Linked notes

Linked Notes

Speak to the user in their language; these instructions are in English for the agent.

A generative skill that turns a session's insight into a small set of atomic notes under learn/notes/, each capturing exactly one idea, written in the user's own words, with citations to the artefact and links to related notes. The notes outlast the session and compose into a growing knowledge base.

Quick start

User: /linked-notes "event sourcing"
Agent: Project read first. Stack: TypeScript + EventStoreDB 23.x.
       I'll guide you through writing 3–6 atomic notes — one idea each, in your
       own words, with citations. You write; I check atomicity, drift, and
       citations. Start by telling me the first idea you want to capture.

After the session, files exist at learn/notes/<slug>.md, each containing one note with frontmatter, body, citations, and links to other notes.

Workflows

Mode detection

Inspect the argument:

• Code-mode — argument is a file or PR. Notes capture what this code teaches (patterns, invariants, surprises).
• Topic-mode — argument is a concept. Notes capture one idea at a time about the concept, anchored to the project's actual stack version.

Note format

See NOTE-FORMAT.md for the exact file location, frontmatter, body structure, and the three non-negotiable properties (atomic, in-the-user's-words, cited).

Session protocol

1. Project read (topic-mode). Read package.json / lockfile to identify versions. Citations must match the version in use.
2. One note at a time. Ask the user for the first idea they want to capture. Single sentence, declarative, in their own words.
3. Atomicity check. Read what the user wrote. Does it contain more than one idea? If yes, ask which one to capture first; the rest become a queue for follow-up notes.
4. Citation check. For every claim in the note about external behaviour (an API behaves this way, a feature works like this), ask the user to supply or accept a citation (file:line or doc link). Use Read / WebFetch / context7 to verify the citation is accurate before writing it into the file.
5. Wording check. Compare the body to known docs/code. If it is a near-quote, ask the user to restate. Real notes are paraphrases — paraphrasing is itself an act of learning.
6. Link suggestion (gentle). After writing the note, check learn/notes/ for existing notes with overlapping tags. Suggest 1–3 candidate links — the user accepts or rejects each. Never auto-link.
7. Write the file. Use the Write tool. Confirm filename and path with the user if it would collide with an existing slug.
8. Loop or close. Ask whether the user wants to capture another idea or close. If close, summarise: N notes written, M new links."

Linking discipline

• Links are markdown wikilinks: [[other-note-slug]].
• Bidirectional manually: if note A links to B, propose adding the reciprocal link to B.
• Don't create "index" notes. The folder structure + tags are the index.
• Don't backlink to source raw files: those belong in Sources, not in links.

File hygiene

• Slugs are kebab-case, derived from the title. Collisions: append -2, -3.
• The agent never modifies an existing note's body in the same session — only adds links (with user confirmation).

Workspace mode

If CURRICULUM.md exists in the current directory, the skill is running inside a Scenario B workspace. In workspace mode:

• Notes go to notes/<slug>.md (not learn/notes/).
• After each note, propose adding new terms to reference/glossario.html: "This introduces the term X — add it to the glossary?" The user decides; the agent writes the entry.

If CURRICULUM.md is absent (legacy mode), notes go to learn/notes/ as before.

Anti-patterns

• Don't write compound notes ("here's a note that covers X, Y, and Z") — non-atomic notes break the linking model and become unreusable; the act of forcing yourself to split them into atomic units is itself a learning act.
• Don't author body prose for the user — the agent prompts, verifies, formats; never writes the user's idea for them, because notes that aren't in the user's own words don't trigger the paraphrase-as-learning act that justifies the format.
• Don't cite from memory — every citation is verified against the real source before being written into the note, otherwise the durable artefact propagates borrowed confidence into the future.
• Don't auto-link based on string match — always propose, never auto-link; linking choices are the user's because the connections are their knowledge graph, not the agent's.
• Don't mix atomic notes with raw notes or sketches — atomic-only is what makes the collection composable; other captures (drafts, outlines, scratch) go elsewhere.

Governing principles (this skill satisfies all five)

1. Agent withholds — the agent prompts, checks, formats; never authors body content.
2. Student speaks first — every note is the user's words.
3. Artefact is the judge — every citation verified against the real source before being written.
4. Source fidelity — version anchoring + citation to authoritative sources on every external claim. Policy: ../../docs/sources.md.
5. Exit is a transfer test — the paraphrase requirement (own words, not quotes) is itself a transfer demonstration.