All Prompts and the Full Pipeline — A Complete Kit for Starting in Your Own Field

This is the final post of the series.

I've collapsed everything from posts 1-5 into a template kit you can reproduce directly: prompts, tool stack, pipeline, checklists. All of it in this one article.

If, by the end of reading, you feel "let me have Claude read a forgotten long-form document from my own field today," the series has done its job.

Full Pipeline

┌──────────────────────────────────────────────┐
│ STEP 1: Discovery                             │
│  - Use WebSearch to narrow the field          │
│  - Trusted indexes (Google Patents/Wikipedia) │
│  - Build a candidate list of 5-10             │
└────────────────┬───────────────────────────┘
                 ↓
┌──────────────────────────────────────────────┐
│ STEP 2: Filtering                             │
│  - Candidate-narrowing prompt                 │
│  - Down to one                                │
└────────────────┬───────────────────────────┘
                 ↓
┌──────────────────────────────────────────────┐
│ STEP 3: Extraction                            │
│  - WebFetch the full text                     │
│  - Extraction prompt → structured info        │
└────────────────┬───────────────────────────┘
                 ↓
┌──────────────────────────────────────────────┐
│ STEP 4: Modern Translation                    │
│  - Modern-translation prompt → table          │
│  - Past ⇔ present correspondence              │
│  - The most powerful prompt in the series     │
└────────────────┬───────────────────────────┘
                 ↓
┌──────────────────────────────────────────────┐
│ STEP 5: Grading                               │
│  - Grading prompt → "right / wrong / neutral" │
│  - Re-evaluate the past against present facts │
└────────────────┬───────────────────────────┘
                 ↓
┌──────────────────────────────────────────────┐
│ STEP 6: Pitfall Check                         │
│  - Anti-fabrication prompt                    │
│  - Context-forcing prompt                     │
│  - Translation-consistency prompt             │
└────────────────┬───────────────────────────┘
                 ↓
┌──────────────────────────────────────────────┐
│ STEP 7: Publish                               │
│  - Primary sources required                   │
│  - No position-talking                        │
│  - Full prompt disclosure                     │
│  - Failures included                          │
└──────────────────────────────────────────────┘

Below is every prompt for every step.

STEP 2: Candidate-Narrowing Prompt

Purpose: pick one out of 5-10 candidates.

For the following [N] [genre] candidates, pick one based on the criteria
below and give three reasons.

Selection criteria:
1. High structural similarity to modern [modern technology]
2. Confirmed expired/retired so it can be freely excavated
3. Off the contemporary mainstream — dropped out of the industry's
   collective memory

Candidates:
[candidate 1 summary]
[candidate 2 summary]
...

In post #2 (Patent Archaeology #1), this picked ZISC.

STEP 3: Extraction Prompts

Purpose: pull structured information out of the primary source.

For patents (post #2 ZISC)

Extract the following from this patent:
1. Patent number, grant date, filing date, inventors, assignee
2. Status (Expired or not) and expiration date
3. Abstract
4. Main Claim 1 (independent claim 1)
5. The problem it solves
6. The proposed solution mechanism
7. Application domains and industries
8. Cited prior art
9. Forward citation count
10. Description of key Figures, and which one best represents the
    mechanism

For standards (post #4 Token Ring)

For [standard name], extract:
1. Year of standardization, year retired/inactive
2. Key inventors and driving companies
3. Core mechanism (the standard's unique key concept)
4. Why it lost in the market
5. Relationship to modern [related technology]
6. Whether it is being re-evaluated for AI workloads / HPC
7. Spec size and length

For government documents (post #5 ALPAC)

For [government document name], extract:
1. Full title, year, publisher
2. Why the report was commissioned
3. Committee members (key people)
4. State of [field] research at the time
5. Main conclusions (recommendations, listed)
6. Policy impact of the report
7. Relationship to the [field] winter
8. Later evaluation
9. Length and availability of the report
10. Whether the report's calls have aged well, viewed against the modern
    [equivalent technology]

For corporate IR (post #3 Samsung)

From [company name]'s history, especially [decade] [business area]
development, extract:
1. Year of [business] entry, first major product
2. Major milestones in [decade]
3. Response to crisis (bubble crash, financial crisis)
4. Key strategic decisions
5. Year of entry into [later business]
6. When the relationship with major customers (Apple, etc.) began
7. Pivot timing toward [present main business]
8. Response to AI / new technology
9. Relationships with competitors
10. Generational succession of CEOs and senior leadership

STEP 4: Modern-Translation Prompt (the most powerful prompt in the series)

Purpose: render past long-form material into a present-day correspondence table.

Translate the technical mechanism (or key concept) of [past document]
into the everyday vocabulary of a [field] researcher in 2026. Show, in
a table, which element corresponds to which concept in modern papers.

This single prompt produced:

• Post #2 ZISC: Manhattan distance ⇔ L1 distance, daisy chain ⇔ systolic array
• Post #4 Token Ring: control token ⇔ credit-based flow control, ring topology ⇔ Fat Tree
• Post #5 ALPAC: "didn't reach production quality" ⇔ pre-Transformer reality, "humans superior" ⇔ correct until the early 2000s

The past ⇔ present correspondence table falls out in one shot. This is the strongest weapon in the series.

STEP 5: Grading Prompt

Purpose: evaluate past claims against present facts.

