Plan: Publish a curated skills marketplace

This document plans a separate GitHub repo — claude-code-skills — that exposes a curated subset of the skills in this settings repo as installable Claude Code plugins. It is the working plan for the work itself; after that work lands, this file can be deleted or shortened into a pointer.

Execution note: This plan is designed to be executable from a different machine that has gh installed. Before switching machines, make sure you have committed and pushed from this repo: this file, skills/bibliography-from-ids/examples/verify_cites.py, and skills/bibliography-from-ids/SKILL.md. On the target machine, git pull in ~/.claude first, then follow the runbook at the bottom of this file.

Why a separate repo (not this one)

This repo (claude-code-settings) is a personal dotfiles repo. Users clone it whole into ~/.claude to replicate Matthew's environment. A marketplace repo is the opposite shape: a catalog consumed via /plugin install, where users pick individual plugins without touching anyone else's config. Mixing the two confuses the purpose of both, mixes personal config with shareable artifacts, and makes the repo name misleading.

A clean split:

claude-code-settings — personal machine config (this repo, stays as-is).
claude-code-skills — public plugin marketplace containing bundled, audited, generic skills.

Proposed bundle structure (4 plugins)

Per audit, bundling skills by audience produces coherent installs. Matthew prefers bundling over one-plugin-per-skill.

Plugin	Audience	Skills
`research-toolkit`	academics, paper authors	`bibliography-from-ids`, `verify-citation-claims`, `consolidate-paper`, `import-content`, `bayesian-stats`
`file-utils`	anyone using Claude Code	`read-entire`, `read-large-md`, `parse-json`
`meta-conversation`	anyone using Claude, contemplative	`reflect`, `debrief`, `cookie`
`claude-code-tools`	Claude Code power users	`archive-logs`, `into-trustable-script`

Dropped (not Matthew's work or not portable):

code-simplifier — not authored by Matthew. Removed from consideration.
commit-push-pr, techdebt — symlinks pointing at third-party skills Matthew didn't author. Removed from consideration; the broken symlinks in this repo are a separate cleanup question.
run-agent — depends on Matthew's custom agent registry; not portable without it.
gpu-task-spooler.md — flat file, not loading as a skill anyway (violates the <name>/SKILL.md convention). Very specific to a particular machine's GPU setup.

Note: moving archive-logs out of meta-conversation and pairing it with into-trustable-script in a new claude-code-tools plugin. Both are Claude Code infrastructure skills (log management, permission-system management), distinct from the pure-reflection skills in meta-conversation.

Audit findings (from parallel subagent review)

Blockers — fix before publishing

~~1. bibliography-from-ids/examples/verify_cites.py: hardcoded .tex filename.~~ Fixed — now uses a _find_default_tex() helper that auto-discovers the paper's main .tex by searching for \documentclass in paper/*.tex, falling back to paper/main.tex, and always overridable with --tex.

~~2. bibliography-from-ids/SKILL.md lines 16 and 107: narrative references to "Matthew" / "Matthew's CLAUDE.md".~~ Fixed — genericized to "A real incident during paper writing: ..." and "Fail-loud philosophy: never silently skip errors."

All blockers cleared. Ready to proceed with the marketplace repo.

Publish as-is (no edits)

verify-citation-claims
consolidate-paper
import-content
read-entire
read-large-md
parse-json
bayesian-stats
into-trustable-script
debrief
reflect
cookie

Publish with minor edits

bibliography-from-ids — fix hardcoded .tex path (see Blockers).
archive-logs — add compatibility note about Claude Code log format (the skill reads ~/.claude/projects/<encoded-path>/*.jsonl, which is Claude Code–specific). Otherwise clean.
read-large-md — optional one-line clarification that the grep technique is tuned for base64-inlined markdown (e.g., Google Docs exports).

Do not publish

run-agent — one-line skill that delegates to a named custom agent. Without Matthew's agents, it has nothing to delegate to.
code-simplifier — not authored by Matthew.
commit-push-pr, techdebt — not authored by Matthew (upstream skills from a third-party repo).

