glebis/publish-skill

Publish Skill to Site

Automate the full pipeline for publishing a skill to the claude-skills-site Astro website.

Paths

• Skills repo: ~/ai_projects/claude-skills/
• Site repo: ~/ai_projects/claude-skills-site/
• Site skills: ~/ai_projects/claude-skills-site/src/content/skills/*.mdx
• Site bundles: ~/ai_projects/claude-skills-site/src/content/bundles/*.mdx

Invocation

/publish-skill <skill-name>              # New skill — generates a hero image by default
/publish-skill <skill-name> --update     # Regenerate existing .mdx (+ hero image)
/publish-skill <skill-name> --dry-run    # Show what would be written, no commits, no image
/publish-skill <skill-name> --no-image   # Skip hero-image generation

Workflow

Execute these phases in order. Stop and report on any error.

Phase 1 — Locate and Read Source

1. Check ~/ai_projects/claude-skills/<skill-name>/SKILL.md exists. Also check ~/.claude/skills/<skill-name>/SKILL.md.
2. If neither exists, report error: "SKILL.md not found for <skill-name>. Ensure the skill directory exists in the skills repo or ~/.claude/skills/."
3. If both exist, compare them. If ~/.claude/skills/ version is newer (by mtime), sync it to the skills repo with rsync -av --exclude='*.egg-info' --exclude='__pycache__', commit the change, and push. This ensures the skills repo always has the latest version.
4. Read the SKILL.md (preferring the skills repo copy after any sync). Extract:
   • name from YAML frontmatter
   • description from YAML frontmatter
   • Body content (everything after the frontmatter closing ---)
5. Also check if ~/ai_projects/claude-skills/<skill-name>/README.md exists — if so, read it for supplementary context.

Phase 2 — Check for Existing MDX

1. Check if ~/ai_projects/claude-skills-site/src/content/skills/<skill-name>.mdx already exists.
2. If it exists and --update was NOT passed: stop and ask for confirmation. "The skill <skill-name> already has an .mdx on the site. Pass --update to regenerate, or confirm to overwrite."
3. If it exists and --update was passed: read the existing .mdx. Preserve these fields from the existing file (do not regenerate):
   • apothecary_name
   • hero_image
   • auto_activity
   • auto_last_synced (will be updated to today)

Phase 3 — Generate MDX Frontmatter

Generate each field:

name — Use the skill directory name (kebab-case). Must match the filename.

tagline — One sentence, max 80 characters. Summarize what the skill does in plain language. End with a period. Do not start with "A skill that..." — lead with the action verb or the thing it produces.

apothecary_name — A poetic 2-4 word name in the style of an old apothecary shop. Follow the existing pattern:

• "The Listener's Ear" (meeting-intelligence)
• "The Scholar's Deep Draught" (deep-research)
• "The Artificer's Bench" (developer-tools)
• "The Therapist's Grimoire" (cognitive-toolkit)
• Use "The [Noun]'s [Noun]" or "[Noun] of [Noun]" patterns.

bundle — Select the best-fit bundle by analyzing the skill's purpose against these categories:

| Bundle key | Name | Covers | |---|---|---| | meeting-intelligence | Meeting Intelligence | Transcripts, meetings, recordings, action items, video/audio processing | | communication | Communication | Email, messaging, Telegram, Google Workspace, Zoom | | research | Research | Web research, search, image sourcing, knowledge retrieval | | content-publishing | Content & Publishing | Image generation, presentations, PDFs, reports, brand assets, visual tools | | personal-analytics | Personal Analytics | Health data, dictation analytics, browsing history, self-reflection, hardware signals, knowledge visualization | | thinking-strategy | Thinking & Strategy | Decision frameworks, JTBD, dialogue modes, cognitive tools, self-design | | developer-tools | Developer Tools | TDD, LLM CLI, issue tracking, session search, code tools | | lab-consulting | Lab & Consulting | Lab meetings, client discovery, retrospectives, demos |

If the skill could fit multiple bundles, prefer the one where it adds the most differentiation. If genuinely unsure, ask the user.

tags — 3-6 lowercase kebab-case tags. Derive from the skill's domain, tools used, and output types. Check existing skills for tag reuse.

accent_color — Pick from the validated set: amber (default), blue, cyan, green, orange, purple, violet. Use amber unless the skill has a strong thematic reason for another color (e.g., hardware/signals → cyan, mental health → violet, nature/strategy → green).

auto_description — Copy the description field from the SKILL.md frontmatter verbatim. Truncate at 200 characters if longer.

auto_triggers — Extract 0-3 trigger phrases from the SKILL.md description (e.g., "queries like", "check my email"). Leave empty [] if no clear triggers.

auto_tools — Leave as [].

auto_last_synced — Today's date in YYYY-MM-DD format. Run date +"%Y-%m-%d" to get it.

auto_last_commit — Run git -C ~/ai_projects/claude-skills log -1 --format="%Y-%m-%d" -- <skill-name>/ to get the last commit date for that skill directory.

auto_activity — Leave as [] for new skills. Preserve from existing .mdx on --update.

install_command — Leave as "".

repo_path — The skill directory name (same as name).

dependencies — Extract from SKILL.md requirements/prerequisites sections. List Python packages, CLI tools, or APIs needed. Use [] if none.

hero_image — Leave commented out (# hero_image:) at this stage. Phase 6.5 generates a hero image by default and rewrites this field to the generated path. It stays commented only if --no-image was passed (or generation fails and no repo screenshot exists).

