SKILL 32 / PKM · ROUNDUP

Personal Productivity Skills Roundup

Meeting notes · inbox · notes · calendar — you can install the plumbing, not the judgment

2026-10-03 · BigCat

Conclusion first for this category: install the plumbing layer, let the official memory layer do its thing, and write the output layer yourself. Below are 11 real, installable or readable entries, each with a link, an install command, and a when-not-to-install. One of them (apple-mcp, 3.1k stars) was archived read-only on 2026-01-01 — that is the hardest honest cost in this issue.

Same series → Skill 9 MCP Server Roundup · Skill 8 AGENTS.md and Portability · Skill 6 Official Skill Authoring

// WHY THIS MATTERS

Meeting notes, inbox, note cleanup, calendar: the four things that look most delegable to AI, and the category where off-the-shelf products help least. The reason is simple. Coding skills have an objective oracle (do the tests pass). Office skills have a deterministic format (is the docx still valid). But what belongs in this meeting's notes is known only to you. So read the table in three layers — pick the wrong layer and you will shop for output skills by star count, then demand that a pipe sound like you.

// 01

The shortlist: 9 candidates + 1 counter-example + 1 tool for building your own

Claim: three different kinds of thing share one table. Locate the layer first, then argue about installing.
LAYER ENTRY KILLS FOR YOU SKIP IT WHEN ─────────────────────────────────────────────────────────────────────────────────────── OUTPUT 1 notion-workspace-plugin (Notion, 1st party) pre-reads / notes / reports not on Notion 2 anthropics/skills · internal-comms update / memo / FAQ format your formats differ MEMORY 3 auto memory + CLAUDE.md (built in) re-explaining who you are — (already there) 4 basicmachines-co/basic-memory conclusions evaporating you have note rules 5 @modelcontextprotocol/server-memory forgetting across sessions you want production PIPE 6 MarkusPfundstein/mcp-obsidian hand-pasting from the vault vault is local 7 Doist/todoist-mcp typing todos by hand todos live elsewhere 8 nspady/google-calendar-mcp hunting for a free slot you just want today 9 Readwise official MCP highlights never reused not on Readwise ANTI 10 supermemoryai/apple-mcp (once Notes / Mail / more) archived — skip BUILD 11 anthropics/skills · skill-creator writing skills by vibes one-off work only

Install: every command you can copy

# 1 Notion first-party plugin (4 skills + 10 slash commands)
/plugin marketplace add makenotion/claude-code-notion-plugin
/plugin install notion-workspace-plugin@notion-plugin-marketplace

# 3 Built in: the toggle and the viewer are the same command
/memory

# 4 A local markdown knowledge graph
uv tool install basic-memory --prerelease=allow

# 6 Obsidian (install the Local REST API plugin first, grab the API key)
uvx mcp-obsidian   # set OBSIDIAN_API_KEY / OBSIDIAN_HOST / OBSIDIAN_PORT

# 7, 8, 9 — todos / calendar / highlights
npx @doist/todoist-mcp
npx @cocal/google-calendar-mcp
claude mcp add --transport http readwise https://mcp2.readwise.io/mcp

Entry by entry: what it kills, when to skip it

Both traps here are about names. (1) "Obsidian MCP" returns at least six same-named repositories, forks included; the one named here is the 4.5k-star MarkusPfundstein/mcp-obsidian. (2) "Official Notion skills" maps to several same-named notion-mcp-server repos plus one archived mirror. Aggregators and search results are for discovering candidates only — verify every name against the GitHub repo tree or the vendor's own page.
// 02

Why these three layers cannot be compared together

Claim: the three layers have completely different acceptance tests, so one set of criteria will pick wrong.
PIPE judge on: coverage · are writes reversible · who maintains it · tool count failure mode: connected, but the model picks the wrong tool — or one delete_file erases your note buy off the shelf? — yes, and you should. No "house style" lives here MEMORY judge on: where it lives · who can read it · does it fight your notes failure mode: two indexes each half full, and you cannot say which to trust buy off the shelf? — start with the built-in one; add more only if short OUTPUT judge on: does it sound like you — there is no second criterion failure mode: grammatically perfect output that you would never have written buy off the shelf? — take its structure only; fill the style slot yourself

Two classic ways to get this wrong: shopping for output skills by star count (you install another company's status-update template, rewrite half of it every time, and would have been better off without it), and demanding that a pipe sound like you (which produces the conclusion "MCP is useless" — it was only ever moving data in).

That is also what SKILLS.md means when it says this category is the one you should build yourself. Not because off-the-shelf quality is bad, but because the output layer's acceptance test lives inside you. Which is why the structure most worth copying is internal-comms: one SKILL.md hanging four sample files — swap the samples, swap the house style.

// 03

A three-minute inspection of the SKILL.md or README

Claim: six criteria are specific to this category, and two of them barely matter elsewhere.

How to spend the three minutes: one minute on the tool list (that answers 1, 2, 5 and 6 at once), one on the top of the repo page and the commit history (3), one on whether the directory tree has a sample slot (4).