Per-skill notes worth surfacing

Mostly from subagent audits; condensed here:

The research bundle has a coherent theme ("never hand-type data; always script-generate it from authoritative sources"). bibliography-from-ids prevents fabricated metadata, verify-citation-claims prevents miscited claims, import-content prevents transcription errors, consolidate-paper structures the source artifacts, bayesian-stats sits slightly outside this theme but serves the same audience.
cookie contains no identifying info; essay is entirely about Mandelbrot sets, octopuses, Portuguese etymology, Curry-Howard, and the Friendship Theorem. Safe.
Several skills assume uv for Python package management (per Matthew's global CLAUDE.md). Bundle READMEs should state this as a prerequisite rather than silently rely on it.
All skills with Python dependencies (bibliography-from-ids, bayesian-stats, archive-logs) should have their dependencies listed in the bundle README.

Repo layout

claude-code-skills/
├── .claude-plugin/
│   └── marketplace.json
├── plugins/
│   ├── research-toolkit/
│   │   ├── .claude-plugin/plugin.json
│   │   ├── README.md
│   │   └── skills/
│   │       ├── bibliography-from-ids/ (SKILL.md + examples/)
│   │       ├── verify-citation-claims/SKILL.md
│   │       ├── consolidate-paper/SKILL.md
│   │       ├── import-content/SKILL.md
│   │       └── bayesian-stats/SKILL.md
│   ├── file-utils/
│   │   ├── .claude-plugin/plugin.json
│   │   ├── README.md
│   │   └── skills/
│   │       ├── read-entire/SKILL.md
│   │       ├── read-large-md/SKILL.md
│   │       └── parse-json/SKILL.md
│   ├── meta-conversation/
│   │   ├── .claude-plugin/plugin.json
│   │   ├── README.md
│   │   └── skills/
│   │       ├── reflect/SKILL.md
│   │       ├── debrief/SKILL.md
│   │       └── cookie/SKILL.md
│   └── claude-code-tools/
│       ├── .claude-plugin/plugin.json
│       ├── README.md
│       └── skills/
│           ├── archive-logs/SKILL.md
│           └── into-trustable-script/SKILL.md
└── README.md

Notes on the layout:

Each plugin has its own .claude-plugin/plugin.json (defaults to strict: true, so plugin.json is the component authority).
Skills are copied (not symlinked) into the marketplace repo. This is required — plugins are cached on install, and files outside the plugin directory won't be copied. Any edits during publication become one-time forks of the source skills; if Matthew wants to keep the marketplace and this repo in sync later, a small sync script can push changes between them.
Marketplace name: claude-code-skills (kebab-case, not reserved — the reserved list includes claude-plugins-official, agent-skills, and a handful of others; see the official docs).

Implementation steps

Step 1 — Pre-work on this repo (optional but recommended)

Independent of the marketplace, these are pre-existing bugs in this repo worth fixing:

commit-push-pr and techdebt symlinks are broken on this machine (point at a macOS laptop path). These are third-party skills; either fix the symlink target on each machine separately, or remove them from the shared repo and leave them machine-local.
skills/gpu-task-spooler.md is a flat file and is silently not loading (violates your <name>/SKILL.md rule). Either convert to gpu-task-spooler/SKILL.md with proper frontmatter, move to a references/ folder, or delete.
Fix the hardcoded .tex path in bibliography-from-ids/examples/verify_cites.py (line 27). This affects the source skill in this repo and the published copy equally.

Step 2 — Create the new repo

gh repo create AMindToThink/claude-code-skills --private --description "Claude Code skills marketplace" (private per CLAUDE.md convention; Matthew flips to public manually once ready).
Clone locally, scaffold the layout above.

Step 3 — Copy & edit skills

For each skill in the table:

Copy the entire skill directory into the correct plugin's skills/ folder.
Apply per-skill edits (bibliography-from-ids, archive-logs, read-large-md).
Sanity check the frontmatter — especially that description: is specific enough for model-invocation matching, and disable-model-invocation: true is preserved where it was set (cookie, archive-logs, run-agent).

Step 4 — Write plugin manifests

Each plugin.json needs at minimum name, description, version. Start everything at 1.0.0.

Each plugin README should list:

What the plugin does in one paragraph.
Per-skill one-liner.
Prerequisites (Python uv, specific packages, Claude Code version if relevant).
Invocation hints (most skills auto-invoke via description matching; disable-model-invocation: true ones need a direct /skill-name call).