Phase 4 — Generate MDX Body

Write the body following this exact structure:

## What it does

[2-3 sentences describing the skill's primary function. Be specific about what it produces or enables. Reference concrete tools, APIs, or protocols if relevant.]

## Key features

[Bulleted list of 4-6 distinguishing features. Each bullet: **Bold label** — explanation. Focus on what makes this skill interesting, not obvious capabilities.]

## When to use

[1-2 sentences describing the trigger scenarios. Start with "When..." to match the existing pattern.]

Additional sections to include only when relevant:

• ## Signals or ## Modes — if the skill has named operational modes or signal types
• ## How it works — if the skill uses a non-obvious protocol (MIDI, MCP, etc.)

Do NOT include: installation instructions, quick start commands, code blocks, or dependency lists. Those belong in the SKILL.md and README, not the site .mdx.

Phase 5 — Validate

Before writing, verify:

1. Bundle exists: Confirm ~/ai_projects/claude-skills-site/src/content/bundles/<bundle>.mdx exists.
2. No duplicate: If not --update, confirm no .mdx already exists (handled in Phase 2).
3. Accent color valid: Must be one of: amber, blue, cyan, green, orange, purple, violet.
4. Name matches: The name field, repo_path, and filename must all match.
5. Tagline length: Must be under 100 characters.

If any check fails, report the specific error and stop.

Phase 6 — Write Files

1. Write the .mdx file to ~/ai_projects/claude-skills-site/src/content/skills/<skill-name>.mdx.
2. Update the bundle: Read the target bundle .mdx. Add the skill name to the skills: array if not already present. Append at the end of the array.
3. If --dry-run: Instead of writing, display the full .mdx content and the bundle change, then stop.

Phase 6.5 — Generate Hero Image (default; skip with --no-image)

This phase runs by default. Skip it only if --no-image was passed, or under --dry-run (which shows planned output without generating).

Generate a hero image using the /codex skill. Codex has a built-in image_gen tool that produces images without needing an API key. The image follows the site's apothecary visual language established in the original Codex design session. Fallback: if /codex is unavailable or its image_gen fails, regenerate with the gpt-image-2 or nano-banana skill using the same prompt and the same 16:9 target size/path.

Prompt template — fill in SKILL_NAME, APOTHECARY_NAME, SCENE_BACKDROP, SUBJECT, COMPOSITION_DETAIL, and AVOID_EXTRA:

Use case: stylized-concept
Asset type: 16:9 skill hero image for the Claude Skills site
Primary request: Create a skill-specific hero artwork for the skill "SKILL_NAME", apothecary name "APOTHECARY_NAME". No readable text.
Style reference: Match the existing main site hero style: dark apothecary workbench fused with developer terminal artifacts, hand-tinted copperplate engraving, technical diagram overlays, tactile paper grain, oxidized copper, medicinal green, ink brown, warm bone, tiny electric-blue signal accents.
Scene/backdrop: SCENE_BACKDROP
Subject: SUBJECT
Composition: Wide 16:9, COMPOSITION_DETAIL, quiet dark negative space left, no border.
Lighting: Low raking desk light, sharp engraving detail, warm copper highlights.
Avoid: no readable words, no logos, no human figures, AVOID_EXTRA, no stock-photo look, no neon sci-fi.

Exact prompts from existing hero images (use as style/structure reference):

tdd ("The Test Anvil"):

• Scene: A heavy workbench inside an apothecary cabinet, with brass measuring instruments, drawers, glass vials, and translucent command cards.
• Subject: A small blacksmith anvil used as a testing station, with three distinct stages represented by red wax seal, green reagent glow, and polished brass refactor tool; tiny checklist slips and code-bracket glyphs etched into metal.
• Composition detail: anvil and three-stage testing ritual centered-right
• Avoid extra: no generic fantasy forge, no smoke-heavy scene

jtbd ("The Hiring Compass"):

• Scene: A researcher's apothecary desk with interview cards, opportunity maps, tiny sample vials, brass compass, and drawer labels represented only by abstract marks.
• Subject: A brass compass pointing through customer-job constellations, annotated review-mining cards, small scales for scoring opportunities, and translucent terminal cards with abstract brackets.
• Composition detail: compass and opportunity map centered-right
• Avoid extra: no corporate stock imagery, no generic charts, no bright flat UI, no blurry glow

skill-studio ("The Automation Architect"):

