ecosystem-update

A Claude Code skill that runs daily, scans the sources you care about for new patterns, diffs them against your current setup, implements the quick wins, and writes a dated report.

Originally built to keep my personal Claude Code config current with the broader community ecosystem. Now generic — point it at any set of URLs and it'll do the same job.

What it does

1. Reads your current state — config files, installed agents, skills, MCP servers (paths configurable)
2. Fetches sources — URLs from sources.yaml, organized into tiers (daily / weekly)
3. Extracts candidates — discrete items it found that you might want to add
4. Diffs against current state — buckets each candidate as HAVE, PARTIAL, MISSING, or CONFLICTS
5. Scores remaining items — Impact / Effort / Alignment-with-philosophy
6. Writes a dated markdown report to ~/.claude/reports/ecosystem/YYYY-MM-DD.md
7. Implements quick wins automatically — high-impact, low-effort, philosophy-aligned changes (with backups + hard limits, configurable)
8. Saves dedup state so it never reports the same item twice

The dedup state is the key thing — it's why running this daily doesn't drown you in noise.

Why you'd use it

You're building something on top of a fast-moving ecosystem (Claude Code, agentic frameworks, web frameworks, ML tooling, security advisories — anything where the community ships new patterns weekly). You want to stay current without spending an hour every day reading.

Cron this once. Read a 2-minute report each morning. The skill does the rest.

Install

git clone https://github.com/Chaddacus/ecosystem-update
cd ecosystem-update
./install.sh

install.sh drops the skill into ~/.claude/skills/ecosystem-update/, scaffolds sources.yaml and config.yaml from the example templates on first run, and creates the state / reports / backups / logs directories.

Then edit:

• ~/.claude/skills/ecosystem-update/sources.yaml — the URLs you want to track
• ~/.claude/skills/ecosystem-update/config.yaml — adjust paths if your Claude home isn't ~/.claude

Then in Claude Code:

/ecosystem-update

For a dry run (report only, no auto-implement):

/ecosystem-update --dry-run

Alternate: dev install (symlink the repo)

If you want to hack on the skill body itself, use --link instead — that symlinks SKILL.md and the example yamls back to the repo so your edits go straight to the source:

./install.sh --link

Updating

cd <wherever you cloned it>
git pull
./install.sh   # idempotent; preserves your sources.yaml and config.yaml

Customize

sources.yaml

The URLs you want scanned, organized into tiers. Tiers control how often each source is fetched:

tier_1_daily:
  - url: https://github.com/some-org/awesome-thing
    extract: "new patterns, new modules added to the catalog"
  - url: https://creatorblog.example.com/
    extract: "creator's latest tips and patterns"

tier_2_daily_optional:  # skipped if last_run < 24h ago
  - url: https://arxiv.org/search/?query=relevant+topic&order=-announced_date_first
    extract: "papers from the last 24h with applicable patterns"

tier_3_weekly:           # skipped if last_run < 7d ago
  - url: https://some-comprehensive-catalog.example/
    extract: "comprehensive catalog updates"

Pre-built example configs in examples/:

• ai-engineering.yaml — Claude Code, MCP, agent patterns, AI eng research
• security-advisories.yaml — CVEs, security blogs, advisories
• web-framework.yaml — frontend framework releases, RFC discussions

Copy any of these to sources.yaml to get started.

config.yaml

Paths, scoring weights, and hard limits:

paths:
  state_file: ~/.claude/state/ecosystem-update-last-run.json
  report_dir: ~/.claude/reports/ecosystem/
  backup_dir: ~/.claude/backups/

scoring:
  quick_win_threshold: 2.0   # Priority = Impact / Effort
  build_queue_threshold: 1.0

implement_quick_wins: true   # set false for report-only behavior

hard_limits:
  never_touch:
    - ~/.claude/CLAUDE.md     # constitutional policy doc — never auto-edit
  forbid_new_files: true      # auto-implement never creates new files
  forbid_body_rewrites: true  # only frontmatter additions to existing files

philosophy_gate: true         # require one-sentence justification per recommendation

Philosophy gate

Every recommendation has to pass a one-sentence test:

"Can I prove in one sentence that an existing primitive cannot satisfy this requirement?"

If the answer is "no" or needs more than one sentence — rejected as overengineering. This is how the skill stays useful instead of recommending shiny things you'd never need. Disable in config.yaml if you want — but it's the heart of why this tool doesn't waste your time.

How the dedup actually works

After every run, state_file (default ~/.claude/state/ecosystem-update-last-run.json) gets a seen_items array of slugified item titles. On the next run, candidates whose slug matches the list are silently bucketed as HAVE and skipped.

This is what makes daily runs sustainable — without it, you'd see the same items every time the catalog source updated.

Schedule it

Cron at 8am daily:

0 8 * * * cd ~/.claude && claude --print "/ecosystem-update" >> ~/.claude/logs/ecosystem-update.log 2>&1

Or use the Claude Code scheduler if you have one wired:

/schedule daily 8am /ecosystem-update

The skill is headless-safe: no interactive prompts, deterministic exit, idempotent on same-day reruns.

What I run it on

I built this for my personal Claude Code config. It's been running daily for 47+ consecutive days as of writing, zero misses. Every morning I get a 1-page markdown report at ~/.claude/reports/ecosystem/YYYY-MM-DD.md with quick wins already applied to my config and a build queue of bigger items I might pick up.

That's why this is generic now — the loop works, the philosophy gate keeps the noise down, and there's no reason it should only watch Claude Code sources.

License

MIT.

Credit

Concept and original implementation by Chad Simon. Open source so you can adapt it to whatever ecosystem you live in.