name

digital-garden

description

Turn raw voice-dump transcripts (from Claude Code voice mode or the Claude web/app voice feature) into personal digital-garden essays — dated, first-person, cross-linked — and publish them into a lightweight React+Vite personal site. Bootstraps a new site from a bundled template if none exists yet, or slots into an existing digital-garden-style repo. Use when the user has voice memos, rambling transcripts, or stream-of-consciousness dumps they want turned into a personal site or blog. Triggers: digital garden, voice dump to site, personal site from transcript, turn my voice notes into a blog, build my digital garden.

Digital Garden Skill

A voice dump is raw material, not a draft. This skill's job is judgement: figure out which parts of a rambling transcript are actually an essay, write that essay in the person's own voice without inventing anything they didn't say, and slot it into a real site that looks like a personal digital garden — not a corporate blog.

Reference implementation this skill is modeled on: Sai Yashwanth's digital garden (source in ~/Dev/DigitalGarden). Read a few of its public/content/*.md files if you want to recalibrate voice before writing — short, personal, dated, unpolished-but-clear, heavy on cross-linking to the author's own past pieces.

Inputs

Ask for (or locate in the working directory):

The voice-dump transcript(s) — raw text from Claude Code voice mode, the Claude web/app voice feature, or any other dictation. Expect filler words, tangents, repetition, and multiple unrelated ideas mixed into one dump. Do not clean this up before reading it fully.
The target site — where the essays should live. If unspecified, use the current working directory.
Identity basics, only if bootstrapping a brand-new site (name, one-line bio, social links). Don't ask for these if the target site already exists — read them from its public/static/about.md instead.

Process

Step 1 — Detect: existing garden or bootstrap new

Check the target directory for public/content/ and public/static/article.md.

Exists → this is an existing digital garden. Work inside it, matching its established conventions (read 2-3 existing public/content/*.md files and the article.md index to learn its section names, dating style, and voice before writing anything new).
Doesn't exist → copy this skill's template/ directory into the target, then fill in the placeholders (__SITE_NAME__, __TAGLINE__, __GITHUB_URL__, __TWITTER_URL__, etc. — grep for __ to find them all) using the identity basics gathered above. Tell the user to run npm install && npm run dev to preview once content is in.

Step 2 — Segment the dump into essay-worthy ideas

Read the whole transcript before touching it. A single dump often contains zero, one, or several distinct essays. For each candidate:

It needs a real point — an opinion, a lesson learned, an event, a reflection. "Testing testing" or a dump that trails into someone's grocery list is not an essay; discard it.
Don't force one transcript into exactly one essay. Split cleanly if the person jumped topics; merge if they circled back to the same point twice.
If nothing in the dump clears the bar, say so — don't manufacture an essay out of thin material.

Step 3 — Write each essay in the site's voice, not a generic one

Rules, non-negotiable:

First person, no invention. Every claim, number, event, or opinion must trace back to something actually said in the transcript. Rambling can be restructured into coherent paragraphs; content cannot be added. If the person trails off or contradicts themselves, resolve it the way they clearly meant, or leave a note rather than guessing.
Short and direct. Look at the reference site's essays — most are 300-800 words, plain paragraphs, no bullet-heavy listicle structure unless the person actually spoke in a list.

Format (matches the reference site exactly):

# <Title>

*Written on <Month DD, YYYY>*

<optional one-line summary/hook, plain or italicized>

<body paragraphs>

Use today's date unless the transcript names a different date.

Cross-link. Scan the existing article.md index (or the essays already written this session) for topically related pieces and link to them inline, e.g. [the last time I wrote about this](./slug). This is what makes it a garden and not a pile of blog posts.
No fabricated media. Only reference an image if the user actually provides one. Don't invent placeholder screenshots.
Title and slug. Title is short and specific, never clickbait. Slug is kebab-case ASCII derived from the title, matching the flat/foldered convention already used in public/content/.

Step 4 — Save and index

Write each essay to public/content/<slug>.md (or into a subfolder if it's clearly part of a series, mirroring how the reference site nests twitter_essays/, ai_playbooks/, etc.).
Update public/static/article.md: insert a line *Mon DD, YYYY* __[Title](./slug)__ in reverse-chronological order within the right section. Default to whatever the site's general/miscellaneous section is called (e.g. "Mini Essays and Articles"); only use a "Series" section if this essay is explicitly part of a multi-post arc, and only touch "Pinned" if the user asks.

Step 5 — Identity pages are curated, not a dumping ground

If the transcript reveals genuine bio or life-event material (a new job, a milestone, a project shipped), it may belong on about.md, home.md, or timeline.md — but these pages are the person's curated self-presentation. Propose the addition and show the exact diff; never overwrite these files without the user confirming first.

