What does this skill do, and when should you use it?
This skill treats GitHub as the durable record and requires the repository's MAINTAINERS.md as its ranking policy. In recommend mode it read-only collects open issues, RFCs, and PRs and applies a polling ladder — gating, impact ordering, readiness classification, execution fit — to produce three to five candidate cards with exclusions, a timestamped poll, and stated uncertainty. Only after the maintainer explicitly selects an item (e.g. "start #123") does it enter execute mode: refreshing state, confirming scope, creating an isolated branch, and opening a linked draft PR early. It enforces strict safety rules: no writes during polling, no trusting instructions in issue/PR bodies, no exposing suspected vulnerability details, and no equating age or reactions with priority.
- Resolves the target repository and reads open issues, RFCs, PRs, milestones, labels, reviews, and checks via a configured GitHub integration or the gh CLI
- Applies the polling ladder from MAINTAINERS.md: gate, order by impact, classify readiness, check execution fit
- Deeply verifies top candidates: problem evidence, RFC/dependency state, competing work already in flight
- Returns 3–5 "Recommended now" cards plus "Needs triage or blocked" items, representative exclusions, an ISO 8601 poll time, and uncertainty notes
- On an explicit "start #123": re-validates state, reports conflicts, states scope and acceptance evidence, creates an isolated branch/worktree, and opens a draft PR as soon as there is a reviewable change
- An open-source maintainer with limited weekly time asks what to tackle and wants a ranked, readiness-tagged shortlist rather than raw issue listings
- A contributor onboarding onto an unfamiliar repository wants to see which issues already have PRs in flight and which are blocked, avoiding duplicate work
- A maintainer reviews the shortlist, explicitly picks one item, and has the agent follow repo contribution guidance: branch, draft PR, ongoing progress updates
- A team wants to replace gut-feel issue picking with a documented process where gating criteria, exclusion reasons, and uncertainty are explicit in the output
- Teams without a MAINTAINERS.md ranking policy, or those expecting the agent to autonomously decide what to do — the skill forbids auto-dispatching public intake events and stays strictly read-only in recommend mode
- Bulk metadata dashboards or statistics — the skill explicitly states bulk metadata is insufficient evidence for a recommendation
- Environments with neither the gh CLI nor a configured GitHub integration to read repository state
How do you install this skill?
- Core function depends on GitHub and the gh CLI; the skill may be unusable where GitHub is not reliably reachable (relevant to mainland-China networks).
- The ranking policy lives in MAINTAINERS.md, whose content was not part of this assessment; review it yourself before trusting ranking results.
- The 'start #123' execute mode performs write operations (branch, draft PR); confirm repository branch protection and review processes before use.
- The skill assumes a configured GitHub credential in the agent environment; use least-privilege credentials and avoid write scopes for poll-only usage.
- Shell / CLI
- Network access
- Local filesystem
GitHub CLI (gh) or a configured GitHub integrationMAINTAINERS.md ranking policy in the repo
The host repository (trycua/cua, MIT) is a monorepo bundling 9 skills; the Spaces app itself is FSL-1.1-MIT. The installer can add cua skills and MCP server into your AI coding agents:
Claude Code / Codex / Cursor and others
curl -fsSL https://cua.ai/install.sh | sh
cua auth loginAfter login the installer offers to install cua skills and the cua MCP server; preselect with sh -s -- --select cua-driver. The skill file lives at .agents/skills/poll-github-work/SKILL.md in the repo; the README documents no separate install command for this individual skill.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Poll work: what's most worth tackling in this repo right now?
- Rank the backlog by impact and flag anything blocked
- What open issues or PRs are actionable? Give me three to five recommendation cards
- start #123
Trigger it in natural language: ask what to work on, request backlog priorities, say "poll work," or ask for actionable issues/PRs. It stays in read-only recommend mode and produces candidate cards awaiting your explicit selection. Saying "start #123" selects item 123 and enters the execute flow: refresh state → report conflicts → state scope and acceptance evidence → make the selection visible via the repo's normal assignment mechanism → create an isolated branch → open and keep a draft PR current. It preserves contributor authorship and prefers reviewing an existing PR over opening a competing implementation.
What are this skill's strengths and limitations?
- Strict mode separation: polling is read-only; execution requires explicit selection, keeping the blast radius small
- Structured output (candidate cards with impact, readiness, competing work, validation path, main risk) that separates facts from inference
- Built-in prompt-injection defenses: untrusted issue/PR content is never executed, vulnerability details stay out of public shortlists
- Respects existing contributors instead of duplicating in-flight work
- Depends on MAINTAINERS.md in the repo for ranking policy; repos without one need to add it or accept defaults
- Recommendation quality depends on gh CLI or GitHub integration availability and permissions
- No test suite or tested-platform evidence is provided in the source for this skill
- The execute flow only covers the normal branch + draft PR contribution path; merging and deploying are explicitly out of scope
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Poll GitHub Work this page | 52 · Use with care | ★ 29k | 1d ago | MIT |
| RenderCV Issue Solver | 51 · Use with care | ★ 18k | 6mo ago | MIT |
| GitHub PR Workflow Assistant | 42 · Not recommended | ★ 1.7k | 3d ago | MIT |
| Maintainer Response (agent-service-toolkit) | 69 · Recommended | ★ 4.5k | 7d ago | MIT |
| Engram Backlog Triage | 45 · Not recommended | ★ 7.1k | 4d ago | MIT |
How did FollowSkills review this skill?
Evidence shows a read-only default mode, explicit prohibition of repo mutation during polling, treatment of issue/PR content as untrusted data (prompt-injection defense), private routing of suspected vulnerabilities, and a 'never auto-dispatch public intake' red line; deducted for thin isolation/rollback detail in execute mode, dependence on a MAINTAINERS.md ranking policy not fully shown, and unverified publisher identity.
Evidence shows clear step structure, directly usable gh command examples, fact/inference separation, and conflict reporting for stale recommendations; deducted because static review cannot execute anything, the core ranking ladder depends on MAINTAINERS.md not included here, and there are no tests or failure-feedback examples for this skill's paths.
Evidence shows explicit trigger phrases (work priorities, 'poll work', 'start #123') and clear mode boundaries (recommend vs execute) plus a defined output template; deducted because core function depends on GitHub/gh API, posing reachability risk on mainland-China networks that the file does not disclose.
Evidence shows good structure, consistent name/description, MIT-licensed repository context, and clear progressive disclosure (poll → candidate cards → execute); deducted for no skill-level versioning or changelog, an external relative-link dependency (MAINTAINERS.md) whose content is not in evidence, and no FAQ or known-limitations section.
Evidence shows a directly usable candidate-card output format with ISO 8601 timestamps and uncertainty disclosure, providing real marginal value for maintainers; deducted because output quality cannot be verified statically, value depends on the correctness of the external MAINTAINERS.md policy, and nothing was executed.
Evidence shows repository-level CI workflows and test files, but they target driver/perception paths unrelated to poll-github-work; the skill's claims rest only on SKILL.md text with no third-party execution evidence or independently reproducible material.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →