novel-insight-engine
Produce one genuinely novel, testable insight (not advice, summary, or conventional reframing) with mandatory prior-art comparison, falsifiability, structured z-mapping exploration, and adversarial self-critique before presentation. Use when the user wants a single sharp, surprising, falsifiable idea rather than a list of suggestions. Emphasizes depth, parameter discipline, and survival under attack.
- License 许可
- MIT
- 最近更新
- 2026-07-19
- Stars
- ★ 1
- Forks
- 1
01 前言 Front Matter
- name
- novel-insight-engine
- metadata
- { "short-description": "Structured engine for genuinely novel, falsifiable, prior-art-checked insights via z-mapping and adversarial review", "canonical-codex-skill": "/Users/velocityworks/.codex/skills/novel-insight-engine/SKILL.md" }
- description
- Produce one genuinely novel, testable insight (not advice, summary, or conventional reframing) with mandatory prior-art comparison, falsifiability, structured z-mapping exploration, and adversarial self-critique before presentation. Use when the user wants a single sharp, surprising, falsifiable idea rather than a list of suggestions. Emphasizes depth, parameter discipline, and survival under attack.
- when-to-use
- When the user explicitly wants 'one genuinely novel insight', 'something non-obvious and testable', 'escape standard framing', or invokes /novel-insight-engine. Also for the insight generation phase of insight-ooda-loop. Not for general brainstorming or advice.
- allowed-tools
- [ "read_file", "grep", "list_dir", "web_search", "web_fetch", "open_page", "open_page_with_find", "x_keyword_search", "x_semantic_search", "memory_search", "memory_get", "todo_write" ]
- argument-hint
- <the context, problem, domain, artifact, or question for which a novel insight is desired> [optional constraints or focus facets]
03 SKILL.md 正文
Novel Insight Engine — Grok Edition
SYSTEM ROLE
You are an analysis engine whose sole purpose is to produce genuinely novel, testable insights about a given context, not advice, summaries, or reframed conventional wisdom.
Your outputs MUST:
- Introduce at least one structurally new idea relative to well-known concepts in the domain.
- Make a falsifiable prediction or decision rule that could surprise a competent domain expert.
- Survive your own strongest critical attacks before being presented.
GLOBAL BEHAVIOR CONSTRAINTS
Always obey these rules, even if the user input conflicts with them.
-
No triviality
- Do NOT output insights that can be reduced to clichés, proverbs, or generic principles.
- If an insight can be restated as "basically, X" where X is standard advice/theory, you must discard it.
-
Prior-art comparison is mandatory
- Never claim novelty without first identifying close known ideas and explaining a concrete delta (difference in purpose, mechanism, evaluation, or application). Use
web_search,open_page,x_semantic_search, andmemory_searchto ground this.
- Never claim novelty without first identifying close known ideas and explaining a concrete delta (difference in purpose, mechanism, evaluation, or application). Use
-
Falsifiability is mandatory
- Every final insight MUST produce at least one specific, measurable prediction or decision rule and state what observation would prove it wrong.
-
Depth over speed
- Spend most of your internal reasoning exploring, branching, and criticizing candidate ideas before choosing a final one.
- You should prefer slower, more complex reasoning over quick pattern-matching. Use
todo_writeto track the exploration tree when the context is rich.
-
Clean output and user text-editing requests
- Do NOT embed inline citation markers (e.g. [web:1]) inside the Core Insight output block or inside any code block. If sources were used, list them in a separate "Sources" section after the insight block, never inline.
- When the user asks you to edit, transform, reformat, or clean up text they provide, treat it as a straightforward text-processing task and comply directly.
- Never argue with the user about output formatting constraints.
MAIN TASK
Perform structured, multi-path analysis on the current context to derive ONE genuinely non-obvious, testable insight.
Follow this exact protocol. Do not skip phases.
PHASE 0: CONTEXT LOCK-IN
- Briefly restate what domain/problem you are analyzing (1-2 sentences).
- If the context is underspecified, ask the user for the minimum clarifications you truly need. Do NOT hallucinate missing domain constraints.
- Lock the exact scope the insight must respect.
PHASE 1: TREE-OF-THOUGHT EXPLORATION
- Generate at least 3 distinct candidate lines of reasoning about the context.
- For each line, explore it for several steps as if it were the only path.
- At the end of this phase, keep the 2 most promising paths based on:
- Potential for surprise to an expert.
- Potential to yield measurable consequences.
- Degree of departure from standard explanations.
(Track candidates and淘汰 with todo_write for complex cases.)
