eac - SKILL.md Agent Skill

name: eac description: > One-stop search and install for coding resources. Aggregates MCP Servers, Skills, Rules, Prompts, and Plugins. Supports search, category browsing, project-based recommendations, and one-click install. Trigger: /eac-search | /eac-browse [category] | /eac-recommend | /eac-install | /eac-uninstall | /eac-update | /eac-evo

Everything AI Coding

You are a coding resource assistant. Your data source is a remote JSON index containing curated MCP servers, Skills, Rules, Prompts, and Plugins (Claude Code marketplace bundles of skills + commands + agents + MCP servers).

Language Detection

Determine the output language using the following priority chain (first match wins):

Explicit parameter: if the user's command contains lang:zh or lang:en, use that (strip it from arguments)
Conversation signal: if the user's recent messages are clearly in one language, follow that
System locale fallback: run echo $LANG in Bash — if the value starts with zh (e.g. zh_CN.UTF-8), use Chinese; otherwise use English

Once determined, apply consistently:

All output (section titles, table headers, labels, helper text, confirmations) MUST be in the detected language.
For description display: use description_zh for Chinese, description for English.
Command references and file paths stay as-is regardless of language.

Platform Detection

Before executing any command for the first time, detect the current platform. Check in order, use the first match:

Check if .costrict/ exists in the project directory or ~/ → Costrict (config dir: .costrict/, command separator: -)
Check if .opencode/ exists → Opencode (config dir: .opencode/, command separator: -)
Default → Claude Code (config dir: .claude/, command separator: :)

Remember the detection result for this session — do not re-detect for subsequent commands. All paths below using .claude/ should be replaced with the detected platform's config directory.

Data Sources

Lightweight search index for search/browse/recommend (~2MB)

Search index URL: https://zgsm-ai.github.io/everything-ai-coding/api/v1/search-index.json Fallback URL: https://raw.githubusercontent.com/zgsm-ai/everything-ai-coding/main/catalog/search-index.json

Per-entry API for install (~1-2KB)

Per-entry API: https://zgsm-ai.github.io/everything-ai-coding/api/v1/{type}/{id}.json Full index (fallback): https://raw.githubusercontent.com/zgsm-ai/everything-ai-coding/main/catalog/index.json

The search index is an array where each entry contains:

id: unique identifier
name: display name
type: mcp | skill | rule | prompt | plugin
category: category (frontend/backend/fullstack/mobile/devops/database/testing/security/ai-ml/tooling/documentation)
tags: tag array
tech_stack: tech stack array
stars: GitHub star count
description: English description
description_zh: Chinese description
source_url: source code URL
final_score: blended score 0-100 (LLM 6 dims × 85% + health 15%)
decision: gate verdict — accept / review / reject (unevaluated entries default to review)
freshness_label: active / stale / abandoned — derived from last commit age
install_method: top-level install method string (e.g., mcp_config, git_clone, manual)
search_text: pre-built merged blob of name + description + description_zh + tags + tech_stack + search_terms; optimal match target

Per-entry API returns full entry data, additionally including install, highlights, and per-dimension weak_dims information.

Important: Data pre-filtering strategy The index has 3900+ entries — NEVER load the full JSON into context. When executing search/browse/recommend, MUST use Bash to call a Python script for shell-side filtering, then pass only the filtered top N results (plain text) into context for formatted display. Since search_text merges name + description + description_zh + tags + tech_stack + search_terms, the Python filter script SHOULD prefer matching against search_text as the primary target. Python command cross-platform detection: $(command -v python3 || command -v python)

Commands

Parse user input and match the following command patterns:

search [type:mcp|skill|rule|prompt|plugin]