// 04

The honest cost

Claim: the more completely you install this category, the likelier some morning finds half your pipes broken.

1. Abandonment, and of high-star projects

supermemoryai/apple-mcp: 3.1k stars, archived read-only on 2026-01-01. The Notion skills mirror team-attention/notion-skills-for-claude: archived 2025-12-22. A personally maintained integration layer depends on an upstream API not changing, and upstreams never promise that. So star counts are worth less here than in any other category — look at the maintainer.

2. The OAuth tax is paid in real time

google-calendar-mcp wants you to create a Google Cloud project, enable the API and configure OAuth desktop credentials, and its README says outright that test-mode tokens expire after seven days. To save yourself "open the calendar and look," you bought a weekly re-authorization. Do that arithmetic before installing.

3. Two indexes fighting

basic-memory has its own SQLite index and ~/basic-memory directory; your Obsidian vault has its own. Run both and "that decision from last week" exists as half a record on each side. There must be exactly one source of truth: either point basic-memory at the vault directory, or keep only one.

4. The largest cost is not in the tools: the output layer cannot be outsourced. Notice where the official docs draw auto memory's boundary — it records only what you taught it (corrections, preferences, decisions) and explicitly skips what is derivable from the codebase and what your CLAUDE.md already says. The vendor does not decide what is worth remembering for you. So if you expect that installing a skill will make meeting notes come out the way you write them, that expectation is doomed in this category: the ceiling on off-the-shelf value is skipping the steps where you open the app, not having your judgment done for you.

// ONE-PAGE SHOPPING CARD

Install these three first: (1) leave /memory on and write a twenty-line CLAUDE.md — free, first-party, applies to every project, and because auto memory skips what your CLAUDE.md already says, the two complement each other. (2) Install exactly one pipe, the one matching your note system: Obsidian → uvx mcp-obsidian (install obsidian-local-rest-api first); Notion → /plugin install notion-workspace-plugin@notion-plugin-marketplace. (3) anthropics/skills · skill-creator — turn the thing you re-explained three times this week into one skill. That is this category's real answer.

Add as needed: Doist/todoist-mcp (if your tasks really are in Todoist) · the Readwise official MCP (if you have highlights in flight) · basicmachines-co/basic-memory (if you have no note system yet, it can be that system) · anthropics/skills · internal-comms (when you need update or announcement formats — copy the structure, swap the samples).

Explicitly do not install: supermemoryai/apple-mcp (archived) · @modelcontextprotocol/server-memory (it calls itself a basic implementation; read it as a model) · basic-memory alongside Obsidian as two indexes · three servers that can all "search notes" at once · any output-layer template with someone else's house style baked in.

// GLOSSARY

PKM (personal knowledge management)
The pile of private context: notes, highlights, meeting records, calendar.
Pipe layer / memory layer / output layer
This issue's three layers. Pipes move data, memory persists across sessions, output produces the finished thing. Different acceptance tests, so never compare across them.
auto memory
The notes Claude Code writes for itself, in four kinds: user, feedback, project, reference. Machine-local, isolated per repository, and only the first 200 lines or 25KB load.
Sample slot (examples/)
The directory under an output-layer SKILL.md left for you to fill with your own house style. Whether that slot exists decides whether the skill can ever be yours.
Context tax
Every server you connect parks a batch of tool definitions in context, consuming it and raising the odds of picking the wrong tool.
Abandonment
A repository archived read-only. The most common death for a personally maintained integration layer, and unrelated to star count.

// THREE THINGS I CANNOT SETTLE

Why does "build it yourself" hold for this category but not for coding?
The difference is whether an objective oracle exists. Coding criteria live outside the repo — tests pass or they do not, types check or they do not — so a stranger's code-reviewer can be useful to you directly. This category's criteria live in your head: nothing arbitrates "what belongs in this meeting's notes" or "why this highlight matters." The inference is deflating but useful: the ceiling on off-the-shelf value here is skipping the steps where you open the app. If you want judgment done for you, you have to write it — which is why skill-creator sits in the top three of the shopping card.
Auto memory already ships. Do I still need something like basic-memory?
Two questions decide it: where it lives, and whether a human can read it. Auto memory lives under ~/.claude/projects/<project>/memory/, isolated per repository and machine-local (the docs state it is not shared across machines or cloud environments), and it writes only when it judges something worth keeping — it is for the model. basic-memory settles into markdown you can open in Obsidian — it is for humans too. So the question is not which is stronger but whether a person needs to read it. If you already have a vault, the answer is usually: do not add a second store, connect the vault.
Why does everything feel slower after installing a pile of pipes?
Two mechanisms stack. First, tool count: thirteen for calendar, seven for notes, nine for memory — thirty definitions riding along, read every turn. Second, and this is the one that actually costs time, overlap: when mcp-obsidian's search and the memory server's search_nodes are both present, the model must first decide where to look, and one wrong choice burns a whole round trip without telling you it chose wrong. The fix is crude and effective: one path per job; turn the overlapping one off.

// FURTHER READING