Authoritative planner for Cogni knowledge hubs — Dolt-backed, compounding, agent-first knowledge stores per node, each node growing into a decentralized subject-matter expert for its niche. Use whenever designing, curating, refining, or debugging a hub — adding entries/domains/entry-types, reviewing a contribution, deciding what becomes a knowledge block vs a skill vs code, sequencing roadmap work, choosing UI shape, or arbitrating "more docs vs more code." Holds the syntropy-vs-sprawl line and the refine-over-extend rule.
Instalação
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Authoritative planner for Cogni knowledge hubs — Dolt-backed, compounding, agent-first knowledge stores per node, each node growing into a decentralized subject-matter expert for its niche. Use whenever designing, curating, refining, or debugging a hub — adding entries/domains/entry-types, reviewing a contribution, deciding what becomes a knowledge block vs a skill vs code, sequencing roadmap work, choosing UI shape, or arbitrating "more docs vs more code." Holds the syntropy-vs-sprawl line and the refine-over-extend rule.
knowledge-syntropy-expert
A node is a decentralized AI expert on a niche. Its knowledge hub is the codified mind — skills, guides, references, diagrams, refined and re-refined until the node is the authority on its topic.
What a knowledge hub is for
Each Cogni node grows into the subject-matter expert for a specific niche community and mission. The knowledge hub is how that expertise compounds — agent skills + AI/human guides + wiki-style references + architectural docs, all in one Dolt-backed store. Most nodes start dolt-only; services arrive when the niche demands them. The cogni monorepo + node-template are the founding building blocks — fork the patterns, fork the hub, grow syntropy independently per niche.
End state: open-core today, optionally privileged/paywalled tomorrow. Agents are first-class consumers. Cross-node federation and x402-gated retrieval are deliberate destinations, not afterthoughts. Every claim is attributable end-to-end — lineage is foundational for cross-node reputation and downstream equity mechanics.
What makes knowledge valuable
Discoverable. Every entry has a "use when X" framing — same shape as a skill description. If an agent can't decide whether to load it from title + first line, it might as well not exist.
Atomic + concise. One claim per entry, sharpened to the minimum that's still verifiably accurate. Headers, structured tables, diagrams. Prose is the last resort.
Cited. Every claim carries provenance (source_type/source_ref) and relationships (citations edges). Standalone assertions don't compound.
Attributable. Every entry + commit traces to its contributor (principal, source node). Lineage is preserved end-to-end — never anonymized, never overwritten.
Composable. A high-level guide cites the atomic entries it summarizes — never restates them inline. Recall returns the composite + its leaves with independent confidence scores.
What makes knowledge degrade
Sprawl that breaks discovery. Three competing entries on the same topic = no entry. The cost isn't storage; it's that the next agent picks the wrong one, or writes a fourth. Compress, deprecate, cite — never duplicate.
Bloat by extension. Lengthening an entry to "cover more cases" makes it less recallable and less verifiable. Edge cases and nuance live in their own atomic entries, joined by citation edges — not in growing paragraphs.
High-certainty action on low-confidence rows. Confidence is load-bearing. 30 = draft means "starting point, don't bet on it." Agents that act on drafts as if canonical produce drift the system can't recover from.
AI-readable artifacts surfaced to humans. Long slugs, ISO timestamps, always-true columns. Humans don't review what they can't scan; un-reviewed knowledge stays at draft.
The action hierarchy
When working with the hub, prefer in this order. Each step is preferable to the next:
RECALL — search the hub. Does it already know this? (knowledge-syntropy.md recall protocol.)
REFINE — found a related atom that's slightly off, stale, or bloated? Sharpen it in place. Commit raises confidence and shortens the entry. This is the most valuable move; most knowledge work should look like this.
CITE — the new claim is a relationship between existing atoms, or an example/edge case of one? Write a citations edge, or a new atomic entry that cites the parent. Never inline.
WRITE ATOMIC — no existing atom covers it. Write a new sharp entry with "use when" framing, then cite anything related.
EXTEND — anti-pattern. Adding paragraphs to an existing atom to "cover more cases" is bloat. If a refinement doesn't fit, the new content is a sibling atom, not an addendum.
Decision tree
Walk top-to-bottom. Stop at the first match. Mirrors the action hierarchy.
Does the hub already know this? Recall first. If yes and the existing atom is sharp — cite it, don't restate.
Is the existing atom muddy, stale, or bloated? Refine in place — shorten, sharpen, raise confidence. New commit. This is the default move.
Is the new claim a relationship between existing atoms, or an example/edge case? Write a citations edge (supports / contradicts / extends / supersedes) or a sibling atom that cites the parent. Never inline "companion to X" prose.
Is this a brand-new atomic claim? Write a sharp knowledge row with entry_type, registered domain, full provenance, "use when" framing in the title/content.
Is this a composite — guide / playbook / skill spanning multiple atoms? Write the composite row + outgoing citations to its constituents. Composite confidence inherits from leaves; don't fake it.
Missing entry_type or domain? Add the entry-type to the syntropy spec (same PR), or register the domain via the registry (not a code change).
Fundamentally new shape — lifecycle, indexes, relationships not modelable as citations? Propose a new table with a syntropy spec amendment in the same PR.
Need a new .md doc under docs/spec/knowledge-* or docs/design/knowledge-*? Almost never. Append to an existing section, or — better — write it as a knowledge entry in the hub itself.
If you reach step 8, you're sprawling in git, where humans can't recall it. Knowledge belongs in the hub.
Non-negotiable invariants
(full list in the specs — these are the ones that get violated)
REFINE_OVER_EXTEND — sharpen atomic entries; never lengthen them. Edge cases get a sibling atom + citation, not a new paragraph. This is the most important syntropy rule.
RECALL_BEFORE_WRITE — search before researching, research before writing. Skipping this is the second-largest entropy source.
ATTRIBUTION_TRACEABLE — every entry + commit traces to its contributor (principal, source node). Lineage is never anonymized, never overwritten. Foundation for cross-node reputation and downstream equity mechanics.
DOLT_IS_SOURCE_OF_TRUTH — Postgres search index is derived and rebuildable.
AUTO_COMMIT_ON_WRITE — every write commits via the capability layer.
EXTERNAL_WRITES_TO_BRANCH — bearer agents → contrib/*; only session users merge to main.
CROSS_LINKS_ARE_EDGES_NOT_COLUMNS — a relationship between any two entities (knowledge↔knowledge and work-item↔knowledge) is a single citations row, never a ref-column duplicated on each endpoint (that fractures — the two copies drift). The work-item↔knowledge link generalizes citations to allow a work-item endpoint (id-shape task./bug./…\d+); it is authored via the contribution cite op (so links are curated, validated both-endpoints-exist-on-main) and read both directions via the indexed citing_id/cited_id. Work-item lifecycle (claim/status/PR) stays direct-write (autonomy); cross-links are curated knowledge. work_items holds no link data. Full rationale + roadmap: hub entry work-knowledge-write-planes (meta) + work item story.5017.
No work-item fan-out. Capture next steps as prose on the current item, not a fan of follow-up tasks.
No parallel docs. Refine an existing section in place, or write the content as a knowledge entry.
No backwards-compat shims. Refactor in place.
UI must not leak storage shape. Slugs, ISO timestamps, always-true columns belong in <details> or out entirely. Humans need title + relative time + citation chips.
Canonical sources
This skill is the synthesis; the docs hold the detail. When they disagree, fix whichever is stale — they should reinforce, not duplicate.
What
Where
Live status, scorecard, top-3 asks, active PR stack