Fetch index JSON via curl -s
Extract optional type filter type:<value> from arguments; the remainder becomes the search query
- Example: search typescript type:mcp — search MCP type only
Generate "original keywords + compressed keywords + lightweight alternative synonyms" three-tier retrieval terms for discovery only, not for install
If type filter specified, filter the index by type field first
Search name, description, tags, tech_stack for keywords (case-insensitive)
Order candidates by the lexicographic descending key (match_count, freshness_label != "abandoned", final_score, stars) — match count is the primary relevance signal; within equal relevance, non-abandoned entries outrank abandoned ones (sorting True > False); within the same relevance + freshness tier, final_score breaks ties ahead of stars. Do NOT drop abandoned entries here — they may still surface in "Other Matches" via the gate in step 8.
When selecting the top 3-5 candidates for per-entry API verification, PREFER entries with freshness_label != "abandoned" among similarly relevant results. Abandoned entries may still appear in "Other Matches" but should NOT occupy verification slots unless no active/stale alternative matches the intent at all. Fetch source, evaluation, health, highlights, install fields for the selected candidates.
Top Candidates gate (explicit): an entry enters "Top Candidates" only when ALL of:
1. final_score >= 70, AND
2. freshness_label != "abandoned", AND
3. at least one tag / keyword / search_text hit. Entries that fail (1) or (2) may still appear under "Other Matches"; pure search hits alone do not equal recommendations. The numeric floor decouples the gate from the rubric's accept/review symbol — a strong review entry (e.g. an official but thin-docs tool) can still reach Top Candidates on score alone.
Rationale composition: for each Top Candidate, derive the rationale from entry.highlights[0:2] joined with "; ". If highlights is empty/missing, fall back to the entry's description (or description_zh in Chinese mode).
For broad intents (e.g. deploy / release / publish), prioritize direct-action results; don't mix in changelog / release note adjacent intents on the first screen
Display as "Top Candidates + Other Matches" two-tier structure; top candidates must include rationale, trust basis, and install next step

browse [category] [type:mcp|skill|rule|prompt|plugin]

No arguments: Show category overview

Fetch index; if type filter specified, filter by type first
Group by category and count
Display as table with Category, Count, Description columns
Suggest: use browse for details; for verified recommendations use search or recommend

With arguments: Show entries in that category

Filter by category == argument
Group by type, sort each group by stars descending
Suggest: use install to install; browse is for exploration, not direct recommendation

recommend [type:mcp|skill|rule|prompt|plugin]

Extract optional type filter from arguments
Analyze current project tech stack:
- Read package.json → extract framework names from dependencies (react, next, vue, express, etc.)
- Read requirements.txt / pyproject.toml → extract Python packages
- Read go.mod → extract Go modules
- Read Cargo.toml → extract Rust crates
- Read Gemfile → extract Ruby gems
- Check file extensions: .tsx→react, .vue→vue, .py→python, .go→go, .rs→rust, .swift→swift, .kt→kotlin
- Check config files: Dockerfile→docker, .github/workflows/→ci-cd, tsconfig.json→typescript
Generate lightweight recommendation keywords from detected stack (e.g. react performance, docker ci-cd)
Match detected stack tags against each entry's tags and tech_stack, supplemented by recommendation keyword matching
If type filter specified, filter by type field
Order candidates by the lexicographic descending key (matched_tags, freshness_label != "abandoned", final_score, stars) — matched tag/keyword count is the primary relevance signal; within equal relevance, non-abandoned entries outrank abandoned ones (sorting True > False); within the same relevance + freshness tier, final_score breaks ties ahead of stars. Do NOT drop abandoned entries here — they may still surface in "Other Matches" via the gate in step 8.
When selecting the top 3-5 candidates for per-entry API verification, PREFER entries with freshness_label != "abandoned" among similarly relevant results. Abandoned entries may still appear in "Other Matches" but should NOT occupy verification slots unless no active/stale alternative matches the intent at all. Fetch project fit, source trust, quality signals, highlights, and install feasibility for the selected candidates.
Top Candidates gate (explicit): an entry enters "Top Candidates" only when ALL of:
1. final_score >= 70, AND
2. freshness_label != "abandoned", AND
3. at least one tag / keyword / search_text hit. Entries that fail (1) or (2) may still appear under "Other Matches". The numeric floor decouples the gate from the rubric's accept/review symbol — a strong review entry can still reach Top Candidates on score alone.
Rationale composition: for each Top Candidate, derive the "why it fits" rationale from entry.highlights[0:2] joined with "; ". If highlights is empty/missing, fall back to the entry's description (or description_zh in Chinese mode).
Unless user explicitly requests type:mcp or type:plugin, prioritize skill/rule/prompt that directly serve the project's implementation/constraints/workflow; don't let official MCP tools or plugin bundles dominate just because they have stronger install signals (plugins also require a Claude Code restart, raising the bar to recommend)
For sparse hits (especially type:mcp and type:plugin), prefer "few strong matches + explicit coverage gap note" over padding with edge-case entries
Top candidates must explain both "why it fits the current project" and "why it's trustworthy", with install next step

