zach-term-miner

zach-term-miner · Term Mining

What this skill solves

With the same model, if you write the prompt in plain language, AI can only give you what you explicitly ask for; switch to the professional terms of this field, and AI will also fill in the dimensions you didn't think of but that this term includes by default. The problem is that most people can't speak the jargon of a field they're not familiar with.

What this skill does is that preliminary step:Before doing the task, first spend 30 minutes figuring out the terms AI can understand, then use the terms to rewrite the prompt.

An example. Analyzing a batch of product reviews to find directions for product improvement: a plain-language prompt will make AI list user groups, scenarios, and pain points; add the single term JTBD (Jobs To Be Done), and AI will automatically fill in "what job users hire the product to do", "functional, emotional, and social needs", and "existing alternatives". Shorter wording, more accurate output.

Input: three sentences, at most one round of follow-up questions

The user only needs to state clearly:

1. What to do: a specific task, not a broad topic.
2. What you already know: the materials you have on hand and the concepts you already understand.
3. Which field: the industry or professional direction.

Handling missing items:

• When one or two items are missing,ask only one round of follow-up questions, at most two questions, and give 2–3 options for each question and mark the recommended one.
• If the user says "just go ahead", doesn't answer, or the conversation already has enough information, write the missing items into task.assumptions, and continue.
• Do not ask a second round of follow-up questions. The promise the user reads is "just a brief introduction is enough"; asking more than one round of follow-up questions is a breach of that promise.

Process

Step one: work backward to candidate terms

Starting from the task, list 10–15 candidates. Ask one question about each term:When AI sees this term, will it change the way it outputs? Keep the ones that will, delete the ones that won't.

Always delete: tool and software names, decorative jargon, famous frameworks unrelated to this task, and common-sense terms the user is already using.

By default, filter down to 5, at most 8. 5 corresponds to something you can finish reading in 30 minutes.

Step two: verify sources

Find a source for each term, from highest to lowest credibility:

1. Original papers or standards documents
2. Original authors, authoritative institutions, classic works
3. Official documentation
4. Reliable secondary explanations (textbooks, systematic explanations by well-known practitioners)

Write the verification result into source_status：

Status	Meaning
verified	Found a level 1–3 source, and the definition is consistent with the context of this task
source_defined	Only found a level 4 source; the definition is usable but is not considered an industry conclusion
private	The user's own wording, not an industry consensus, and must be marked as such
unverified	Could not be verified; goes only into the "to be verified" section, not into the formal terms

When there is no internet access or search tool: Label all industry terminology unverified, display a notice at the top of the HTML saying "No online verification was performed this time; please confirm the source yourself." Still produce the complete output. Do not withhold the deliverable just because verification cannot be done.

Read references/term-selection-rubric.md for the complete screening and verification rules.

Step 3: Write the before-and-after prompt comparison

This is the most important output of this skill. For each term (or group of related terms), write one comparison set:

• plain_prompt: How the user would write it in plain language based on their task.
• term_prompt: How to write it after adding terminology. It should be shorter and directly usable.
• added_dimensions: What additional dimensions the AI will fill in that the user did not mention after seeing the terminology. This column is the basis for the user to judge "whether this term is worth learning."

For rewriting rules, read references/prompt-rewrite-guide.md。

Step 4: 30-minute learning order

Assign minutes to each term (totaling about 30), and explain which to learn first and why. Principle: first learn framework words that can change the prompt structure, then learn standard words for modifying judgment.

Step 5: Generate output

1. Follow references/term-prompt-pack.schema.json to write JSON.
2. python3 scripts/validate_pack.py pack.json Validate with
3. python3 scripts/render_pack.py pack.json output.html . Render into a single-file offline HTML.
4. Open the HTML and self-check: it opens offline, does not scroll horizontally at phone width, does not break across pages when printing, and has no errors in the browser console.
5. Give only the HTML to the user, with a three-sentence explanation: how many terms, how many have been verified, and which prompt can be used directly.

Stop conditions

Stop after outputting the HTML.Do not execute the original task. If the user wants to execute it, take term_prompt and use it; that is another conversation.

Hard rules

• Do not invent terminology, sources, years, or authors. If not found, mark unverified。
• The user's own jargon can be included, but must be marked private and state "not an industry consensus."
• Separate inference from fact: definitions come from sources, and "why this affects this task" is your judgment; do not mix them together.
• Write the definition of a term as one sentence a person can understand, then give the original English name.
• Do not put Plan, requirements documents, or knowledge base accumulation in the HTML. These are not the concern of this skill.

Files

• references/term-selection-rubric.md: How to screen candidate terms and verify sources
• references/prompt-rewrite-guide.md: Plain language → terminology-version rewriting rules and examples
• references/term-prompt-pack.schema.json: Output data contract
• scripts/validate_pack.py: JSON and HTML validation
• scripts/render_pack.py: Render offline HTML
• scripts/test_pipeline.py: Regression tests (three examples)
• examples/: Rendered example HTML