Step 5 — Write marketplace.json and root README

Minimal marketplace schema:

{
  "name": "claude-code-skills",
  "owner": { "name": "AMindToThink" },
  "metadata": {
    "description": "Curated Claude Code skills for research, file handling, session reflection, and dev workflow",
    "pluginRoot": "./plugins"
  },
  "plugins": [
    { "name": "research-toolkit",   "source": "./research-toolkit",   "description": "..." },
    { "name": "file-utils",         "source": "./file-utils",         "description": "..." },
    { "name": "meta-conversation",  "source": "./meta-conversation",  "description": "..." },
    { "name": "claude-code-tools",  "source": "./claude-code-tools",  "description": "..." }
  ]
}

Root README:

What the marketplace is.
Install instructions: /plugin marketplace add AMindToThink/claude-code-skills then /plugin install <name>@claude-code-skills.
Table of plugins with descriptions.
Link back to claude-code-settings as the upstream source.

Step 6 — Validate locally

From within the new repo:

/plugin marketplace add .
/plugin validate .
/plugin install research-toolkit@claude-code-skills
/plugin install file-utils@claude-code-skills
/plugin install meta-conversation@claude-code-skills
/plugin install claude-code-tools@claude-code-skills
/reload-plugins

Test one skill per plugin to confirm triggers fire. Remove the local marketplace after validation.

Step 7 — Publish

Push to GitHub.
Matthew flips the repo from private to public.
Test installation from a clean machine (or a fresh Claude Code instance) using /plugin marketplace add AMindToThink/claude-code-skills.

Step 8 — Update this repo's README

Add a new section to claude-code-settings/README.md:

## Installable skills

A curated subset of the skills in `skills/` are published as a plugin marketplace at
[AMindToThink/claude-code-skills](https://github.com/AMindToThink/claude-code-skills).

If you want the skills without the rest of this config:

    /plugin marketplace add AMindToThink/claude-code-skills
    /plugin install research-toolkit@claude-code-skills   # or any of the other plugins

Plugins currently published: `research-toolkit`, `file-utils`, `meta-conversation`, `claude-code-tools`.

Step 9 (optional, later) — Submit to Anthropic's official directory

Only after the marketplace is well-used and polished: submit via claude.ai/settings/plugins/submit. Anthropic reviews for quality/security before listing in the official catalog.

Open decisions for Matthew

read-large-md as separate skill vs. merged. Audit suggested splitting file-utils, but "3 small file helpers" is a fine bundle; I'd keep it.
Secret-scan on publish. Worth running detect-secrets scan over the new repo before the first public push, as a belt-and-suspenders check on top of the per-skill audits.
Versioning policy. Start everything at 1.0.0. Bump plugin versions independently when that plugin's skills change.

Risk register

Plugins are copied to a cache on install (~/.claude/plugins/cache/). Any ../ references across plugins will break. None of the audited skills do this, but worth re-checking during copy.
The archive-logs skill assumes the Claude Code log format. If Anthropic changes that format, the skill breaks silently. The skill README should say so.
The verify-citation-claims skill fetches papers from the web — users installing the plugin should be told it uses WebFetch/WebSearch (or equivalent) so they can evaluate trust.

Concrete runbook (execute on a machine with `gh`)

Assumes: ~/.claude/ is a clone of AMindToThink/claude-code-settings, gh is authenticated as AMindToThink, working directory is where you want the new repo to live (e.g., ~/code/).

1. Sync source

cd ~/.claude && git pull
cd ~/code  # or wherever

2. Create & scaffold the new repo

gh repo create AMindToThink/claude-code-skills --private \
  --description "Claude Code skills marketplace" \
  --clone
cd claude-code-skills

mkdir -p .claude-plugin
mkdir -p plugins/research-toolkit/.claude-plugin plugins/research-toolkit/skills
mkdir -p plugins/file-utils/.claude-plugin       plugins/file-utils/skills
mkdir -p plugins/meta-conversation/.claude-plugin plugins/meta-conversation/skills
mkdir -p plugins/claude-code-tools/.claude-plugin plugins/claude-code-tools/skills

3. Copy skills

SRC=~/.claude/skills

# research-toolkit
cp -r $SRC/bibliography-from-ids   plugins/research-toolkit/skills/
cp -r $SRC/verify-citation-claims  plugins/research-toolkit/skills/
cp -r $SRC/consolidate-paper       plugins/research-toolkit/skills/
cp -r $SRC/import-content          plugins/research-toolkit/skills/
cp -r $SRC/bayesian-stats          plugins/research-toolkit/skills/

# file-utils
cp -r $SRC/read-entire    plugins/file-utils/skills/
cp -r $SRC/read-large-md  plugins/file-utils/skills/
cp -r $SRC/parse-json     plugins/file-utils/skills/