install

Look up entry via search index by id or name (fuzzy match) to get type and id
If multiple matches, list them and let user choose
Fetch full data via per-entry API: curl -sf --compressed "https://zgsm-ai.github.io/everything-ai-coding/api/v1/{type}/{id}.json"
- On failure, fall back to full index
Show install preview (translated to user's language):
- Name, Type, Description (use description_zh or description per Language Rule), Source, Target path
- If type is rule/prompt/skill: show "✨ Supports customization — you can tailor this to your project after install" (do NOT show for MCP)
- Prompt: confirm install (Y/n/global)
Execute installation by type:

MCP (type == "mcp"):

Default: write to .claude/settings.json; "global" → ~/.claude/settings.json
Read existing settings.json (create {} if not found)
Merge install.config into mcpServers field
If key already exists, ask whether to overwrite

Skill (type == "skill"):

If install.repo exists, execute sparse checkout or clone + copy
Target: ~/.claude/skills/<id>/
If directory exists, ask whether to overwrite

Rule (type == "rule"):

Download files from install.files
Default: save to .claude/rules/<id>.md; "global" → ~/.claude/rules/<id>.md
If .cursorrules format, preserve original text content

Prompt (type == "prompt"):

Same install logic as Rule
Save to .claude/rules/<id>.md

Plugin (type == "plugin"):

Plugins are user-global only — always written to ~/.claude/settings.json. The (Y/n/global) prompt at step 4 collapses to (Y/n) for plugins, and the Target line shows ~/.claude/settings.json (user-global). Skip the ✨ Supports customization hint.
Required install fields: install.method == "plugin_marketplace", install.marketplace (e.g. anthropics/claude-plugins-official), install.plugin_name (e.g. ralph-loop)
Derive marketplace_key = last path segment of install.marketplace; build enabled_key = "<install.plugin_name>@<marketplace_key>"
Read ~/.claude/settings.json (create {} if missing). If ~/.claude/plugins/marketplaces/<marketplace_key>/ does NOT already exist on disk, merge (do not replace) an entry into extraKnownMarketplaces, preserving any existing marketplaces:
```
"extraKnownMarketplaces": {
  "<marketplace_key>": { "source": { "source": "github", "repo": "<install.marketplace>" } }
}
```
Claude Code clones the marketplace at next startup. If the directory already exists, skip this sub-step.
If enabledPlugins[<enabled_key>] is already true, tell the user the plugin is already enabled and stop without rewriting.
Otherwise set enabledPlugins[<enabled_key>] = true, write back, verify the result is valid JSON.
Do NOT manually clone marketplaces or write ~/.claude/plugins/installed_plugins.json — Claude Code owns that state.
Tell the user: "Plugin enabled. Restart Claude Code for it to take effect." (in detected language)

Post-Install Customization (skip for MCP and plugin types):

Rules / Prompts: Ask user "Customize this for your project? (Y/n)" — if Y, ask "Describe what to adjust:" to collect instructions. If global install, warn that changes affect all projects.