• Scene: A precise apothecary drafting bench with brass calipers, small drawers, vellum workflow maps, index cards, and translucent terminal recipe cards.
• Subject: An architect's compass, interview question cards, a flowchart map, and small vial-like modules arranged into a structured automation blueprint.
• Composition detail: detailed lower and right areas, some quiet negative space for page text overlay if needed
• Avoid extra: no modern stock-photo look, no blurry glow, no readable UI text

telegram-telethon ("The Daemon's Relay"):

• Scene: An apothecary communications bench with relay coils, brass switches, labeled drawers with abstract marks, sealed message tubes, and glass vials.
• Subject: A daemon relay apparatus: brass telegraph key, paper-plane-like message glyphs etched on translucent cards, audio waveform phials, and blue signal lines running between drawers and terminal cards.
• Composition detail: relay apparatus and message tubes centered-right, darker negative space left
• Avoid extra: no Telegram logo, no brand marks, no modern phone mockup

agency-docs-updater ("The Publisher's Engine"):

• Scene: A compact apothecary publishing press connected to terminal recipe cards, archive drawers, reels, and document trays.
• Subject: A brass engine that transforms a meeting transcript scroll into three outputs: a video reel glyph, a bound documentation sheet, and a clean index card; conveyor belts, gears, redaction veil strips, and synchronized blue trace lines.
• Composition detail: publishing engine centered-right
• Avoid extra: no YouTube logo, no brand marks, no modern UI screenshot, no generic printer

vision-bench ("The Judge's Loupe"):

• Scene: A dark apothecary judging bench with comparison trays, specimen drawers, translucent image plates, and brass optical instruments.
• Subject: A large judge's loupe examining two image plates side by side, small scoring weights and calibration vials, vision-model glyphs as abstract eye diagrams, and blue signal traces connecting evaluation cards.
• Composition detail: loupe and image comparison plates centered-right
• Avoid extra: no realistic eyeballs, no modern UI screenshot, no generic camera stock image

Subject crafting guidelines:

• Map the skill's core action to a physical apothecary/workshop metaphor
• Each subject should have: a central object (anvil, compass, loupe, engine), supporting detail objects (cards, vials, glyphs), and subtle digital traces (blue signal lines, terminal brackets)
• The metaphor should be immediately recognizable to someone who knows what the skill does
• Keep subject descriptions to 1-2 sentences

Steps:

1. Craft all template fields (scene, subject, composition, avoid) based on the skill's purpose
2. Invoke the /codex skill with the assembled prompt, requesting a 16:9 image saved to ~/ai_projects/claude-skills-site/public/images/SKILL_NAME-hero.jpg
3. Verify the image was created and is approximately 1672x941 pixels (the standard size for existing heroes)
4. Update the .mdx frontmatter: change # hero_image: to hero_image: "/images/SKILL_NAME-hero.jpg"
5. Stage the image file alongside the .mdx in Phase 8

If Codex is not available or image generation fails, warn the user and continue without the image. The skill publishes fine without a hero image.

Phase 7 — Show Diff and Confirm

1. Run git -C ~/ai_projects/claude-skills-site diff to show all changes.
2. Present the diff to the user.
3. Ask: "Publish these changes? This will commit and push to claude-skills-site." Wait for confirmation.

Phase 8 — Commit and Push

1. Stage the changed files (include the hero image — generated by default):

   git -C ~/ai_projects/claude-skills-site add \
     src/content/skills/<skill-name>.mdx \
     src/content/bundles/<bundle>.mdx \
     public/images/<skill-name>-hero.jpg   # omit only if --no-image / no image was produced
   

2. Commit with message: feat: add <skill-name> skill to site
   • For --update: feat: update <skill-name> skill on site
3. Push: git -C ~/ai_projects/claude-skills-site push
4. Report the commit hash and confirm success.

Error Recovery

| Error | Recovery | |---|---| | SKILL.md not found | Check both paths, report which were tried | | Bundle .mdx not found | List available bundles, ask user to pick | | .mdx already exists (no --update) | Ask user to confirm overwrite or pass --update | | Git push fails | Show error, suggest git pull --rebase first | | accent_color invalid | Default to amber, warn user | | Skill already in bundle array | Skip bundle update, not an error |

MDX Template

---
name: "<skill-name>"
tagline: "<tagline>"
apothecary_name: "<apothecary_name>"
bundle: <bundle>
tags: [<tags>]
accent_color: <color>
# hero_image:
auto_description: "<description>"
auto_triggers: [<triggers>]
auto_tools: []
auto_last_synced: "<date>"
auto_last_commit: "<date>"
auto_activity: []
install_command: ""
repo_path: "<skill-name>"
dependencies: [<deps>]
---

## What it does

<body>

## Key features

<features>

## When to use

<when>