# Kb Lessons Bridge > Bridges reading-derived workflow insights into the self-improvement loop. Reads a kb-compile output note (project:knowledge-base, type:reference/zettel) or a Knowledge Intake synthesis note's "## Workflow implications" / "## Candidate actions" sections, filters for process/workflow-relevant insights (discarding general domain knowledge), and emits them as REC-ID blocks in the exact schema self-im… - **Type:** Skill - **Install:** `agentstack add skill-pfeff-claude-skills-kb-lessons-bridge` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [pfeff](https://agentstack.voostack.com/s/pfeff) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [pfeff](https://github.com/pfeff) - **Source:** https://github.com/pfeff/claude-skills/tree/main/skills/kb-lessons-bridge ## Install ```sh agentstack add skill-pfeff-claude-skills-kb-lessons-bridge ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # kb-lessons-bridge — reading → REC bridge ## The gap this closes Two loops exist and don't talk to each other: - **Reading → knowledge**: `kb-capture → kb-compile` turns highlighted Readwise sources into vault notes (`type: reference` / `type: zettel`, `project: knowledge-base`). CoS's Knowledge Intake mode also writes synthesis notes with `## Workflow implications` and `## Candidate actions` sections. - **Experience → process improvement**: `lessons-learned → self-improvement → skillify`. But `self-improvement`'s *only* input is lessons-learned documents — it globs `Generated/*Lesson*.md` and regex-scans for `#### REC-\d+:` blocks. It never looks at KB notes. Reading-derived workflow implications currently dead-end at operator grooming. This skill is the missing pipe: it does not reimplement `lessons-learned`'s Five-Whys discussion or `self-improvement`'s scan/analyze/apply cycle — it only produces a **drop-in** lessons-learned note so those two skills work on reading-derived insight exactly as they already work on session-derived insight. ## Trigger **Standalone / explicit invocation only.** This skill is never auto-wired to the end of `kb-compile` or CoS Knowledge Intake — it runs only when explicitly invoked (natural-language request or a future `/kb-lessons-bridge` command). Chaining it automatically would remove the confirm-before-write gate's meaning (see Step 4) and is out of scope for this skill. ## Schema this must match (verified against source, not assumed) Read directly from `${CLAUDE_PLUGIN_ROOT}/skills/self-improvement/operations/scan-recommendations.md` (Step 2 extraction regex) and `${CLAUDE_PLUGIN_ROOT}/skills/lessons-learned/operations/single-session-analysis.md` (body template): - **Discovery glob**: `ls -1t "$OBSIDIAN_VAULT_PATH"/Generated/*Lesson*.md` — the output filename **must contain the literal substring `Lesson`**. - **REC block regex** (`self-improvement` extracts on this exact pattern): ``` #### REC-(\d+): (.+) **Category**: (.+) **Effort**: (.+) **Impact**: (.+) ``` followed by a `**Target**:` line, free-text description, a `**Implementation**:` line, and a `---` separator before the next REC. Category must be one of `Workflow | Commands | Utilities | Skills` to survive `self-improvement`'s Step 3 category filter (Documentation / User Practices / External Tools get excluded there). - **Frontmatter** (per `lessons-learned`'s Phase 4): tags **must** include both `generated_note` and `lessons_learned`; `keywords:` as capitalized, space-separated wikilinks. Do not invent new tags — pull the vocabulary from the `obsidian-notes` skill. - **Filename base pattern**: `lessons-learned` itself writes `Generated/YYYYMMDDHHmm-Lessons Learned - {topic}.md`. This skill's own output (Step 5 below) adds a `Reading Bridge -` segment ahead of `{topic}` so a bridge-generated note is distinguishable from a session-derived one at a glance — the added segment doesn't change the `*Lesson*.md` discovery glob match (both still contain the literal substring `Lesson`). - **REC numbering is per-note, not global** — `self-improvement` scans `#### REC-\d+:` within each lessons file independently, so a bridge-generated note numbering `REC-001, REC-002, ...` from 1 in its own file is safe even though other lessons-learned notes also start at `REC-001`. **Do not add a bridge-specific REC-ID prefix** (e.g. `REC-R1`) — that would break the `REC-(\d+)` capture group other tooling relies on. Provenance instead rides the **`**Source**:`** line (per-REC) plus a `[[Reading Derived]]` frontmatter keyword (note-level) — both additive, neither touches the regex. Nothing above is reimplemented logic — it is the literal contract the two existing skills already speak. This skill's only job is to produce a note that satisfies it, sourced from reading instead of a live session. ## Inputs One or both of, discovered by **Globbing the vault directly** (never the plugin tree — this skill has no source data of its own, only the vault does): 1. **A kb-compile output note** — `Notes/YYYY/MM/*.md` with `type: reference` or `type: zettel` and `project: knowledge-base` frontmatter (written by `kb-compile`). 