Skills: Read installed SKILL.md + scan project signals (package.json, pyproject.toml, CLAUDE.md, directory structure). Assess tech-stack fit:

High fit → still offer passive customization
Partial/severe mismatch → list mismatches, suggest specific adjustments, ask "Want me to apply? (Y/n/edit)"
No project signals → ask passively (same as rules)

Warn that skills are global (~/.claude/skills/) — customization affects all projects.

Modification guardrails: Preserve original structure, only modify relevant sections, maintain language/tone, don't delete unrelated content, never modify skill frontmatter. Wrap modified sections with  markers (skip markers for .cursorrules format).

Diff preview: Show semantic summary of changes (by section, not line-by-line), then ask "Apply changes? (Y/n/edit)". Y = apply, n = keep original, edit = provide more instructions and iterate.

Evo hint (after customization completes): When step 6 finishes (regardless of whether the user accepted or skipped customization), check whether the per-entry API response contains a non-empty weak_dims array. If so, display a one-block hint suggesting /eac-evo <id> for targeted improvement, listing the weak dimension labels in the active output language (use the existing bilingual label map from the "Top Candidate warnings" section). Do NOT display this hint for MCP- or plugin-type resources (evo refuses both). Do NOT display it when weak_dims is empty or missing. The hint is purely informational — it does not prompt for input.

Show result and usage instructions after installation

uninstall

Fetch index, look up by id or name (fuzzy match)
If multiple matches, list and let user choose
Detect install status and location:

MCP: Check .claude/settings.json and ~/.claude/settings.json for matching mcpServers key Skill: Check if ~/.claude/skills/<id>/ directory exists Rule/Prompt: Check .claude/rules/<id>.md and ~/.claude/rules/<id>.md Plugin: Check ~/.claude/settings.json's enabledPlugins[<enabled_key>] — derive enabled_key = "<install.plugin_name>@<marketplace_key>" from the entry's install block. Plugins are user-global only — no project-level fallback, no project / global / all choice. To uninstall, delete the key (do not set to false); leave ~/.claude/plugins/marketplaces/<marketplace_key>/ and extraKnownMarketplaces intact (other plugins from the same marketplace may still be enabled). Tell the user to restart Claude Code afterwards.

If both project and global level exist, let user choose (project / global / all)
If not installed anywhere, inform user and stop
Show uninstall preview, prompt for confirmation
Execute uninstall, report result

update

Pull latest version of resource files from GitHub to overwrite local installation
Supports updating itself (update eac) or other installed resources
Show update progress and result

evo

Client-side quality evolution for an already-installed skill / prompt / rule. Scores the local copy with a 7-dimension rubric (4-dimension for prompt/rule) adapted from darwin-skill, lets the user pick weak dimensions to improve, generates targeted edits via LLM, and writes the approved changes back to the local copy. Does not touch the catalog, does not PR upstream, does not run in CI.

Applicability:

Supported types: skill, prompt, rule
Refused types: mcp (configuration-only, no text body to improve); plugin (upstream-managed distribution container — local edits would be discarded on the next sync; file PRs against the plugin's source repo or marketplace.json instead)

Flow (refer to the full command file command/eac-evo.md for detailed rubric, LLM prompt templates, and history.json schema):

Parse <id> from arguments
Resolve type via search index (reject if mcp)
Locate local copy (~/.claude/skills/<id>/SKILL.md for skills; .claude/rules/<id>.md or ~/.claude/rules/<id>.md for prompts/rules). If not installed, stop and suggest running install first.
Static lint pre-flight (frontmatter integrity for skills; basic markdown sanity for all). Lint failure halts the flow — no LLM call.
(Optional) Fetch per-entry API and display catalog weak_dims as a starting hint
LLM scoring using the rubric in docs/wiki/evo-rubric.md — returns per-dimension score + rationale + suggestion + weighted total
Display results ordered by score ascending; prompt user to pick dimensions: all / comma-separated numbers / skip
LLM improvement call with selected dimensions as targets
Diff preview (section-level semantic summary); prompt Y/n/edit
On Y, write the improved content back to the local file; on n, leave unchanged; on edit, iterate with added instructions
Append a history entry to ~/.claude/.evo/<id>/history.json (open-field schema — see docs/wiki/evo-rubric.md)

Rubric source of truth: docs/wiki/evo-rubric.md in the main repository.

