Menu
Classify any GitHub URL, issue/PR number, or freeform task and route it to the right downstream skill (clud-fix for bug-shaped issues, clud-pr for features / freeform / PR triage). One-line dispatcher; no scope of its own.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
File a deeply-researched GitHub issue via investigate โ investigate โ post, returning a summary plus the issue URL. Files without asking for confirmation.
Spin up a fast Linux build container for a Rust + soldr + zccache, Python (uv), or C++ (CMake + ccache) project using the bundled `docker-build` tool family. Uses anonymous Docker volumes for build state and a read-only bind for source โ the one rule that turns Docker-Desktop's 20-minute cold-build into a sub-30-second warm cycle.
Diagnose and recover a wedged Docker Desktop (engine pipe/socket absent while the UI stays alive, WSL/Docker startup failures) and answer Docker VM disk-growth / memory-pressure questions. Read-only `doctor` first; every restart/reset is confirmation-gated and preserves images/volumes; storage disks are resolved from Docker Desktop's real config (never assumed) and never compacted or deleted automatically.
Coordinate dependent cross-repo changes under a repo-local .extern-repos/ checkout convention.
Show a git diff in a native OS webview window (Beyond Compare-style dual-pane with file picker on the left) via the bundled git/clud-git-diff.py tool. Invoke when the user asks to visually review changes between revisions.
Worktree, branch, process-audit, and quarantine playbook extracted from /clud-pr so /clud-fix-quick and other skills can reuse it. Use this when the user asks for worktree cleanup, stale-branch teardown, or a process audit on a leftover dir.
| name | clud-do |
| description | Classify any GitHub URL, issue/PR number, or freeform task and route it to the right downstream skill (clud-fix for bug-shaped issues, clud-pr for features / freeform / PR triage). One-line dispatcher; no scope of its own. |
| triggers | ["When the user says \"/clud-do <url-or-task>\"","When the user invokes \"/loop clud-do <url>\"","When the user gives an issue/PR URL but isn't sure which sub-skill to invoke","When a meta-skill or harness needs an entry point that classifies before dispatching"] |
Thin router on top of [[clud-fix]] and [[clud-pr]]. Reads any GitHub issue URL, PR URL, issue/PR number, or freeform task; classifies the input; dispatches to exactly one downstream skill; recovers cleanly when the first dispatch was a mis-triage.
/clud-do does not code, does not branch, does not set /goal, and does
not own any worktree of its own. The child skill owns all of that.
For code changes, the child skill ([[clud-fix]] or [[clud-pr]]) preserves
RED -> GREEN: identify or add the focused failing test/repro first,
implement the scoped change, then rerun that focused signal until it
passes before broad gates. /clud-do itself never makes code changes,
so the rule applies through delegation โ never directly.
https://github.com/<owner>/<repo>/issues/<num>.https://github.com/<owner>/<repo>/pull/<num>.#<num> (issue) or !<num> (PR) when the current checkout
unambiguously identifies the repository. Resolve <owner>/<repo> with
gh repo view before acting.add a structured changelog generator.Optional user hint may prefix the argument: /clud-do fix <url> forces
the /clud-fix route; /clud-do feat <url> forces the /clud-pr route.
User hints win outright over every other classifier signal.
This skill defaults to GitHub and the gh CLI for backwards compatibility. URL inputs from other forges are classified by URL prefix and routed to the matching native CLI. Bare numbers (#<N>) without an explicit prefix resolve their forge from the current worktree's git remote get-url origin.
| Forge | URL prefix(es) | Native CLI | Vocabulary |
|---|---|---|---|
| GitHub | github.com/<o>/<r>/(issues|pull)/<N> | gh | issue / PR (#N) |
| GitLab | gitlab.com/<g>/<p>/-/(issues|merge_requests)/<N> and self-hosted variants | glab | issue / merge request (MR) (!N) |
| Bitbucket | bitbucket.org/<o>/<r>/(issues|pull-requests)/<N> | none official; REST API | issue / PR (#N) |
| Gitea | <host>/<o>/<r>/(issues|pulls)/<N> | tea | issue / PR (#N) |
| Forgejo | <host>/<o>/<r>/(issues|pulls)/<N> (same patterns as Gitea) | forgejo-cli (early) or tea | issue / PR (#N) |
| Self-hosted GitLab / Gitea / Forgejo | same patterns under custom domains | same CLI | same vocabulary |
The classifier returns {forge, kind, owner, repo, number, host} for any URL input.
When the input is a bare #<N> or <N>:
git remote get-url origin in the current worktree.gh pr view for GitHub, glab mr view for GitLab, etc.).Prefixes in the invocation force a specific forge and skip remote inference: github:<N> / gitlab:<N> / bitbucket:<N> / gitea:<N> / forgejo:<N>.
