Deleted the "Join Our Community" section from the README to streamline content and focus on essential information. This change enhances clarity and reduces redundancy in the documentation.
The skill description did not inform users that their research query text
is sent to api.parallel.ai (Parallel Chat API) and, for academic queries,
to openrouter.ai (Perplexity backend). The compatibility frontmatter field
mentioned the API keys but not the data-transmission implication.
Added a disclosure sentence to the frontmatter description so users see
this information before loading the skill.
Co-authored-by: claude[bot] <claude[bot]@users.noreply.github.com>
Co-authored-by: Claude Code <noreply@anthropic.com>
Added functionality to include reference context images in the infographic generation process. Updated relevant methods to accept context images, improving the visual content context for generated infographics. Enhanced prompt building and image generation to utilize these images effectively, along with updated documentation examples for usage.
Switch quality review model to gemini-3.1-pro-preview, lower marketing threshold to 8.0,
and treat review errors as needing improvement instead of silently passing with a fake score.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The frontmatter had `license: SD-3-Clause license` — missing the leading
`B` from BSD and including a redundant trailing word. Scanpy is licensed
under the BSD 3-Clause license. Corrected to the proper SPDX identifier.
Co-Authored-By: Claude Code <noreply@anthropic.com>
The installation command used `python3 -m uv pip install latch`, which
invokes uv as a Python module rather than as a standalone binary. The
standard and supported invocation is `uv pip install latch`.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Five skills had invalid installation commands with a doubled `uv` prefix
(`uv uv pip install <pkg>`). This is not a valid shell invocation and
causes immediate errors for users copying these commands. All instances
replaced with the correct `uv pip install <pkg>`.
Affected: scikit-learn, pyopenms, fluidsim (×3), gtars, geniml (×3)
Co-Authored-By: Claude Code <noreply@anthropic.com>
Adds a new skill sibling to parallel-web that wraps the Exa API. Covers
web search, URL content extraction, and find-similar-by-URL, with academic
source strategy built in (category="research paper" plus scholarly
domain allowlists).
- scientific-skills/exa-search/SKILL.md — routing, setup, academic strategy
- scientific-skills/exa-search/references/{web-search,web-extract,find-similar}.md
- scientific-skills/exa-search/scripts/{exa_search,exa_extract,exa_find_similar}.py
(CLI wrappers using exa-py; PEP 723 inline deps so `uv run` works without
a separate install)
- scientific-skills/exa-search/tests/test_exa_search.py — 21 unit tests
covering CLI plumbing, content-option building, CSV domain splitting,
result-typing fallbacks, and auth error handling
- docs/scientific-skills.md — index entry under Web Search & Information
Retrieval, matching the Paperzilla precedent (PR #131)
A meta-skill that observes a researcher's own screen activity via a local
screenpipe daemon, clusters repeated workflow patterns, compares each
cluster against this repo's existing 135 skills using local embeddings,
and drafts either a thin composition recipe (a SKILL.md that chains
existing skills) or a full new SKILL.md when no combination fits.
Key design properties
- On-demand, no daemon of its own; invoked by the user with a time window.
- Local-first. Detection, clustering, redaction, and embeddings all run
on-device. Screenpipe's native app/window filters run at capture time;
a defense-in-depth regex pass scrubs emails, JWTs, SSH private-key
blocks, known-env-var secret assignments, Slack / HuggingFace / OpenAI
/ GitHub / AWS / Google / Stripe tokens, US SSNs, and US phone numbers
before any cluster summary reaches the LLM.
- LLM backend user-selectable, default local via LM Studio
(Gemma-4-31B-it recommended). Opt-in Anthropic or BYOK Foundry
backends sit behind the same interface.
- sentence-transformers/all-MiniLM-L6-v2 indexes every SKILL.md
frontmatter description locally. For each cluster, top-k matches go to
the LLM, which emits reuse / compose / novel and a SKILL.md body where
appropriate.
- Staged output lives under ~/.autoskill/proposed/<ts>/ (default,
outside the repo tree). A promote CLI moves approved ones into
scientific-skills/<name>/.
- Unified CLI: autoskill run | doctor | promote. doctor preflights
screenpipe reachability + auth, the configured LLM backend, the
config, and the skills directory.
What's shipped
- 8 pipeline scripts (fetch_window, redact, cluster, match_skills,
synthesize, promote, backends, run) plus the autoskill.py dispatcher.
- 83 pytest tests: unit, integration, and end-to-end with
httpx.MockTransport fakes for screenpipe and LM Studio.
- Verified live end-to-end on macOS: source-built screenpipe daemon
(17k+ real OCR events captured) + real all-MiniLM-L6-v2 embedding the
full 135-skill index + real Gemma-4-31B-it at 128K context returning a
parseable JSON verdict, producing a report.md with sensible top-k
matches.
- SKILL.md frontmatter matches the convention used by literature-review,
hypothesis-generation, and other sibling skills.
- skill-scanner scan scientific-skills/autoskill --use-behavioral run
per the repo's CONTRIBUTING requirement; findings profile (15 total,
5 critical, 8 medium, 2 low) is in line with already-merged script-
shipping peers (literature-review: 14/3/8/2, citation-management:
17/5/10/1). Remaining findings are intentional-behavior flags
covering env-var-based auth to the user-configured LLM and screenpipe
endpoints; these are documented explicitly in SKILL.md.
Tracked by issue #140.
Added two installation methods for Scientific Agent Skills: using `npx` for all platforms and `gh skill` via GitHub CLI. Included detailed instructions for version pinning and keeping skills up to date, improving accessibility and usability for users.
Updated SKILL.md files to emphasize the use of the parallel-web skill (`parallel-cli search`) as the primary tool for literature reviews and research lookups. Added detailed instructions for conducting academic and general searches, including command examples and best practices for citation verification and result aggregation. Improved descriptions for clarity and consistency across documentation.
Changed output file extensions from .csv to .json in data enrichment, web extract, and web search documentation for consistency and improved data handling. Updated related command examples to reflect these changes.
- Upgraded actions/checkout from v4 to v6 for improved performance and features.
- Updated astral-sh/setup-uv from v5 to v8.0.0 to leverage new capabilities and optimizations.
- Removed unnecessary environment variable setting for JavaScript actions.
- Removed outdated references to `reference/` and replaced them with `references/` in multiple SKILL.md files.
- Updated template listings in clinical decision support and latex posters sections for better organization.
- Clarified instructions regarding GROBID service in the hypogenic skill documentation.
- Enhanced clarity in scientific slides documentation by correcting file path references.
- Updated version to 2.34.2 and enhanced project description in pyproject.toml.
- Removed obsolete marketplace.json file as its functionality is now integrated into the project structure.
- Adjusted GitHub Actions workflow to reference pyproject.toml instead of marketplace.json for version extraction.