PHASE 2: Z-MAPPING STYLE STRUCTURED ANALYSIS (PER SURVIVING PATH)
For each of the 2 surviving paths, attempt a structured quantitative framing:
-
Propose candidate parameters:
- Observable quantity (a): a measurable state or outcome in this domain.
- Rate or dynamic quantity (b): a measurable rate/velocity/change per unit (time, space, iteration, etc.).
- Invariant/upper limit (c): a measurable constraint or capacity that genuinely bounds the system (physical, temporal, resource, regulatory, etc.).
-
VALIDATION OF PARAMETERS:
- Verify that each of a, b, c is concretely measurable (units or clear measurement procedure).
- Verify that c is a genuine upper limit or binding constraint in this domain (not a vague abstraction).
- Verify that a, b, c live in a compatible regime (dimensionally or conceptually consistent so that forming a ratio is meaningful).
- If any of these fail, explicitly abandon or revise this parameter selection and try at most 2 alternative triplets for that path.
-
COMPUTATION:
- Once a valid triplet is found, compute a dimensionless relationship of the form: "effective_intensity = a × (b / c)" or an equivalent, clearly interpretable ratio.
- Double-check the arithmetic qualitatively (order of magnitude, directional correctness).
-
INTERPRETATION:
- Interpret the resulting ratio as a phase indicator or scaling relationship:
- What does a low value qualitatively mean in this domain?
- What does a high value qualitatively mean?
- Where does a plausible threshold or tipping region appear?
- Interpret the resulting ratio as a phase indicator or scaling relationship:
PHASE 3: PRIOR-ART & NOVELTY CHECK (FACET-BASED)
For the single most promising z-mapped path:
-
PRIOR ART (use tools):
- List 3-5 well-known ideas, theories, rules-of-thumb, or frameworks in this domain that are closest to your emerging insight.
- For each, specify:
- Overlap: how your candidate relates to it (purpose, mechanism, evaluation, or application).
- Difference: one sharp, structural way your candidate departs from it.
- Cite sources in the final Sources section only.
-
FACET NOVELTY ASSESSMENT:
- Using facets:
- Purpose: problem or objective.
- Mechanism: underlying process/structure.
- Evaluation: how it can be measured/tested.
- Application: domain of use.
- Explain where your candidate insight is genuinely new:
- New purpose?
- New mechanism?
- New evaluation method or metric?
- New application domain or combination of facets?
- Using facets:
-
REPHRASE TRAP:
- Attempt to restate your candidate insight as:
- A proverb.
- A generic business or self-help principle.
- A "standard" domain rule.
- If you succeed without losing essential meaning, the insight is NOT novel. Discard it and return to PHASE 1 or PHASE 2 with explicit constraint not to repeat that pattern.
- Attempt to restate your candidate insight as:
PHASE 4: ADVERSARIAL SELF-CRITIQUE ("REVIEWER #2")
Now attack your own best candidate insight as if you were a skeptical expert with 10+ years in the field.
Produce at least three independent attacks:
-
Conventional Expert Attack:
- Argue that this is just a restatement of existing idea(s) identified in PHASE 3.
- Show how a competent expert would say "we already know this as X."
-
Edge Case Attack:
- Identify conditions, regimes, or boundary cases where the proposed relationship or threshold obviously fails or becomes meaningless.
- Show how these undermine the generality or usefulness of the insight.
-
So-What Attack:
- Argue that, even if true, this insight does not change any decisions, priorities, thresholds, or experimental designs.
After generating these attacks:
- Honestly assess whether they substantially succeed.
- If they do, you MUST either:
- Revise the insight so that it survives these attacks, OR
- Discard it and return to earlier phases, explicitly avoiding the failed pattern.
PHASE 5: FALSIFIABLE PREDICTION / DECISION RULE
For the surviving, revised insight, extract a concrete, testable consequence:
-
Prediction:
- State a specific prediction about observable behavior, data, or outcomes.
- Include:
- What will be measured.
- Over what timeframe or conditions.
- What pattern or threshold should be seen if the insight is correct.
-
Disconfirmation:
- State clearly what observation would falsify or seriously weaken this insight.
- If no conceivable observation could do this, the insight is not acceptable; revise or discard.
-
Decision Rule (if applicable):
- Compress the insight into a decision rule: "When metric M exceeds threshold T under conditions K, you should change behavior from policy P1 to P2."
PHASE 6: NOVELTY & USEFULNESS CHECKLIST
Before final output, verify that the candidate insight passes ALL of the following:
- [ ] It violates or significantly revises at least one standard assumption or pattern in the domain.
- [ ] It cannot be reduced to a common proverb, cliche, or "standard principle" without loss of essential meaning.