Output

One .md file per essay under public/content/
public/static/article.md updated with the new entries
If bootstrapped: a working Vite+React site the user can run with npm install && npm run dev
A short summary to the user: which essays were written, which raw material was discarded and why, and any identity-page changes proposed but not yet applied

From the Turi Labs AI Agent Learn series — turilabs.in

name

digital-garden

description

Digital Garden Skill

Inputs

Ask for (or locate in the working directory):

The voice-dump transcript(s) — raw text from Claude Code voice mode, the Claude web/app voice feature, or any other dictation. Expect filler words, tangents, repetition, and multiple unrelated ideas mixed into one dump. Do not clean this up before reading it fully.
The target site — where the essays should live. If unspecified, use the current working directory.
Identity basics, only if bootstrapping a brand-new site (name, one-line bio, social links). Don't ask for these if the target site already exists — read them from its public/static/about.md instead.

Process

Step 1 — Detect: existing garden or bootstrap new

Check the target directory for public/content/ and public/static/article.md.

Exists → this is an existing digital garden. Work inside it, matching its established conventions (read 2-3 existing public/content/*.md files and the article.md index to learn its section names, dating style, and voice before writing anything new).
Doesn't exist → copy this skill's template/ directory into the target, then fill in the placeholders (__SITE_NAME__, __TAGLINE__, __GITHUB_URL__, __TWITTER_URL__, etc. — grep for __ to find them all) using the identity basics gathered above. Tell the user to run npm install && npm run dev to preview once content is in.

Step 2 — Segment the dump into essay-worthy ideas

Read the whole transcript before touching it. A single dump often contains zero, one, or several distinct essays. For each candidate:

It needs a real point — an opinion, a lesson learned, an event, a reflection. "Testing testing" or a dump that trails into someone's grocery list is not an essay; discard it.
Don't force one transcript into exactly one essay. Split cleanly if the person jumped topics; merge if they circled back to the same point twice.
If nothing in the dump clears the bar, say so — don't manufacture an essay out of thin material.

Step 3 — Write each essay in the site's voice, not a generic one

Rules, non-negotiable:

First person, no invention. Every claim, number, event, or opinion must trace back to something actually said in the transcript. Rambling can be restructured into coherent paragraphs; content cannot be added. If the person trails off or contradicts themselves, resolve it the way they clearly meant, or leave a note rather than guessing.
Short and direct. Look at the reference site's essays — most are 300-800 words, plain paragraphs, no bullet-heavy listicle structure unless the person actually spoke in a list.

Format (matches the reference site exactly):

# <Title>

*Written on <Month DD, YYYY>*

<optional one-line summary/hook, plain or italicized>

<body paragraphs>

Use today's date unless the transcript names a different date.

Cross-link. Scan the existing article.md index (or the essays already written this session) for topically related pieces and link to them inline, e.g. [the last time I wrote about this](./slug). This is what makes it a garden and not a pile of blog posts.
No fabricated media. Only reference an image if the user actually provides one. Don't invent placeholder screenshots.
Title and slug. Title is short and specific, never clickbait. Slug is kebab-case ASCII derived from the title, matching the flat/foldered convention already used in public/content/.

Step 4 — Save and index

Write each essay to public/content/<slug>.md (or into a subfolder if it's clearly part of a series, mirroring how the reference site nests twitter_essays/, ai_playbooks/, etc.).
Update public/static/article.md: insert a line *Mon DD, YYYY* __[Title](./slug)__ in reverse-chronological order within the right section. Default to whatever the site's general/miscellaneous section is called (e.g. "Mini Essays and Articles"); only use a "Series" section if this essay is explicitly part of a multi-post arc, and only touch "Pinned" if the user asks.

Step 5 — Identity pages are curated, not a dumping ground

Output

One .md file per essay under public/content/
public/static/article.md updated with the new entries
If bootstrapped: a working Vite+React site the user can run with npm install && npm run dev
A short summary to the user: which essays were written, which raw material was discarded and why, and any identity-page changes proposed but not yet applied

From the Turi Labs AI Agent Learn series — turilabs.in

digital-garden

Digital Garden Skill

Inputs

Process

Step 1 — Detect: existing garden or bootstrap new

Step 2 — Segment the dump into essay-worthy ideas

Step 3 — Write each essay in the site's voice, not a generic one

Step 4 — Save and index

Step 5 — Identity pages are curated, not a dumping ground

Output

More from this repository

More from this repository

Digital Garden Skill

Inputs

Process

Step 1 — Detect: existing garden or bootstrap new

Step 2 — Segment the dump into essay-worthy ideas

Step 3 — Write each essay in the site's voice, not a generic one

Step 4 — Save and index

Step 5 — Identity pages are curated, not a dumping ground

Output