cannot be caught by textual inspection; they must be rendered as images for visual models to "see" them. A single model might be perfunctory or make errors → use 2+ models for cross-validation. When issues are found, use coordinate data to double-check before fixing.

# Choose one of the following environment variables (recommended order: free → cheap & fast → backup)
export ZHIPU_API_KEY=...             # ① GLM-4V-Flash (free, ~1.6s/page) — preferred
export DASHSCOPE_API_KEY=sk-...      # ② qwen3-vl-flash (Bailian Qwen Visual Flash, cheap & fast)
export MIMO_API_KEY=...              # ③ Xiaomi Mimo 2.5 (native multimodal)
export DEEPSEEK_API_KEY=...          # ④ DeepSeek Vision (backup — DeepSeek is mainly used for writing; minimize vision usage)

python3 scripts/vision_check.py report.docx    # docx → render key pages → multi-model review (auto-fallback: ① → ② → ③ → ④)

Three Steps for Delivery

1. Deliver the three-piece set: "report.docx" + "calculation model.xlsx" + "preparation notes.docx". If the client needs a PDF, add it.
2. When delivering, include two short instructions: (a) "Open the docx in Word and press F9 to refresh the table of contents and cross-references." (b) "For calculation model.xlsx, only fill in the blue cells; the formulas are fixed — do not modify them." These two sentences avoid 80% of client follow-up questions.
3. Ask for one sentence of feedback at the end: "Is this version ready to proceed? Any part needs revision?" → Extends the conversation and reduces friction. Do not ask the client "Are you satisfied?"

IV. Revision and Terminology Alignment (Handling Client Markups / Number Updates)

Revision Reception Process (Don't Start Making Changes as Soon as You Receive a Client Document)

The most common mistake in revision: immediately finding "the place where the number changed" and making a local fix → introducing a terminology conflict → getting hit back. Calibrate the alignment first.

1. Differentiate the revision type:
   • Type A (Simple text polish): Typos, phrasing, flow tuning
   • Type B (Number changes, e.g., total investment changed from 10 billion to 8 billion, or revenue phasing adjusted)
   • Type C (Structural overhaul): Client wants to change the project positioning, logic argument, or industry classification → recalculate the impact scope immediately
2. Align terminology first (especially important for Type B/C!):
   • If the investment structure changes, confirm whether the new numbers use the same calculation bases (e.g., "construction area" from building permit area vs. red-line land area → two different things)
   • If the timeline changes, clarify: "preparatory phase" vs. "construction phase" vs. "trial operation phase" — what does each include?
   • If the financing solution changes, clarify: equity vs. debt ratio, what counts as the company's own investment (land conversion? existing equipment evaluation?)
3. Scope the ripple effect: Use report_qa.py cache to build a cache, then search the full text for old terminology/old numbers to locate all affected positions. Do not rely on memory.

Terminology Alignment Output (Always Send a Confirmation Before Revision)

After discussing with the client, produce a terminology alignment table:

Terminology	Original Version Value	Revised Version Value	Affected Chapters (Scope)
Total Investment	10 billion yuan	8 billion yuan	Overview + Investment Estimation + Financial Evaluation + Conclusions
Construction Period	3 years	4 years	All chapters
Main Product Output	500,000 tons/year	350,000 tons/year	Market + Technology + Financial

Always send this table to the client for confirmation before modifying a single word. This avoids "the number was changed but the premise wasn't updated" causing the client to send revision instructions three times.

Revision Workflow

# Step 1: Build cache (improves subsequent grep speed ~14x)
python3 scripts/report_qa.py cache current_draft.docx

# Step 2: Search for old terminology/number all instances (including table cells!)
python3 scripts/report_qa.py grep current_draft.docx "old total investment|original production capacity|old construction period"

# Step 3: After all replacements are made, do a regression check: search for old number → should be 0 hits
python3 scripts/report_qa.py grep current_draft.docx "old total investment"

# Step 4: Update baseline.json assertion set
# (add the new value as a required item and the old value as a banned word)
python3 scripts/report_qa.py check current_draft.docx baseline.json  # exit code 0 = passed

Typology of Client Revision Feedback and Approaches

Client Feedback	Implicit Meaning	Approach
"The numbers don't look right, check them"	The client finds the numbers illogical (usually the total doesn't match the sum of components) but hasn't identified where	Check if the total is consistent with the breakdown. If financial tables differ from the text, check the formula first
"Add this paragraph here"	The client wants to include a policy/tender requirement	Confirm the policy source (official document/link). If not found, tag it as "pending client confirmation"
"Format is not professional enough"	Softer sign that an additional polish round is needed	Re-run docx_polish, pay attention to table border thickness and header style unification
No response for 3 days	The client might be shopping around or has internal disputes; low priority	Trigger a professional follow-up: "Preparation complete; if you need to adjust any points, let us know anytime"

