SuperAPI AI Directory

nix-knowledge-protocol

General & other

View repository

Updated 10-15 21:40

About this skill

Purpose
  • Handle raw materials, time-bound thinking, and long-term reusable knowledge separately, to avoid them being silently flattened during the organization process.
  • Make knowledge objects possess three qualities at the same time: human-readable, AI-retrievable, and source-auditable.
  • Improve the long-term reuse value of knowledge work through "object typing + progressive distillation + provenance preservation," rather than only pursuing what reads smoothly in the moment.
Applicable scenarios
  • Turn chats, meeting notes, web pages, papers, and videos into notes or knowledge objects
  • Summarize, synthesize, and generalize multiple source materials
  • Update long-term reusable explanatory documentation, knowledge base pages, or concept indexes
  • Organize raw materials into tutorial, how-to, reference, explanation, overview
  • Supplement an existing knowledge base with claims, sources, relationships, open questions, or navigation structure
  • Design the structure, naming, hierarchy, and distillation strategy of knowledge objects
Inapplicable scenarios
  • The main task is implementation, debugging, refactoring, or code review
  • The main task is recommendation, prioritization, or trade-off decision-making
  • The main task is rewriting AGENTS, SKILL, prompt architecture, or routing descriptions
  • The main task is blog polishing for publishing channels, reader experience optimization, or article structure diagnosis
First determine the type of knowledge object

Do not treat all "organization" as the same kind of work. First determine which type of object the current task is closest to, then decide how to organize it:

  • Time-bound snapshot: records the judgments, confusions, plans, troubleshooting process, or interim conclusions at a certain point in time
  • source digest: extracts, compresses, and annotates a single source for easier later review
  • grounded claims: distills claims that can be examined independently from one or more sources, while preserving supporting evidence
  • reference: an index or explanation of stable concepts, people, tools, terms, interfaces, and nouns
  • explanation / synthesis: synthesizes multiple materials around a question or topic to provide a structured understanding
  • how-to / tutorial: a step-based knowledge object that helps readers complete a task
  • overview / map: provides navigation, boundaries, and entry paths for a set of notes or a topic

If the user does not specify an object type, prioritize the most conservative type that is most conducive to future reuse, and state this clearly.

Core constraints in the AI era

1. The body serves humans, the underlying layer serves AI

  • By default, treat the body as an interface for humans to read and review.
  • Provenance, claim layering, source mapping, and authorship attribution are the underlying layer; the body should not be written as a schema dump.
  • If machine-friendly information would significantly harm reading fluency, move it down into frontmatter, appendices, notes, fields, or separate blocks.

2. Preserve sources first, then synthesize

  • Distinguish: original source wording, observed facts, organized phrasing, synthesized judgments, guesses / suggestions.
  • Any synthesized conclusion should be able to answer "where does it come from."
  • For negative universal judgments in the summary layer (such as "zero citations", "never appeared", "no one has mentioned"), do not rely on them directly; first return to the original traces to do minimal sampling verification, then decide whether to continue using that judgment.
  • If the sources are insufficient to support a strong conclusion, lower the strength of the conclusion; do not use smooth wording to cover up weak evidence.

3. Temporal semantics cannot be lost

  • Temporal snapshots preserve the context of "why it was thought this way at the time"; do not rewrite them according to hindsight.
  • Store evergreen knowledge and stage-specific judgments separately or express them in separate sections.
  • If the content has a shelf life, version dependency, or time sensitivity, mark it explicitly.

4. AI assists thinking, it does not replace thinking

  • AI can help extract, compare, compress, and reorganize, but do not by default complete all cognitive processing on the user's behalf.
  • When a task essentially requires the user's stance, judgment, or value trade-offs, do not disguise AI-generated content as conclusions the user has already endorsed.
  • For highly uncertain material, prioritize preserving tensions, open questions, and candidate explanations rather than forcing them into a neat answer.

5. Normalize multimodal material first, then enter the knowledge layer

  • First convert audio, video, images, and slides into a citable text evidence layer, then synthesize.
  • During normalization, preserve traceable locators such as source type, title, time, and paragraph / page / timestamp.
  • Do not write summaries directly based on vague impressions.
Workflow

1. Clarify the request and the reuse goal

Before starting, at least clarify:

  • What knowledge object the user wants to obtain
  • Who this object is mainly for: personal review, team sharing, AI retrieval, external readers, or serving multiple parties at the same time
  • What the raw materials are, which are credible sources, and which are only drafts or casual thoughts
  • Whether the current output is for one-time consumption or will continue to be updated in the future
  • Which content must retain its original form, and which may be distilled or reorganized

2. First layer the materials

At least divide the materials into these layers:

  • raw source: original text, original words, excerpts, records
  • observed facts: factual content that has been observed
  • grounded claims: claims that can be independently examined and their basis
  • synthesis: understanding after cross-source integration
  • open questions / tensions: unresolved conflicts, gaps, uncertainties

Do not jump directly from raw source to polished summary.

3. Choose the smallest but correct distillation path

Prefer to use the following default path:

  • source -> digest -> grounded claims -> synthesis / reference / how-to

Only shorten the path when the user explicitly wants only a brief summary, or when the material itself is extremely simple.

