SkillNote synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
SkillNote is a self-hosted procedural registry: authors retain reusable skill descriptions and bodies, and shipped adapters make them available to external agents. Its strongest supported feedback contribution is narrower than its improvement claims: retained usage and comments can change the order and membership of a later candidate response without rewriting a procedure. Whether that improves future task capacity remains unresolved. The claims that agents learn repeated patterns and that the registry improves over time are doctrine, not demonstrated outcomes. See EPI-CLM-pattern-learning, EPI-CLM-registry-improves and MEM-RTE-feedback-catalog.
The evidence is SRC-1, the SkillNote repository at commit 7ea2fc7a8435af64d1c080d8765c0eabba1121ea, with an applicability cutoff of 2026-10-14. The whole-system boundary includes browser authoring, backend state, imports and archives, CLI maintenance, host hooks, OpenClaw instructions, MCP, the claude.ai connector and feedback. It is code-grounded but selectively inspected. External hosts, their models and context semantics, remote source/package bytes and deployed state are not frozen. No target execution, host transcript, actual skill corpus or causal experiment was supplied. Thus implementation and policy can be characterized, but successful operation and benefit cannot. See the boundary and Source register.
From a convention to retained guidance
The representative workflow turns a persistent convention into a reviewed registry skill, then distributes it to a configured Claude Code project. An explicit user request starts authoring; for unsolicited creation, the skill-push prompt requires permission before drafting. An optional inherited-model skill-creator role can draft and check trigger quality, brevity and examples. Its availability does not establish a mandatory specialist launch, fixed handoff or isolated child return. The foreground author checks slug existence, offers collection choices, shows a complete preview and requests approval before POST or PATCH. Invalid collection names are surfaced for user correction; a slug collision directs the author toward PATCH. The delivered authoring result is committed registry content and a viewing link, not a completed later task. See the workflow account and RT-RTE-registry-write.
User review and backend admission are different gates. Review is a host-followed natural-language policy, with trigger suitability as a named criterion; it is not observed truth acceptance. Backend schema, collection and conflict checks can reject structurally invalid writes, but receive no required approval receipt or independent task oracle. Direct callers, imports, archives and connector policies have other admission paths. Moreover, report verification found a separate synchronous prompt-watch hook that retains pending sentence drafts automatically and emits authoring pointers. No automatic later draft-adoption chain was established. Consequently, the approval policy cannot be generalized to every draft or registry write. See RT-RTE-registry-write, EPI-OBJ-skill-push and the report-verification limits.
Current skill rows hold natural-language descriptions and instructions alongside symbolic IDs, collections and source fields. Saving creates a content snapshot; restore creates another snapshot, while set-latest reselects an existing one. History omits extra frontmatter and origin fields and does not require an edit rationale. Published ZIPs are independent retained products, not interchangeable with Markdown snapshots. Browser-local fallback can survive an API error without registry admission. Reverse import can stage a non-current body, automatically adopt it under policy, or expose a resolution choice; rejecting staged content can delete its bytes while preserving an audit identifier. These distinctions determine what can be restored, distributed or inspected later. See MEM-OBJ-skill, MEM-OBJ-content-version, MEM-OBJ-skill-version and MEM-RTE-inbound-resolution.
From retention to a later agent
SkillNote owns database, HTTP, archive and file operations, not the enclosing task loop. For Claude Code, hooks fetch collection-filtered current rows and write transformed SKILL.md files plus a manifest into project discovery paths. Collection matching selects content before host discovery; wildcard supply is untargeted. Codex has a related handler with its own project lock and stdout policy. MCP instead exposes current rows as named tools and returns requested instruction text. CLI add and update install selected active archives, or generated current bundles for never-published skills, through agent adapters. Checksum and archive-entry checks constrain bytes and extraction, not procedure correctness or the agent's subsequent actions. The browser connector uses queued operations and a logged-in browser session for remote delivery, with best-effort group retirement. See RT-RTE-project-sync, RT-RTE-mcp-delivery, RT-RTE-bundle-install, MEM-RTE-install-maintenance and RT-RTE-browser-sync.