All gh examples elsewhere in this skill are GitHub-specific. Substitute the matching native CLI per forge:
gh issue view <N> โ glab issue view <N> โ tea issues show <N> โ Bitbucket REST: curl ... /repositories/<o>/<r>/issues/<N>gh pr view <N> โ glab mr view <N> โ tea pulls show <N> โ Bitbucket REST: curl ... /pullrequests/<N>gh pr merge <N> --squash โ glab mr merge <N> --squash โ tea pulls merge <N> โ Bitbucket REST: PUT /pullrequests/<N>/mergegh issue create โ glab issue create โ tea issues create โ Bitbucket REST: POST /repositories/<o>/<r>/issuesInternal skill logic can keep saying "PR" generically. User-facing output uses the forge's native vocabulary:
PR #123 merged โ unchanged.MR !123 merged (note the ! sigil GitLab uses instead of # for MR references).PR #123 merged.Never silently translate vocabulary in error messages โ if a GitLab MR is mentioned, the message says MR !123, not PR #123.
Each forge has its own auth model:
gh auth status or GITHUB_TOKEN env var (default).glab auth status or GITLAB_TOKEN / GL_TOKEN.BITBUCKET_TOKEN).GITEA_TOKEN, FORGEJO_TOKEN).If the required CLI or token is missing, emit a clear refusal and stop:
forge-cli-missing: install <cli> to use clud against <forge>
forge-auth-missing: authenticate to <forge> via <cli> auth login
Don't log or persist tokens; rely on the user's existing auth.
gh / glab / tea / etc. is on PATH; refuse if not. Don't bundle tooling.MR !123, not PR #123./goal. The child skill owns its goal, worktree, and terminal sentinels.gh issue view, gh pr view). /loop will re-invoke; the second iteration's
"open PR exists, drive to merge" is a different dispatch than the first
iteration's "implement from scratch."/clud-fix and let its meta workflow handle the children.Read once with gh issue view <num> --repo <owner>/<repo> --json number,state,title,body,labels,comments,url,closedByPullRequestsReferences
(or gh pr view for PR inputs). Then dispatch from this table:
| Input | State | Dispatch |
|---|---|---|
Issue URL or bare #N | OPEN, bug-shaped | /clud-fix <url> |
| same | OPEN, feature-shaped | /clud-pr <url> |
| same | OPEN, meta/parent/burn-down | /clud-fix <url> (its meta workflow owns this) |
| same | OPEN, ambiguous | default to /clud-fix (stricter gate catches mis-routes; recovery branch below handles the rejection) |
| same | CLOSED, resolving PR merged | emit done-terminal (no dispatch) |
| same | CLOSED, no resolving PR | emit done-terminal |
PR URL or bare !N | OPEN | /clud-pr <pr-url> (triage mode) |
| same | MERGED | emit done-terminal |
| same | CLOSED, not merged | /clud-pr <pr-url> (its CLOSED-not-merged path) |
| Freeform task sentence | n/a | /clud-pr <text> |
| Non-GitHub URL / Actions run / Gist / other | n/a | refuse with unreachable-scope sentinel |
Classify the issue as meta/parent/burn-down when labels, title, checklist body, linked issue references, GitHub sub-issue metadata, or explicit user wording identify a tracker โ same rule [[clud-fix]] uses.
[[clud-fix]] does NOT require a bug label or [bug] title prefix โ its
gate is purely content-based. So /clud-do treats labels and titles as
signals, not requirements. Apply in priority order, highest signal
first; stop at the first signal that resolves the input:
/clud-do fix <url>,
/clud-do feat <url>) wins outright.bug โ fix; enhancement / feature / feat โ pr.fix:, feat:)./clud-fix when nothing matches. Its readiness gate
becomes the secondary classifier and the recovery branch below catches
the mis-route cleanly./clud-fix emits a precise terminal when its intake gate rejects:
<clud-fix:terminal kind=investigation-report-posted reason=intake-failed url=<issue-url>>
If and only if the initial dispatch was /clud-fix AND its terminal was
that exact sentinel, /clud-do recovers:
Read the investigation report /clud-fix just posted to the issue.
Use gh issue view <num> --json comments --jq '.comments[-1].body' (or
filter for the most-recent clud-fix intake gate failed banner). The
report enumerates which readiness items were missing
(reproduction / symptom / criterion / scope).
Re-classify with the post-rejection signal. A feature-shaped recovery candidate has:
feat: title, or enhancement /
feature label that the initial pass missed because a stronger bug
signal was present./clud-fix investigation report cites missing reproduction OR
missing symptom (the bug-specific items) but does NOT cite missing
fixed criterion. The latter would mean even the feature shape is
under-specified.One recovery attempt per loop iteration. If the candidate
qualifies, dispatch to /clud-pr <url> and emit:
<clud-do:terminal kind=recovered-to-clud-pr original=clud-fix reason=feature-shaped-issue url=<url>>
If the candidate does NOT qualify (no acceptance criteria AND no feature labels AND the report cites missing fixed criterion too), the issue is genuinely under-specified for either flow. Do NOT re-dispatch. Emit:
<clud-do:terminal kind=routing-failed reason=neither-bug-nor-feature-shape-clear url=<url>>
Suggest the user add either "Steps to reproduce" (bug shape) or "Acceptance criteria" (feature shape), then stop.