V. Financial Model and Four-Statement Cross-Validation (Core Numbers Must Match at a Glance)

Six Tables Must Be Interactive

Table	Core Formula Check
Investment Estimate Table	Total investment = construction costs + other costs + contingency + financing costs (during construction period)
Funding Sources and Uses	Sources total = Uses total; equity ratio ≥ 20% (if required by the lender)
Revenue Estimate Table	Revenue = price × sales volume (cannot be a lump sum without detail)
Total Cost Estimate	Cost items = each cost line item matches the investment estimate
Profit and Loss Table	Profit = revenue - cost - tax; years with negative profit → confirm no dividend distribution
Cash Flow Table	Cumulative net cash flow turns positive = static payback period; IRR vs. benchmark yield

Terminology Cross-Validation Checklist

Check Item	If It Fails, What Usually Goes Wrong
EBITDA / Operating profit / Net profit → amounts are reasonable and progressive	Costs are not broken down, or depreciation + amortization is forgotten
Loan term ≥ construction period + 2 years	Construction period changes → loan repayment schedule not updated
Payback period ≤ loan term	Client can't get a loan because the bank calculates the loan term is shorter than the payback period
Average selling price ≥ unit cost	Pricing below cost without subsidy logic → investment recovery analysis is invalid
Sensitivity analysis covers ±10% / ±20%	Only one scenario → review meeting: "What if it doesn't meet expectations?"

Sensitivity Analysis Requirements (All Five Factors)

Factor	Variation Range	Impact on IRR	Impact on NPV
Selling price	±10%, ±20%
Raw material cost	±10%, ±20%
Construction investment	±10%, ±20%
Production utilization rate	±10%, ±20%
Construction period	±1 year, ±2 years

If the "Sensitivity Analysis" section in a feasibility study report only has one or two factors (usually only investment and revenue), the review meeting will see through it in a second: the IPP (Important Participant) asks "What about raw material prices?" and the reply is "We hadn't calculated that yet."

VI. Common Pitfalls (Itemized, Precisely Labeled)

Pitfalls in Terms of Quantity and Scale

Pitfall	Manifestation	Prevention
Total ≠ sum of components	Table says total investment is 1 billion, but the sum of sub-items is only 950 million	The last row of the investment estimate table must use SUM formula, not manual calculation
Area inconsistency	Text says "total construction area 50,000 sqm" but the technical chapter lists building area 60,000 sqm	Always write construction area with a formula reference; check the table-specific area when crossing chapters
Percentage doesn't add up	"Market share 30%" but the chapter analysis shows the main competitors already have 80%	Market share must be supported by capacity/industry scale calculations, not gut feel
Unit confusion	Investment estimate in "10,000 yuan" but revenue estimate in "billion yuan"	All financial tables set a unified unit; the model automatically scales

Logic-Based Pitfalls

The following logical defects are the most common categories that get caught in review meetings:

Level	Category	Manifestation
L1	No evidence to support a conclusion	"Market demand is strong" — no industry data, no growth rate
L2	The evidence does not support the conclusion	"Electric vehicles have high growth rates, so we will make hydrogen fuel cells" — the evidence is at the industry level, but the conclusion goes to a specific technology path
L3	The evidence contradicts the conclusion	Data shows declining industry profit margins, but the conclusion is "future profitability outlook is good"
L4	Conclusions about A are derived from data on B	Feasibility of an aquaculture farm is based on the aquaculture industry's national data and ignores the data of specific farmed species

Commercial Delivery Pitfalls (Also Business Opportunities)

1. Qualification disclaimer wording: Add in the "Preparation Notes" or at the end: "This report prepares the feasibility study content; its professional opinions are for reference only. The final document requires review and seal by a qualified engineering consulting unit."
2. For projects requiring government approval (forbidden items) : Do not include: (a) planning opinions or land pre-approval documents; (b) environmental impact assessment approval; (c) energy conservation review opinions — these require official department seals, cannot be fabricated, and inserting them is illegal.
3. Cost increase for revision: The first version includes 2 rounds of revision. Additional rounds are charged. Clearly state in the contract: free for text changes; number changes that affect the model recalculate the fee.

VII. Practical Tips and Customizable Module Configuration

Category	Item	Cross-Sell Opportunity
QC Standard	AI hallucination rate ≤ 2% per 1000 words (client random spot check pass rate ≥ 95%)	QC report as an upsell item
Content	Map analysis (macro industrial map + micro logistics flow diagram)	Visual design separate fee
Content	Policy alignment table (project meets which specific articles of which planning)	Policy analysis report separate fee
Content	Project construction conditions Gantt chart / AON chart	Project management plan separate fee
Deliverable	Appendices: investment estimate breakdown book, equipment list signed by supplier	Supplier coordination fee
Delivery	Add version comparison tool (".docx version comparison" delivered along with the draft)	Aids in client review, reduces back-and-forth

Tell This to the Client When Talking Price

• What sets our team apart: We deliver at NDRC standard, not just a translation of the outline chapters. Every conclusion in the report has a source, every number has a corresponding table, and every table has a formula.
• Fast track: (a) The report is AI-finished in one go; (b) it's fully polished; (c) you sign your name. The client only has two jobs: review + confirm.
• Who can't use this: (a) "I want it for free or for 500 bucks"; (b) "Help me fabricate an approval"; (c) "We'll do it ourselves, we just need some templates." → Only sell templates in the above three categories, not custom services.