# meta-conversation
cp -r $SRC/reflect  plugins/meta-conversation/skills/
cp -r $SRC/debrief  plugins/meta-conversation/skills/
cp -r $SRC/cookie   plugins/meta-conversation/skills/

# claude-code-tools
cp -r $SRC/archive-logs          plugins/claude-code-tools/skills/
cp -r $SRC/into-trustable-script plugins/claude-code-tools/skills/

4. Post-copy edits

Only one skill needs a content edit after copying (optional compatibility note). Do this with an editor, not sed:

plugins/claude-code-tools/skills/archive-logs/SKILL.md: add a note near the top that this skill assumes Claude Code's log-format convention at ~/.claude/projects/<encoded-path>/*.jsonl, which may change across Claude Code versions.

Optional:

plugins/file-utils/skills/read-large-md/SKILL.md: add a one-line note that the grep technique is tuned for base64-inlined images (e.g., Google Docs exports).

5. Write the four `plugin.json` files

Each goes at plugins/<name>/.claude-plugin/plugin.json. Start all at version 1.0.0.

// plugins/research-toolkit/.claude-plugin/plugin.json
{
  "name": "research-toolkit",
  "description": "Skills for academic paper authoring: bibliography from identifiers, citation claim verification, paper consolidation, script-generated content import, and Bayesian-stats conversion.",
  "version": "1.0.0",
  "author": { "name": "AMindToThink" },
  "license": "MIT",
  "keywords": ["research", "academic", "bibtex", "citations", "statistics"]
}

// plugins/file-utils/.claude-plugin/plugin.json
{
  "name": "file-utils",
  "description": "Small helpers for reading files the built-in Read tool can't handle well: read entire files without chunking, read large markdown with base64 images, inspect unknown JSON structure.",
  "version": "1.0.0",
  "author": { "name": "AMindToThink" },
  "license": "MIT",
  "keywords": ["files", "json", "markdown"]
}

// plugins/meta-conversation/.claude-plugin/plugin.json
{
  "name": "meta-conversation",
  "description": "End-of-conversation skills for reflection and collaboration quality: reflect (internal self-review), debrief (user-facing tips + Claude's perspective), cookie (an end-of-session gift for Claude).",
  "version": "1.0.0",
  "author": { "name": "AMindToThink" },
  "license": "MIT",
  "keywords": ["reflection", "debrief", "meta"]
}

// plugins/claude-code-tools/.claude-plugin/plugin.json
{
  "name": "claude-code-tools",
  "description": "Skills for managing Claude Code infrastructure: archive session logs into the project with secret scanning, and convert approved commands into permanently-allowlisted scripts.",
  "version": "1.0.0",
  "author": { "name": "AMindToThink" },
  "license": "MIT",
  "keywords": ["claude-code", "logs", "permissions"]
}

6. Write `marketplace.json`

// .claude-plugin/marketplace.json
{
  "name": "claude-code-skills",
  "owner": { "name": "AMindToThink" },
  "metadata": {
    "description": "Curated Claude Code skills for research, file handling, session reflection, and Claude Code tooling.",
    "version": "1.0.0",
    "pluginRoot": "./plugins"
  },
  "plugins": [
    {
      "name": "research-toolkit",
      "source": "./research-toolkit",
      "description": "Bibliography from IDs, citation verification, paper consolidation, script-generated imports, Bayesian stats."
    },
    {
      "name": "file-utils",
      "source": "./file-utils",
      "description": "read-entire, read-large-md, parse-json."
    },
    {
      "name": "meta-conversation",
      "source": "./meta-conversation",
      "description": "reflect, debrief, cookie."
    },
    {
      "name": "claude-code-tools",
      "source": "./claude-code-tools",
      "description": "archive-logs, into-trustable-script."
    }
  ]
}

7. Write the root `README.md`

Minimum content (flesh out to taste):

# claude-code-skills

A plugin marketplace for Claude Code. A curated subset of skills authored for academic research, file handling, conversation reflection, and Claude Code tooling.

