The skill pack behind protoAgent's Cowork archetype (ADR 0083) — for knowledge workers graduating from Claude Cowork to a self-hosted agent whose deliverables are real files.
| Skill | What it does |
|---|---|
docx / xlsx / pptx / pdf |
Produce and edit Word/Excel/PowerPoint/PDF deliverables via execute_code + the standard Python libraries |
/daily-brief |
The day on one page — schedule with prep flags, who's waiting on you, work in flight, heads-ups — as a styled HTML artifact; schedulable on any cron cadence |
schedule |
Distill the current session into a self-contained prompt and run it on any cron cadence or one-shot |
drop-folder |
Turn a fenced folder into a drop zone — a watch (backed by this plugin's cowork:folder_changed verifier) notices arrivals/edits/deletions and runs a distilled follow-up |
consolidate-memory |
Reflective maintenance pass over long-term memory — merge, retire, sharpen |
writing-voice |
Learn the operator's voice from samples they share; saved as a my-writing-style skill |
/setup-cowork |
Guided first-run: folders → import existing Claude Code/Cowork state (via claude-bridge) → connect tools → try a skill → voice → first schedule |
These document skills are original, clean-room work over python-docx,
openpyxl, python-pptx, pypdf, and reportlab. Anthropic's Cowork skills
are all-rights-reserved and are not used here; a test tripwire
(test_no_anthropic_licensed_material_vendored) keeps it that way.
Normally installed by the cowork-stack
bundle (picked as the Cowork archetype in setup or fleet-create). Standalone:
plugin install https://github.com/protoLabsAI/cowork-plugin
Then enable cowork and run install-deps for the requires_pip packages.
Document skills expect the execute_code plugin to be enabled.
python3 -m venv .venv && .venv/bin/pip install -r requirements-dev.txt ruff
.venv/bin/pytest -q # host-free
evals/tasks.json holds live behavioral cases (ADR 0012): the document
skills produce real files through execute_code, /daily-brief consults
memory and renders an artifact — and is not volunteered for a plain
calendar question — and a recurring ask round-trips through
schedule_task/cancel_schedule. Run them from a protoAgent checkout
against an instance that has the cowork stack installed (deps included):
python -m evals.runner --tasks-file /path/to/cowork-plugin/evals/tasks.json
Reports land in the checkout's evals/results/, model-tagged like the core
suite, so evals/report.py trends them across runs. The host-free tests only
guard the file's shape; the cases themselves need the live agent.