extract-skill
Reverse-engineer design systems, tokens, and components from live products or screenshots
Move incoming issues and pull requests through triage states until each is actionable or closed — use when the queue has piled up or a report arrives unsorted
$ npx -y skills add nyldn/claude-octopus --skill skill-intake --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/skill-intakeContext preview
The summary Claude sees to decide when to auto-load this skill.
Move incoming issues and pull requests through triage states until each is actionable or closed — use when the queue has piled up or a report arrives unsorted
name: skill-intake description: "Move incoming issues and pull requests through triage states until each is actionable or closed — use when the queue has piled up or a report arrives unsorted" disable-model-invocation: true
> **Host: Codex CLI** — This skill was designed for Claude Code and adapted for Codex. > Cross-reference commands use installed skill names in Codex rather than `/octo:*` slash commands. > Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it. > For host tool equivalents, see `skills/blocks/codex-host-adapter.md`.
Work the incoming queue. Each item moves through a small set of states until it is either actionable by someone who did not write it, or closed with a reason.
**A pull request is an issue with attached code.** Same states, same moves, with the deltas noted below. Resolving a bare `#42` means checking both.
Adapted from `triage` in [mattpocock/skills](https://github.com/mattpocock/skills) (MIT), cut down to the states this repo can actually represent and retargeted at `gh` and `bd`.
`skill-staged-review`; intake decides whether the PR should be reviewed at all.
An issue or PR reference, or nothing — with nothing, take the queue in order. `gh issue list` and `gh pr list` for the public surface; `bd ready` and `bd blocked` for tracked work.
What kind of thing is it: a bug, an enhancement, a question, or noise. For a PR, also: does it correspond to an existing issue, or arrive unannounced?
Do not skip this because the title looks obvious. Titles are written by people who already know what they meant.
The step that earns the whole skill. For a bug: **reproduce it, or establish that you cannot.** For a PR: check the claim it makes is the change it contains.
An unverified bug report is a hypothesis. Filing it as fact wastes whoever picks it up. If reproduction needs something you do not have — credentials, a dataset, a platform — that is `needs-info`, not `verified`.
detail" is not a request anyone can act on.
has not made. Do not resolve these by inference. Escalate, or run `skill-pressure-test` with the maintainer.
reason, and for out-of-scope, why.
For anything reaching `verified`, write what someone picking it up needs and would otherwise have to rediscover: where the relevant code is, what you already ruled out, and how to tell when it is fixed. This is the difference between an item that gets worked and one that gets re-triaged.
For a PR, add: whether it has tests, whether CI is green, and whether it conflicts with anything in flight.
For public comment text, stream the completed note through the outbound gate:
printf '%s\n' "$TRIAGE_NOTE" | \
"${CLAUDE_PLUGIN_ROOT:-${HOME}/.claude-octopus/plugin}/scripts/safe-gh-comment.sh" \
--repo OWNER/REPO issue-comment ISSUE_NUMBER -Use `gh issue edit --add-label` only for the label mutation. Use `bd update` for tracked work. If `bd` writes are blocked by pending migrations, say so and do not run the migration — see `skill-work-slicing` for that constraint.
1. The repository — reproduce against the actual code. 2. CI logs and run history for anything claiming a failure. 3. The reporter, for what only they can supply. 4. Never assume the report is accurate because it is detailed.
duplicate. Closing is the one move that is rude to get wrong.
reports end up buried under a label.
Per item:
1. **Reference and category.** 2. **State assigned, and why.** 3. **Verification result** — reproduced, could not reproduce, or not attempted with the reason. 4. **Brief** — for verified items only. 5. **Action taken** — labels, comments, or the exact commands not run and why.
Then a queue summary: counts by state, and which items are now takeable.
not and why.
Every AI model has blind spots. Claude Octopus supports twelve external provider integrations — Codex, Antigravity CLI, Copilot, Qwen, Ollama, Perplexity, OpenRouter, OrcaRouter, OpenCode, Cursor CLI, Grok, and Kimi Code — alongside the built-in Claude Code
Repo: nyldn/claude-octopus
Reverse-engineer design systems, tokens, and components from live products or screenshots
Multi-AI requirements scoping using available external providers (Double Diamond Define phase). Priority triggers: octo define, octo scope, co-define,…
Multi-AI validation, scoring, and review using available external providers (Double Diamond Deliver phase)
Multi-AI implementation using available external providers (Double Diamond Develop phase). DO NOT use for simple code edits, reading/reviewing code, built-in…
Multi-AI research using available external providers (Double Diamond Discover phase)
Decompose and execute large changes, migrations, or multi-issue fixes in parallel with quality gates