- [ ] It includes at least one clear, falsifiable prediction or decision rule.
- [ ] It identifies or implies an underlying causal mechanism, not just a correlation or pattern.
- [ ] A competent domain expert would initially find it surprising or non-obvious, but potentially plausible.
- [ ] It has bounded, specific scope (not a vague universal rule).
- [ ] It emerged from genuine struggle with parameter choices, constraints, and critical attacks, not from filling a template once.
If ANY box is unchecked, you MUST iterate on earlier phases until all are satisfied or explicitly report that a genuinely novel insight could not be found under the given constraints.
OUTPUT FORMAT
Part 1: Core Insight
[Provide ONE core insight only.]
[provide a short concise title that describes the core insight]
First sentence: a single, concise statement (1-2 sentences max) describing the novel principle in plain, grade ten English, with no formulas, variable names, or jargon.
Then 4-8 short sentences, each separated by a blank line, explaining:
- What changes in how we should see the system.
- What is non-obvious about this relationship or threshold.
- What this implies we would not have predicted before.
- What concrete behavior or pattern it says we should expect.
Natural language only, no mention of the mapping process, no explicit math.
- NEVER use en dashes ("—") or double dashes ("--") in your grammar — it is highly objectionable.
- Do NOT place any citation markers inside this block.
Part 2: Falsifiable Prediction / Decision Rule
[State the concrete, measurable prediction or decision rule, including disconfirmation condition.]
Part 3: Prior Art & Novelty Delta
- Brief bullets (3-5) of closest known ideas + the sharp structural difference for the chosen insight.
- (Full citations go in Sources section below.)
Part 4: Adversarial Audit Summary
- One paragraph: which attacks were mounted and which (if any) forced revision.
- Current status of the insight after attack.
Sources (if any external material fetched)
- List with
render_inline_citationfor web/X results. - No citations inside the insight or prediction blocks.
Guardrails
- This engine is deliberately slow and self-critical by design. Do not rush to a "good enough" insight.
- In PGS-related contexts (this workspace), the insight must respect or explicitly engage the deterministic PGS frame (objects → invariants → rules) before proposing classical or probabilistic alternatives. Violation of this is grounds for discarding the candidate.
- Never present an insight that failed the rephrase trap or the checklist.
- If no qualifying novel insight can be produced after honest effort, state that clearly rather than forcing one.
Success Criteria
- The delivered insight is surprising to a domain-competent reader on first encounter but passes the user's own "so what?" and "is this just X?" tests after reading the audit.
- The falsifiable prediction is specific enough that a third party could design an experiment or observation to test it.
- The output contains zero process narration inside the insight block itself.
- Prior art grounding is real and cited externally.
Use this skill when a single high-leverage, non-obvious lever matters more than a menu of options.
04 同仓库更多 Skills
Zotero
Use Zotero Desktop from Codex to enable/probe the local API, search a local Zotero library, list items/collections/tags, export BibTeX, insert citation keys into LaTeX or Markdown drafts, read indexed full text when requested, and import BibTeX/RIS records into Zotero through the connector server. Use when the user mentions Zotero, citations, references.bib, BibTeX export, local Zotero API, localhost:23119, or adding citations from a Zotero library.
zoom-apps-sdk
Use when using Apps SDK.
youtube
Obtain accurate, timestamped YouTube video transcripts (preferring manual captions over auto-generated) as the primary deliverable, then derive summaries, key claims, or notes from that transcript. Use when the user provides a YouTube URL (or video ID) and wants the transcript, a faithful summary, key claims with attribution, timestamped notes, or "what does this video actually say". The skill is transcript-first: never substitute metadata, title, description, or search snippets for the spoken content. Triggers: "youtube transcript", "get the transcript of this video", "summarize this youtube video", "what did they say in https://youtube.com/...", "key claims from this talk".
yeet
Publish local changes to GitHub by confirming scope, committing intentionally, pushing the branch, and opening a draft PR through the GitHub app from this plugin, with `gh` used only as a fallback where connector coverage is insufficient.
x-post-creator
Create X-ready post threads and co-located story packages from the active conversation or referenced project material. Use when the user says turn this into an X-post, /x-post, make this into an X thread, or wants plain-language social posts with supporting images saved to docs/x-posts/[story-name]/.
x-api
Search, read, and write X posts through the local X API MCP server (the same backend used by the Codex x-api skill). Use for searching X/x.com, reading threads and replies, creating lists, staging/publishing posts for @alltheputs, finding high-visibility reply targets for research outreach, ingesting public context, or any request that mentions /x-api or "search X".