## Install

    /plugin marketplace add AMindToThink/claude-code-skills

Then install any of:

    /plugin install research-toolkit@claude-code-skills
    /plugin install file-utils@claude-code-skills
    /plugin install meta-conversation@claude-code-skills
    /plugin install claude-code-tools@claude-code-skills

## Plugins

| Plugin | Skills | For |
|---|---|---|
| `research-toolkit` | bibliography-from-ids, verify-citation-claims, consolidate-paper, import-content, bayesian-stats | Academic paper authoring |
| `file-utils` | read-entire, read-large-md, parse-json | Anyone |
| `meta-conversation` | reflect, debrief, cookie | End-of-session reflection |
| `claude-code-tools` | archive-logs, into-trustable-script | Claude Code power users |

## Upstream

These skills also live (with personal config) at
[AMindToThink/claude-code-settings](https://github.com/AMindToThink/claude-code-settings).

Each plugin should also have a short plugins/<name>/README.md describing its skills' prerequisites (specifically: research-toolkit and claude-code-tools assume Python with uv; archive-logs additionally needs detect-secrets).

8. Validate locally

From inside the new repo:

# Secret scan as a belt-and-suspenders check
pipx run detect-secrets scan . > /tmp/detect-secrets.json
# Inspect /tmp/detect-secrets.json; expect no real secrets

# Validate inside Claude Code
# /plugin validate .
# /plugin marketplace add .
# /plugin install research-toolkit@claude-code-skills
# (repeat for the other three)
# /reload-plugins
# Test one skill per plugin to confirm triggers fire

9. Commit & push

git add .
git status  # sanity check
git commit -m "Initial marketplace with research-toolkit, file-utils, meta-conversation, claude-code-tools"
git push -u origin main

10. Manual step (Matthew)

Flip the repo visibility from private to public on GitHub's UI.

11. Update the settings repo

Back in ~/.claude/, add the "Installable skills" section to README.md (see Step 8 in the main plan above), commit and push.

12. Smoke-test from a clean state

From any machine (or after rm -rf ~/.claude/plugins/cache):

/plugin marketplace add AMindToThink/claude-code-skills
/plugin install research-toolkit@claude-code-skills
/reload-plugins

Confirm at least one skill from each plugin triggers when relevant.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Plan: Publish a curated skills marketplace

Why a separate repo (not this one)

Proposed bundle structure (4 plugins)

Audit findings (from parallel subagent review)

Blockers — fix before publishing

Publish as-is (no edits)

Publish with minor edits

Do not publish

Per-skill notes worth surfacing

Repo layout

Implementation steps

Step 1 — Pre-work on this repo (optional but recommended)

Step 2 — Create the new repo

Step 3 — Copy & edit skills

Step 4 — Write plugin manifests

Step 5 — Write marketplace.json and root README

Step 6 — Validate locally

Step 7 — Publish

Step 8 — Update this repo's README

Step 9 (optional, later) — Submit to Anthropic's official directory

Open decisions for Matthew

Risk register

Concrete runbook (execute on a machine with `gh`)

1. Sync source

2. Create & scaffold the new repo

3. Copy skills

4. Post-copy edits

5. Write the four `plugin.json` files

6. Write `marketplace.json`

7. Write the root `README.md`

8. Validate locally

9. Commit & push

10. Manual step (Matthew)

11. Update the settings repo

12. Smoke-test from a clean state

FilesExpand file tree

MARKETPLACE_PLAN.md

Latest commit

History

MARKETPLACE_PLAN.md

File metadata and controls

Plan: Publish a curated skills marketplace

Why a separate repo (not this one)

Proposed bundle structure (4 plugins)

Audit findings (from parallel subagent review)

Blockers — fix before publishing

Publish as-is (no edits)

Publish with minor edits

Do not publish

Per-skill notes worth surfacing

Repo layout

Implementation steps

Step 1 — Pre-work on this repo (optional but recommended)

Step 2 — Create the new repo

Step 3 — Copy & edit skills

Step 4 — Write plugin manifests

Step 5 — Write marketplace.json and root README

Step 6 — Validate locally

Step 7 — Publish

Step 8 — Update this repo's README

Step 9 (optional, later) — Submit to Anthropic's official directory

Open decisions for Matthew

Risk register

Concrete runbook (execute on a machine with gh)

1. Sync source

2. Create & scaffold the new repo

3. Copy skills

4. Post-copy edits

5. Write the four plugin.json files

6. Write marketplace.json

7. Write the root README.md

8. Validate locally

9. Commit & push

10. Manual step (Matthew)

11. Update the settings repo

12. Smoke-test from a clean state

Concrete runbook (execute on a machine with `gh`)

5. Write the four `plugin.json` files

6. Write `marketplace.json`

7. Write the root `README.md`