At the eventual task consumer, prose is guidance interpreted under an external host's instruction hierarchy, not a SkillNote-enforced task invariant. The host chooses what to read and do. OpenClaw's shipped instructions request description-first matching, ordinarily at most three matches and never more than five loaded bodies. These are policy body-count limits, not enforced token budgets. User consent attaches its sidecar to AGENTS. Subagent and compaction hooks supply pointers or catalog reminders rather than a retained task-reasoning summary; model-visible PostCompact delivery remains unresolved. Neither the context-bundle resolver docstring nor the optional authoring specialist establishes an executed isolated work frame. See RT-BAP-host-instructions, RT-CMP-host-model, RT-RTE-openclaw-use and RT-RTE-context-hooks.
Feedback changes availability, not warrant
Completion prompts ask the host for ratings and outcomes. Code checks score ranges and some references; it does not independently verify task success. Comments, ratings and usage events have separate stores and consumers. Rated agent comments also create rating rows, but deleting a comment does not retract that fan-out in the inspected handler. The watcher extracts skill-file reads from host logs into analytics and retains replay state; a read is not proof of application, and those call events are not the usage-event store counted by the context bundle. See RT-RTE-feedback-report, EPI-RTE-comment-check, MEM-OBJ-comment, EPI-RTE-rating-admission and MEM-RTE-trace-reporting.
On a context-bundle request, the server takes an alphabetic window of four times the requested size, orders it by 30-day usage, comment-rating mean and name, then truncates it. The request's task summary does not enter this ranking. Any deprecation warning or low mean yields needs_review; no warning or rating can still yield healthy. These labels classify stored reports, not semantic correctness, and do not withdraw a skill. Only the latest comment's first 200 characters are delivered. Arithmetic conclusions concern retained rows, whose outcome truth remains unchecked. A deployed resolver consumer and successful task matching are unestablished. An additional prompted large-catalog consumer uses analytics top-skills, whose call-count and SkillRating aggregates differ from the bundle's Comment means; the bundle is therefore not the only feedback surface. See EPI-RTE-bundle-metadata, EPI-RTE-bundle-selection, MEM-RTE-feedback-catalog and the report-verification limits.
Feedback-dependent selection is implemented capacity. It does not establish accepted new truth, automatic procedure correction, weight learning or improved future-action capacity. The bundle is recomputed rather than retained as a durable derived procedure. Unknown evaluation consumers and caller/payload production prevent a complete negative about trace learning. Separately, reliable reflection and autonomous content criticism remain unestablished. The runtime supplies no condition-by-condition theory-builder assessment: storage and delivery do not establish consumption of a formulated theory, content-directed criticism or iteration, and no corpus resolves which skill contents qualify as theories. Conditions 1–4 therefore cannot be affirmed collectively or treated as collectively absent. Nor is there evidence that criticism of a consumed theory improved future capacity. See the epistemic bounded conclusion, MEM-OBJ-content-version and profile verification.
The OpenClaw adapter has an automatic update route, but unequal version strings and installation are not an improvement experiment. No independently assessed objective or subsequent operation causally dependent on an evidence-shaped organizational change is established. Exercised self-improvement and successful improvement therefore remain unestablished independently of feedback ranking, reflection and autonomy. See RT-RTE-adapter-update.
Maintenance and scenario-relative assessment
For the scenario of retaining a reviewed team convention and making it discoverable in later configured projects, the source supports persistence, collection scoping and delivery machinery. It does not support assured host application. For the scenario of withdrawing a harmful procedure everywhere, controls are narrower: disabled archives block covered downloads, but current-row MCP and cached copies can remain available. CLI update can skip a withdrawn release without deleting its installation. Sync guards preserve caches on some empty or failed responses, while the interactive picker can independently delete managed files. Neither uniform freshness nor universal revocation follows. Git re-import can change metadata without replacing a same-source body, and refresh apply is a stub. See RT-RTE-bundle-install, MEM-RTE-install-maintenance, RT-RTE-project-sync, RT-RTE-source-import and the report-verification limits.