If an old object already exists, by default update it incrementally rather than tearing it down and rewriting the whole thing. Unless the user explicitly requests overwriting.

4. Organize content by object type

Temporal snapshot

  • Preserve the time, situation, triggering reason, and the uncertainty at the time.
  • Allow immature expressions to remain; do not forcibly rewrite them into evergreen conclusions.
  • If understanding changes later, prioritize adding follow-up notes or supplementary links rather than secretly rewriting the past.

Source digest

  • Record source identity: author / title / time / form / link or source.
  • Extract key viewpoints, key facts, and the parts most relevant to your current question.
  • Mark “why it is worth recording” and “what has not yet been verified.”
  • Avoid disguising a single source’s position as objective consensus.

Grounded claims

  • Each claim should be specific enough to be challenged, cited, or reused on its own.
  • Each claim should include the minimum necessary evidence where possible: source, excerpt, page number, timestamp, related notes.
  • If multiple sources conflict with one another, preserve the conflict rather than choosing only the side that fits best.

Reference

  • Prioritize answering “what is this,” “where are its boundaries,” and “what objects is it easily confused with.”
  • Keep it compact, stable, and low in narrativity.
  • Do not write one-off personal thoughts or temporary debates directly into the main body of the reference.

Explanation / synthesis

  • Give the core question or core conclusion first, then expand on the claims that support it.
  • Make clear which parts come from sources and which parts are your organization and synthesis.
  • If the material is insufficient to reach a conclusion, output “the currently most credible understanding + questions still to be confirmed.”

How-to / tutorial

  • Make clear the preconditions, inputs, steps, expected results, and common failure points.
  • Separate steps from explanation: the main text should serve execution first, with background explanation placed later or in a separate section.
  • Do not cram reference details into the process; link to the reference object when needed.

Overview / map

  • Explain the scope, boundaries, subtopics, and recommended entry path.
  • It is a navigation layer, not a restatement of everything beneath it.
  • If the overview is already outdated and misleading, it is better to trim it to the minimum than to maintain false completeness.

5. Do a dual readability check

After completing the draft, self-check it against two sets of standards:

Is it friendly to humans

  • On first review, can one quickly know what this is, why it is worth reading, and where to start reading
  • Is the structure natural, rather than fragmented into pieces for machines
  • Are the key conclusions, key sources, and key unresolved questions visible at a glance
  • Has the necessary context been preserved, rather than leaving only dry bullet points

Is it friendly to AI

  • Is the object type clear
  • Are time, entities, concepts, relationships, and states explicit
  • Can the sources of key conclusions be traced
  • Are stable knowledge and temporary judgments mixed together
  • Are boundaries left for continued expansion, linking, and citation

6. Clarify authorship attribution and governance boundaries

  • When the output mixes the user’s original text, source excerpts, and AI synthesis, the boundaries should be identifiable.
  • Unless requested, do not overwrite the user’s original judgment with the AI’s voice.
  • For rewrites, merges, or deletions that may change the original meaning, preserve a traceback path or provide an explicit explanation by default.
Common output shapes

Organize a single source

At minimum include:

  • Source identity
  • Key content
  • Relationship to the current topic
  • Unverified points or limitations

Multi-source synthesis

At minimum include:

  • The question to answer
  • Key claims
  • The main basis for each claim
  • Conflicts and uncertainties
  • Current synthesized conclusion or provisional conclusion

Update a knowledge base object

At minimum include:

  • What object is being updated
  • What information has been added
  • Which knowledge is stable, and which is phase-specific information
  • Whether to retain the old wording or link to historical versions

Design the knowledge structure

Must include at least:

  • Division of object types
  • Boundaries of each object type
  • Distillation path
  • How provenance is retained
  • How the human-readable interface and the AI substrate are layered
Output requirements
  • Prioritize delivering directly usable knowledge objects rather than general suggestions.
  • For content that is source-sensitive or highly factual, explicitly report verification status.
  • When citation or attribution is needed, prioritize traceable sources rather than "according to some materials."
  • When you synthesize, restate, or compress, make it clear to readers that this is an edited formulation.
  • If the user requires a strict format, follow that format; this protocol only defines the quality threshold for knowledge work and does not override repository-local formatting standards.
Completion criteria
  • An appropriate knowledge object type has been selected, rather than mixing different tasks into one kind of "organization."
  • The output simultaneously supports human review experience, AI retrievability, and source auditability.
  • Temporal semantics, authorship attribution, and key uncertainties have not been silently smoothed away.
  • Synthesized conclusions can be traced to sufficient evidence, or it is clearly stated that the evidence is insufficient.
  • The result has clear boundaries for future continued updates, linking, or reuse.
Anti-patterns
  • Flattening time-bound snapshots into evergreen knowledge
  • Rewriting raw notes directly into a polished summary, losing the traceback path
  • Mixing source wording, facts, synthesized judgments, and recommendations into one layer
  • Writing the main text as a hard-to-read database export for AI retrieval
  • Deleting key provenance for the sake of readability
  • Disguising a single source's position as consensus
  • Forcing all objects into claim lists, damaging natural reading
  • Outputting overly strong, overly stable synthesized conclusions when there is insufficient evidence
  • Silently overwriting the user's original text or original position without stating that this is an AI rewrite