2. **A Knowledge Intake synthesis note** — a vault `Reference` note (`type: reference`, `area: workflow`) containing `## Workflow implications` and `## Candidate actions` sections. CoS's Knowledge Intake mode writes these; the CoS skill itself lives at `~/.claude/skills/chief-of-staff`, outside the `claude-skills` plugin tree, so this skill never reads CoS's definition — only the vault notes it produces. If invoked with no argument, ask which note(s) to bridge (or accept a path/title directly). ## Eligibility filter — process/workflow-relevant only An insight qualifies for a REC **only if** it proposes a change to how Claude/the operator's system works — mirroring `lessons-learned`'s own Phase 3 INCLUDE/EXCLUDE test: **INCLUDE** — insight suggests: - A new or changed skill operation, slash command, or agent behavior - A workflow/process change (e.g., "always do X before Y") - A tool-usage pattern or permission-whitelist candidate - A documentation gap in an existing skill **EXCLUDE** — insight is: - General domain knowledge with no workflow-change implication (this is what `kb-compile` already captured correctly as a `reference`/`zettel` note — leave it there) - A one-off fact about the subject matter, not about how work gets done - Too vague to name a target skill/command/file even approximately When in doubt, exclude — a missed REC costs nothing (the source note still exists and can be re-bridged later); a spurious REC pollutes `self-improvement`'s queue. ## Steps 1. **Read the source note(s)**. For a kb-compile note, read the full body (reference/zettel notes are short). For a Knowledge Intake note, read only `## Workflow implications` and `## Candidate actions` — ignore the rest of the synthesis. 2. **Extract candidate insights** — one per distinct suggested change. Apply the eligibility filter above; drop anything that's general knowledge. 3. **For each surviving insight, draft a REC block**: - `#### REC-{next-number}: {short title}` (numbered from 1 within this note — see the per-note numbering note above; no bridge-specific prefix) - `**Category**:` one of `Workflow | Commands | Utilities | Skills` - `**Effort**:` Quick / Medium / Complex (estimate) - `**Impact**:` HIGH / MEDIUM / LOW - `**Target**:` best-guess file path under the `claude-skills` repo (validate it exists per `lessons-learned`'s "Validate Target Paths" step — Glob for it; if it doesn't exist, either find the nearest real target or note "no existing target — would need a new skill") - Body: the insight, framed as a workflow change (not raw domain knowledge) - `**Implementation**:` what a PR to `claude-skills` would need to do - **`**Source**:` line** (additive — does not break the regex above, which doesn't require the block to end after `Implementation`): `[[{source note title}]]` plus the kb-compile `source_key` if the origin is a KB note. This is how traceability back to the original reading survives even after the insight is folded into an aggregated bridge note. 4. **Show the drafted REC(s) and ask for confirmation** before writing — same safety gate `lessons-learned` and `skillify` both use. Let the user cut, merge, or edit any REC. 5. **Write the note** once confirmed. **One bridge note per run**, even when multiple source notes were read in step 1 — every REC in the note carries its own `**Source**:` line, so batching multiple sources into a single note never loses per-REC traceability. - Path: `Generated/YYYYMMDDHHmm-Lessons Learned - Reading Bridge - {topic}.md` - Frontmatter: `tags: generated_note, lessons_learned`; `keywords:` wikilinks including `[[Lessons Learned]]`, `[[Reading Derived]]` (marks the note's provenance at a glance), plus the source topic(s); consider adding `sources:` (list of kb `source_key`s) as an extra field for provenance — harmless extra frontmatter, ignored by `self-improvement`'s scan. - Body: minimal — a one-line "Source" section pointing at the originating note(s), then `## Recommendations` with the REC blocks from step 3. Skip Problems/Five-Whys/Positive Patterns sections entirely — there's no session to retrospect on, only insight to route. - Use the `obsidian-notes` skill's CLI recipes exactly as `lessons-learned` Phase 4 does (`property:set` for frontmatter, `append` for body). Fall back to `docs/lessons-learned/` if the vault helper is unavailable. 6. **Report** the note path and REC count. Do not run `/claude-skills:self-improvement` automatically — that stays a separate, explicit invocation. ## What this is NOT - Not a reimplementation of `lessons-learned`'s Five Whys or Phase 1-3 discussion — there is no live session to retrospect on, only a written note to route. - Not a reimplementation of `self-improvement`'s scan/analyze/propose/apply cycle — this skill stops once the REC note is written. - Not a bypass of the confirm-before-write step. - Not a way to sneak general domain knowledge into the improvement queue — that's what `kb-compile`'s `reference`/`zettel` notes are for; leave it there. - Not auto-chained to `kb-compile` or Knowledge Intake — see Trigger above. ## See also - `kb-compile` — produces one of this skill's two input shapes - `lessons-learned` — owns the REC-ID schema and Obsidian note conventions this skill mirrors - `self-improvement` — the consumer; scans `Generated/*Lesson*.md` for `#### REC-\d+:` blocks - `obsidian-notes` — frontmatter/tag vocabulary and CLI recipes used for the actual write ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [pfeff](https://github.com/pfeff) - **Source:** [pfeff/claude-skills](https://github.com/pfeff/claude-skills) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-pfeff-claude-skills-kb-lessons-bridge - Seller: https://agentstack.voostack.com/s/pfeff - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.