Configurable Template Module Inventory

Module Name	Used For
gov-dagang-2023.md	Government investment project outline (11 chapters)
ent-dagang-2023.md	Enterprise investment project outline (10 chapters)
data-request-list.md	Client material request list
report-plan.md	Report writing plan (three columns)
review-checklist.md	Quality review checklist (40+ items)
qa-baseline.json	Baseline assertion set for automated QA
model-template_v2.xlsx	Financial model template (customizable)
report-template.docx	Report document template (cover + body + appendices)

VIII. Engine Architecture

feasibility-report-engine/
├── README.md (this file)
├── SKILL.md (this document, standalone version)
├── references/                         # Reference materials (outlines, templates, checklists)
│   ├── dagang-2023.md                  # Original text of NDRC 2023 Outline
│   ├── gov-dagang-2023.md              # Government investment project version outline
│   ├── ent-dagang-2023.md              # Enterprise investment project version outline
│   ├── data-request-list.md            # Client data request list (three batches)
│   ├── report-plan.md                  # Report writing plan template (three columns)
│   └── review-checklist.md             # Quality review checklist (40+ items)
├── scripts/                            # Automation scripts (zero dependencies, ready to run)
│   ├── gen_report_docx.py              # Generate report skeleton docx
│   ├── docx_polish.py                  # Polish docx (remove AI traces, beautify tables, unify styles)
│   ├── gen_model_xlsx_v2.py            # Generate financial calculation model (v2 professional version)
│   ├── gen_shuoming.py                 # Generate preparation notes
│   ├── face_docx.py                    # Add front matter (cover + signature + TOC)
│   ├── report_qa.py                    # Report QA tool (stats/grep/check/cache)
│   ├── pipeline_check.py               # Pipeline self-check (source/uncertainty/terminology)
│   ├── vision_check.py                 # Visual inspection (multi-model cross-validation)
│   └── convert_pdf.sh                  # Convert docx to PDF (LibreOffice)
└── templates/                          # Editable template files
    ├── model-template_v2.xlsx          # Financial model (blue cells = inputs, formulas pre-built)
    ├── report-template.docx            # Report document template
    └── qa-baseline.json                # QA assertion set (example)

Pro Tip: After cloning this skill, directly edit templates/qa-baseline.json to set global banned words (e.g., trivial quantifiers like "many" and "enough") and must-have items for your team. Equals zero manual QA workload.

IX. Dependency Checklist (Zero Native Python Dependencies; Tools Need These)

Tool	Purpose	Installation
python3 (≥3.8)	Run all scripts	system default / brew install
LibreOffice	Convert docx to PDF (fully automatic)	apt/brew/yum install
(Optional) Word/WPS	Client-side viewing; not needed for writing	—

Why no python-docx / openpyxl / Pillow? Because we parse the docx XML directly and output valid OOXML with proper namespace handling. The financial model also directly constructs XLSX XML. This ensures no version‑compatibility surprises on the client side.

X. Version History and Roadmap

Version	Date	Core Changes
v1.0	2024-03	First publicly released skill, Word + Excel combined template
v1.1	2024-06	Added two-outline type determination logic; financial model v1
v1.2	2024-09	Optimized model matrix writing mode; added visual inspection
v1.3	2024-12	Added revision workflow; terminology alignment table template
v1.4	2025-03	Financial check + sensitivity analysis requirements; investment estimate breakdown v2
v1.5	2025-09	Added citation sanity check + regression check for terminology; quantified QA indicators
v1.6	2025-12 (this update)	New: report_qa.py + pipeline_check.py (pure Python standard library, zero dependencies ready to run); new: three quality gates + pre-delivery verification checklist; new: hard rules for pagination and images + PDF delivery; new: revision and terminology alignment chapter; new: financial check table + sensitivity analysis requirements

Planned Features (Skill Roadmap)

• Lean version: same engine, condensed to 90% common output data
• PPT version feasibility presentation skill
• PDF version feasibility report skill
• Multi-language version (EN/JP/KR, target: overseas investment projects)

---

Three-Sentence Summary for the Client:

1. We write a feasibility study report that passes review — using the NDRC 2023 outline, every chapter checked against it.
2. We deliver a Word report and an Excel financial model (formulas pre-built), so the numbers and text match.
3. You sign your name. We provide the content. No AI traces, fully polished, delivered.

Three-Sentence Summary for the User (Claude/Cursor/Codex that is running this skill) :

1. Determine the outline: government investment (11 chapters) vs. enterprise (10 chapters); getting it wrong means rewriting the entire report.
2. Parallel streams: due diligence (must search) + distributed writing (model matrix) + quality review (three gates) — do not skip any step.
3. Delivery: run scripts/ pipeline to generate docx + xlsx; then run visual inspection. After three times, you'll have the muscle memory.