Claude skills for daily Vipassana practice: one skill, and what it deliberately refuses

Yes, exactly one. It lives at ~/.claude/skills/vipassana-matching/SKILL.md on the operator's machine, the file is 344 lines, and the YAML frontmatter declares allowed-tools: Bash, Read, Edit, Write, WebSearch. It is invoked when the operator says things like 'do matching' or 'check the waitlist.' What it actually does, end to end, is query the Neon Postgres waitlist, find pairs whose UTC sit times are within 60 minutes of each other, optionally web-search a candidate's name to verify they are who they say they are, prepare a list of recommended pairs, wait for the operator to confirm, then create Google Meet rooms and send confirmation emails through Resend. There is no model-call against Anthropic from inside the production website itself; the skill runs in the operator's Claude Code session. The website's own runtime AI footprint is one floating page-assistant chat on the /t/ guide pages, which uses Google Generative AI to answer questions about the public guide content and has no access to the matches table.

Because the technique it would otherwise be coaching is taught in a specific way by a specific lineage, and a Claude Skill is not the right vehicle for that role. Goenka-style Vipassana is transmitted at 10-day residential courses by authorized assistant teachers on a strict donation basis through dhamma.org. Technique instruction is reserved for that setting. A Claude Skill can be useful around the practice, on the logistics layer (who is sitting, when, with whom, has the calendar invite gone out, did they show up), and that is exactly the layer this skill covers. It can also be useful on the content layer of public, non-instructional guide pages, which is the only place this site uses an LLM at runtime. Anything operational about the practice itself, the skill is silent on.

The allowed-tools line is exactly: Bash, Read, Edit, Write, WebSearch. Five tools. Bash is for running the operator scripts (find-matches, send-matches, check-stats). Read is for reading the SKILL.md and the script files themselves. Edit is for occasional surgical changes to the send-matches.mjs PAIRS array. Write is for new query scripts saved into vipassana-cool/scripts/. WebSearch is for verifying a candidate's name and email domain before suggesting them as a buddy. What is not in the list: any Anthropic Messages call as a tool, any Slack or notification tool, any voice or chat tool that could reach the practitioner, any Calendar or Email tool that runs without an operator intermediary. The Calendar API and the Resend send happen inside scripts the operator runs, not as autonomous skill tools.

Step 1, run check-stats.mjs for a full status snapshot of the database (counts of pending, ready, contacted, engaged, matched, plus recent signups). Step 2, run find-matches.mjs to enumerate viable pairs under the ±60 minute UTC cap and applying the prior-pair guard. Step 3, query vipassana_emails for inbound replies that might change a candidate's eligibility. Step 4, present top-recommended pairs to the operator and wait for explicit confirmation. Step 5, for each approved pair create a recurring daily Google Calendar event with a Meet conference and per-person tracking tokens stored in meet_links. Step 6, edit send-matches.mjs with the approved PAIRS array (each entry naming person A id, person B id, and the two tracking URLs). Step 7, run send-matches.mjs which calls /api/admin/matches with sendConfirmation: true and a 2-second rate limit between sends. Step 8, report the result: X pairs sent, X confirming, X matched, X pending. There is no step 9. The skill stops at the report.

It means there is no path from the SKILL.md back to a meditator's inbox or device that does not first pass through the operator. The send-matches.mjs script writes the email body using static HTML builders in src/lib/emails.ts; the skill does not author the email at runtime. The intro template names the operator (Matt) and refers to the meditator's chosen partner by their actual first name and chosen sit time, both pulled from the waitlist row. The body of the email talks about logistics, not technique: opening the same Meet link at the same time every day, optionally playing a Goenka recording over screen-share, sitting for the chosen duration. There is no model-generated coaching, no after-action prompt, no encouragement loop. If a Claude Skill could write the email body, it could insert technique advice, and the design specifically routes around that.

Technically yes. Editorially no, and that decision is the entire reason this page exists. The existing Bash and Write tools could absolutely be wired to a Resend send or an SMS gateway, and a model call could absolutely generate a 'how to work with this difficulty' response. The reason the skill does not do this is not technical limitation but lineage respect. For anything operational about the practice, the right place is a 10-day residential course at dhamma.org with an authorized assistant teacher. A Claude Skill that pretended to fill that role would be giving people the wrong artifact and undermining a tradition whose entire transmission model is designed around in-person, authorized teaching.

Two layers, both deterministic. First, the skill's WebSearch tool is used in step 4 of the Quick Start to look up a candidate's name plus email domain or city. The findings get written to the research_notes column on waitlist_entries so future runs can see what was learned. Second, the auto-match cron at src/app/api/auto-match/route.ts enforces a 24-hour cool-down on contact_count = 0 signups before they enter the auto-pool, gives a 7-day re-try gap between attempts for anyone whose previous match closed, and skips anyone whose contact_count is 10 or more (status 'ready' bypasses the cap). The skill itself respects those database states; it never bypasses them. So a brand-new signup is intentionally invisible to the auto-pipeline for 24 hours, by which point a human (the operator, often via this skill) has eyeballed the row.

Not as part of the matching or the daily sit. The pair gets emails written by static HTML templates, a Google Calendar invite that arrives in their inbox like any other, and a daily Google Meet URL. There is no chat interface, no voice agent, no AI coach the meditator opens during a sit. If a practitioner browses the public guide pages on /t/, they may interact with the floating page assistant chat at the bottom right of those pages; that runs on Google Generative AI inside src/app/api/guide-chat/route.ts, has no access to matches, and is scoped to answering questions about the public text on the page they are reading. The matched daily-sit experience is two humans on a silent Meet call. Nothing else is in the loop.

Six lines of YAML frontmatter, then the body. The frontmatter is: name: vipassana-matching; description: 'Full Practice Buddy matching workflow for vipassana.cool — check submissions, research candidates, identify matches, send confirmation emails, and manage the waitlist'; allowed-tools: Bash, Read, Edit, Write, WebSearch. After the frontmatter, the body opens with 'Trigger this skill when the user asks to:' followed by six trigger phrases, then a Project Context table, then a Person Status Flow ASCII diagram, then a Match Status Flow, then the smart confirmation flow, the auto-expiry rules, the guards, the database schema, the Node.js query pattern, the matching workflow, the admin API reference, and the reusable scripts. None of those sections contain a prompt that would address a practitioner.

If you want logistics help around your own practice (calendar invites, finding a sit partner in your timezone, tracking that you and your partner both showed up to a Meet room), the matching skill on this site already provides that to you indirectly: sign up at /practice-buddy and you are in the pool. If you want a private operator skill on your own machine that does similar work for a small circle of friends, the SKILL.md file in this product is freely readable and the pattern is straightforward to copy. If you want technique guidance for your sits, that is not what a Claude Skill should be giving you, and the only honest pointer this page can offer is dhamma.org and a 10-day residential course with an authorized assistant teacher. Nothing on this site, and no Claude Skill anywhere, is a substitute for that.