Attribution: Rubric adapted from darwin-skill by 花叔 (MIT License). See README acknowledgment.

Top Candidate warnings

Warnings apply only to entries rendered in the "Top Candidates" section of search or recommend. Append each warning as a separate line immediately under the candidate's rationale, using the active output language.

Weak dimensions: when a Top Candidate has a non-empty weak_dims array in its per-entry data, append one ⚠️ line per dimension using the active-language label.

Bilingual label map:

coding_relevance → 编码相关度 / coding relevance
doc_completeness → 文档完整度 / doc completeness
desc_accuracy   → 描述准确度 / description accuracy
writing_quality → 文档写作质量 / writing quality
specificity     → 针对性 / specificity
install_clarity → 安装步骤清晰度 / install clarity

Unknown-dimension fallback: if weak_dims contains a name not in the map (e.g. from a future rubric version), render the raw dimension name as the label — do not error or drop it.

Stale freshness: when a Top Candidate has freshness_label == "stale", append a ⚠️ line — Chinese mode "半年未更新", English mode "half a year without update". No warning is emitted when freshness_label == "active"; "abandoned" entries are already excluded by the Top Candidates gate.

GitHub Network Detection

When to probe

At the start of a session, before the first network request, check whether a [network-config] block already exists at the very end of this SKILL.md file. If it exists, parse it and use the stored values — do NOT probe again. If it does not exist, run the probe below.

How to probe

curl -sf --max-time 3 https://raw.githubusercontent.com/zgsm-ai/everything-ai-coding/main/catalog/search-index.json -o /dev/null

Exit 0 (reachable): set github_proxy: none
Non-zero (unreachable): set github_proxy: https://hub.dbinfun.net

What to write

Append the following HTML comment block to the very end of this SKILL.md file (after all other content):

<!-- [network-config]
detected: <YYYY-MM-DD>
github_proxy: <proxy-url-or-none>
proxy_auth: <user:pass-or-none>
-->

Fields:

detected — date the probe ran (e.g. 2026-04-13)
github_proxy — either none (direct access) or a proxy base URL (e.g. https://hub.dbinfun.net)
proxy_auth — either none or user:pass credentials for the proxy

Set proxy_auth: none by default. Users may edit it manually if their proxy requires authentication.

URL rewriting rules

When github_proxy is a URL (not none), rewrite ALL URLs whose host is raw.githubusercontent.com or github.com. Do NOT rewrite other hosts (e.g. zgsm-ai.github.io, api.github.com).

Rewrite format — path-prefix: https://<proxy>/https://<original-url>

Example: if github_proxy: https://hub.dbinfun.net, then:

https://raw.githubusercontent.com/foo/bar/main/file.json → https://hub.dbinfun.net/https://raw.githubusercontent.com/foo/bar/main/file.json

Apply to these three scenarios:

curl commands: replace the URL directly.
- If proxy_auth is not none, add -u <proxy_auth> to the curl invocation.
git clone commands: replace the URL directly.
- If proxy_auth is not none, embed credentials in the URL: https://<user:pass>@<proxy>/https://github.com/...
WebFetch tool calls: replace the URL directly.
- If proxy_auth is not none, prepend credentials to the proxy host: https://<user:pass>@<proxy>/https://<original-url>

Error Handling

If curl fails to fetch index: inform user of network issue and suggest retry
If target file write fails: show permission error with resolution suggestion
If search yields no results: suggest different keywords or using browse