Take the N main recommendations (or claims) made in [past document] in
[year], and grade each against the reality of modern [present
technology] ([specific present-day technology and date]). Categorize
each as "right," "wrong," or "neutral," and give the basis for the
verdict in 1-2 sentences.

In post #5 ALPAC, this graded the five recommendations as "3 right / 3 wrong." This is the section readers find most valuable.

STEP 6: Pitfall-Check Prompts

Anti-fabrication prompt (mandatory pre-publication)

For the following article draft, list every cited outlet name, person
name, organization name, number, and quote. For each one, classify as:

(A) Primary source verifiable
(B) Confirmed only through secondary citation
(C) Cannot be confirmed (= suspected fabrication)

Items in category C are deletion candidates from the article.

Context-forcing prompt (anti-misreading)

For the following terms, interpret each in the industry context of the
document's publication year (YYYY). If the contemporary meaning differs
from the historical meaning, give both.

[term list]

Translation-consistency prompt (anti-mistranslation)

Below are the original X and translation Y. Check whether the major
numbers, proper nouns, and numerical expressions in Y match X. Report
every discrepancy.

Tool Stack

Every tool used in the series:

Sources WebFetch cannot reach (Claude Code environment constraint):

• SEC EDGAR (403)
• TSMC IR (403)
• Samsung 1990s IR (does not exist)
• Wayback Machine (fetch refused)

These need a different route (direct browser, Bash + curl, or API key). For real ongoing use, a Python script running on your own machine (Mac mini, etc.) is the most reliable.

Checklist for Starting in Your Own Field

□ Pick one specialist field (or area of strong interest)
  Example: FX / medicine / law / semiconductors / education / music /
  cooking...

□ Write down the "long-form material in that field that humans don't
  read but is valuable"
  Example, FX:       central bank statements, IMF reports, 20 years of
                     FOMC minutes...
  Example, medicine: discontinued treatment protocols, retracted papers,
                     out-of-print textbooks...
  Example, law:      old precedents, repealed ordinances, transcripts...

□ Identify web-accessible primary sources
  Example: FOMC minutes → federalreserve.gov (free, HTML)
  Example: old medical papers → PubMed (free) / NLM historical archive

□ Run STEPS 1-3 once for real (discovery → filtering → extraction)

□ Use STEP 4 (modern-translation prompt) to draw out the past ⇔ present
  correspondence table

□ Use STEP 5 (grading prompt) to evaluate the past claims

□ Run all of STEP 6 (pitfall checks)

□ Write the post: primary sources required, no position-talking

□ Publish (personal blog / Substack / X long-form post / dedicated LP)

□ Watch the response, then think about sub-series naming (mine became
  Patent / IR / Standard / Declassified Archaeology)

Run this for one month in one field, and you become the AI archaeologist of that field. Nobody in the world has claimed that title yet. The window for first-mover advantage is open right now.

Series Wrap-Up

Across posts 1-7, the things I wanted to convey:

1. LLM-mediated arbitrage is not just for Amazon (post #1, Gipp case)
2. A 30-year-old patent holds the ancestor of the modern NPU (post #2, ZISC)
3. Companies forget their own greatest achievements; IR is behind walls (post #3, Samsung 1996)
4. Discarded standards weren't "wrong" — they were "30 years too early" (post #4, Token Ring)
5. A single government document can stop a research field for 20 years (post #5, ALPAC)
6. Avoid three pitfalls (fabrication, cost explosion, misreading) and your incident rate drops by orders of magnitude (post #6, Pitfalls)
7. With this prompt set and pipeline, anyone can start today (post #7, this post)

The single theme underneath all of them:

"Humanity has produced enormous volumes of long-form material that humans never read. LLMs can now read it. The first mover takes the territory."

That's the entire bet.

Where This Goes From Here

The series completed its "introduction set" at post #7, but the act of mining forgotten long-form documents continues from here, indefinitely.

My (haruko's) plan:

• Patent Archaeology #2, #3, #4...: dig one expired patent per month
• IR Archaeology #2, #3...: try to break the SEC EDGAR wall via alternate routes
• Standard Archaeology #2: re-evaluate CORBA, WAP, HTTP/1.0
• Declassified Archaeology #2: 1973 UK Lighthill report — the British AI winter
• Possible new sub-series: Bankruptcy Archaeology (final filings of failed companies) / Court Archaeology (old precedents) / Thesis Archaeology (buried doctoral dissertations)

Pace: at minimum four posts per month. This is the real starting line of the series.

Every prompt, every pipeline, every checklist used in the series is collapsed into this post. If you dig something up in your own field, please tell me about it. I am genuinely looking forward to reading those archaeology logs.

References (every tool used in the series):

• Google Patents
• USPTO Open Data Portal
• USPTO Developer Hub
• PatentsView
• IETF RFC Editor
• National Academies Press
• DTIC (Defense Technical Information Center)
• microsoft/markitdown
• simonw/files-to-prompt
• anthropics/claude-code-skills

Series links:

• Post 1 — Introduction
• Post 2 — Patent Archaeology #1
• Post 3 — IR Archaeology #1
• Post 4 — Standard Archaeology #1
• Post 5 — Declassified Archaeology #1
• Post 6 — Pitfalls
• Post 7 — Templates (this post)

→ Read the original Japanese version at haruko's blog

Author: はる子 / @haruko_ai_jp — a non-engineer running 7 web apps with Claude Code and 4 AI assistants in Tokyo.