No re-recovery. /clud-pr's task-to-PR workflow is permissive โ
if it also refuses, that's a real refusal, not a triage error. Emit:
<clud-do:terminal kind=both-skills-refused url=<url>>
Recovery is unidirectional (/clud-fix โ /clud-pr only), cycle-free
by construction.
Other /clud-fix terminals are NOT recoverable โ
unreachable-scope-refused and empty-children-report-posted mean the
work genuinely can't be done in this repo; re-dispatching won't help.
Pass them through verbatim and emit the matching wrap-and-add sentinel.
The existing /loop dynamic-mode contract handles /clud-do with zero
/loop changes:
/loop clud-do <url> parses with rule 3 (no interval) โ prompt =
clud-do <url>./loop wakes; invokes /clud-do. /clud-do reads
issue state, classifies, dispatches to /clud-fix or /clud-pr. The
child runs to its terminal (merged + closed + validated,
investigation-report-posted, unreachable-scope-refused, PR-opened,
etc.)./loop wakes again; re-invokes /clud-do. /clud-do
re-reads issue state โ issue may now be CLOSED, or have an open PR, or
have a merged PR awaiting validation. It re-classifies against the new
state and either re-dispatches (e.g. "open PR exists โ /clud-pr to
drive merge") or emits a done-terminal./clud-do is stateless across loop iterations. No ledger, no
.clud/do/<hash>.json. The state of truth is GitHub. /clud-fix's own
meta-workflow ledger is unchanged.
The loop's natural termination is the same as today: when /clud-do's
done-terminal sentinel fires and the loop's prompt has nothing else to
do, the next iteration's /clud-do emits the same done-terminal again.
The loop user (or an outer /goal) can detect that and stop scheduling
further wake-ups.
/clud-do emits both the child's verbatim sentinel AND a
<clud-do:terminal โฆ> wrapper that includes the child's terminal as a
nested field:
<clud-fix:terminal kind=investigation-report-posted reason=intake-failed url=...>
<clud-do:terminal kind=recovered-to-clud-pr original=clud-fix child-terminal=intake-failed url=...>
Outer /goal evaluators can match on clud-do:terminal exclusively
without learning every child skill's vocabulary; the transcript still
carries the precise child terminal for debugging. Belt-and-suspenders.
Sentinel kinds /clud-do emits:
kind= | Meaning |
|---|---|
dispatched-to-clud-fix | Initial dispatch to /clud-fix succeeded; child terminal passed through. |
dispatched-to-clud-pr | Initial dispatch to /clud-pr succeeded; child terminal passed through. |
recovered-to-clud-pr | /clud-fix rejected at intake; recovered by re-dispatching to /clud-pr. |
done-already | Issue/PR is already in its terminal state (closed + resolving PR merged, PR merged); no dispatch needed. |
routing-failed | Issue is under-specified for both bug and feature shapes. |
both-skills-refused | Recovery dispatch to /clud-pr also refused. |
unreachable-scope-refused | The reachable-scope gate fired before dispatching. |
Every sentinel includes url=<issue-or-pr-url> (or task=<freeform>
for freeform inputs). Emit the sentinel as the FINAL line of the
user-facing response.
Every recovery iteration should print a one-paragraph diagnostic so the user can see why the router second-guessed itself:
/clud-do routed to /clud-fix (signal: no `feature` label, no `feat:` prefix,
body contained "Issue" keyword).
/clud-fix rejected at intake โ investigation report at <comment-url>;
missing: concrete reproduction, falsifiable symptom.
Re-classifying: issue body has an explicit "Acceptance criteria" section
and three checkbox items. Re-dispatching to /clud-pr as a feature
implementation.
That's diagnostic AND trains the heuristic โ next time the user types
/clud-do <similar-url> (or files the same shape of issue), they know
what to do. If the user disagrees with the recovery on a specific issue,
they can add a bug label so the next loop iteration's classifier picks
it up correctly.
/clud-fix โ /clud-pr,
never the other way. If /clud-pr also refuses, emit
both-skills-refused and stop; do not retry.routing-failed is the right terminal.
Same trap as /clud-fix documents at its /goal-wedge failure mode./clud-issue-triage. Closed issues with resolving PRs
route to done-already, not to a closure-audit. Suggest
/clud-issue-triage and stop if the user wants a triage./clud-fix for users who know they want a bug fix.
/clud-do is an additive router; direct /clud-fix <url> calls
continue to work unchanged. Do not insert /clud-do between the user
and /clud-fix when they typed /clud-fix themselves./clud-do only reads; it does not edit
the issue, add labels, or close it. Recovery posts no new comment of
its own โ the /clud-fix investigation report is the user-visible
artifact./clud-fix or /clud-pr โ honour the
request. Use /clud-do only when the user invoked it directly or when
the input shape genuinely needs classification./clud-do.