The assessment would change with linked authoring, approval, synchronization and host-use traces identifying the actual consumed versions; complete cleanup and feedback-consumer tracing; and comparisons that independently vary feedback or revised guidance against a declared task objective. Demonstrating content-directed criticism requires the challenged claim, evidence, blame placement and resulting revision or changed reliance, followed into a next round. Demonstrating improvement requires evidence of better future capacity at the comparison's actual scope, not merely saved feedback or a replaced file. These additions would address the gaps in RT-RTE-registry-write, RT-BAP-host-instructions, EPI-RTE-bundle-selection and RT-RTE-adapter-update.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed operation, corpus or causal comparison | SRC-1; RT-RTE-registry-write, RT-CMP-host-model, EPI-CLM-registry-improves | Frozen source implementation and doctrine; external hosts and deployed state excluded | Successful end-to-end delivery, truthful outcomes, reliable compliance, learning benefit and exercised self-improvement | Version-linked approval/delivery/use traces and objective-specific intervention comparisons |
| Conversational candidate inventory omits prompt-watch | SRC-1; RT-RTE-registry-write, EPI-OBJ-skill-push; report-verification limit | Synchronous hook and pending-draft writer inspected by verifier; bounded search found no automatic later consumer | Complete candidate-acquisition or trace-fed draft inventory; all drafting waiting for approval | A retained route account of extraction, draft storage, dismissal and any later consumer, with linked execution evidence for activation |
| Installed-copy cleanup inventory omits picker paths | SRC-1; RT-RTE-project-sync, MEM-RTE-install-maintenance; report-verification limit | Setup wrapper and picker cleanup inspected by verifier alongside scoped sync/CLI routes | Cross-entry-point cache preservation or complete withdrawal comparison | Integrated picker, sync and CLI cleanup account, including failure branches and host reload behavior |
| Feedback consumer inventory omits prompted analytics top-skills use | SRC-1; RT-RTE-openclaw-use, RT-RTE-feedback-report, MEM-RTE-trace-reporting, MEM-RTE-feedback-catalog; report-verification limit | Verifier traced analytics producer and large-catalog prompt, not host execution | Complete feedback-to-selection lineage; bundle as sole consumer; reliable selection benefit | Separate retained analytics-consumption route and host traces linking its aggregates to later selection |
| Profile units group distinct objects, consumers and admissions | SRC-1; RT-RTE-project-sync, MEM-RTE-inbound-resolution, MEM-RTE-install-maintenance, MEM-OBJ-watcher-state, EPI-RTE-comment-check; profile-verification limit | Supported positive classifications and partial coverage, with coarse units | Faithful unit counts or one-to-one object/consumer/admission comparisons, not supported existence findings | Split classification units preserving each mechanism's scope and evidence basis |
| Opaque payloads and incomplete provenance | SRC-1; MEM-OBJ-skill, MEM-OBJ-skill-version, MEM-OBJ-usage-event | Schemas and admission/delivery branches; archive contents, arbitrary metadata and remote-origin bytes unresolved | Uniform form, authority, authorship, source warrant or rationale preservation | Pinned payload corpus and traced interpretation/provenance at each consumer |
| Context and resolver semantics remain external or unresolved | SRC-1; RT-RTE-context-hooks, RT-RTE-openclaw-use, RT-CMP-host-model | Hook outputs and selection policy, without host traces | Model-visible PostCompact recovery, isolated resolver execution, exact model mutability or enforced loading budget | Host delivery/discovery contracts and invocation traces showing surviving context, actual loaded bodies and returns |
| Selective maintenance and evaluation coverage | SRC-1; RT-RTE-source-import, RT-RTE-browser-sync, MEM-RTE-inbound-resolution, MEM-RTE-install-maintenance, MEM-RTE-trace-reporting | Selected branches; full inspector, conflict detector, shared-install ownership and evaluation consumers not audited | Complete refresh, collision recovery, cross-project cleanup, semantic acceptance or trace-learning absence | Targeted tracing of these available paths and retained accounts of their consequential consumers |