| name | skill-upstream-pr |
| description | Improve an open-source GitHub skill and open a friendly suggestion PR upstream: fork, run skill-auto-improver, attach asm eval before/after metrics. Don't use for local-only skills, authoring from scratch, bulk repos, or registry publish. |
| license | MIT |
| compatibility | Claude Code |
| allowed-tools | Bash Read Write Edit Grep Glob |
| effort | high |
| metadata | {"version":"1.0.0","author":"luongnv89"} |
Skill Upstream PR
You are contributing quality improvements to someone else's open-source skill. The workflow is: fork → clone → improve via skill-auto-improver → push to fork → open a friendly suggestion PR upstream. You do not own the target repo. Every step assumes you are a polite contributor, not a maintainer.
Repo Sync Before Edits (mandatory)
Before any modification to the cloned fork, pull the latest from the fork's tracked branch:
branch="$(git rev-parse --abbrev-ref HEAD)"
git fetch origin
git pull --rebase origin "$branch"
If the working tree is dirty: stash, sync, pop. If origin is missing or conflicts occur: stop and ask the user. A stale fork that PRs yesterday's tree is worse than no PR.
When to Use
- User shares a GitHub URL to a public skill repo and asks to "improve", "level up", or "contribute to" it
- User says "open a PR to this skill" or "suggest improvements upstream"
- The target skill scores below 85/8 on
asm eval and the user wants the upstream author to benefit
Do not trigger for: editing a local skill with no upstream, authoring a brand-new skill, publishing to the ASM registry, or bulk-improving many repos in one pass. One skill → one PR.
Prerequisites
Verify each before cloning anything. Stop and tell the user if any fails.
asm on PATH (command -v asm)
gh on PATH and authenticated (gh auth status)
git on PATH
- Network access to GitHub
- Write access to your own GitHub account (for the fork)
Input
The user provides a GitHub reference to a skill. Accept any of:
https://github.com/owner/repo
https://github.com/owner/repo/tree/<branch>/<path/to/skill>
github:owner/repo[#ref][:path]
Normalize to owner, repo, optional ref, optional path. If path is not given and the repo has multiple SKILL.md files, ask the user which one.
Workflow
Execute phases in order. Do not skip or reorder.
Phase 0 — Fork and clone
cd "$(mktemp -d)"
gh repo fork "$OWNER/$REPO" --clone --remote
cd "$REPO"
git checkout "$REF"
gh repo fork --clone --remote creates the fork under your account, clones it locally, sets origin to your fork and upstream to the original repo. Verify with git remote -v before continuing.
Create a dedicated branch for the improvement:
git checkout -b "skill-upstream-pr/improve-$(date +%Y%m%d-%H%M%S)"
Phase 1 — Locate the target SKILL.md
If the user gave a path, use it. Otherwise:
find . -maxdepth 5 -name "SKILL.md" -type f
If more than one match and no path was given, list the candidates and ask the user to pick. Never guess — a wrong PR is worse than a delayed one.
Set SKILL_PATH to the directory containing the chosen SKILL.md (not the file itself) — asm eval takes a directory.
Phase 2 — Delegate to skill-auto-improver
This skill does not reimplement the improvement loop. Follow the workflow in skills/skill-auto-improver/SKILL.md with $SKILL_PATH as the target:
- Phase 0 of that skill: capture
.asm-improver/baseline.json
- Phase 1:
asm eval --fix
- Phases 2-4: category-by-category loop with the 85/8 floor
- Phase 5:
.asm-improver/report.md
If the baseline already passes 85/8, stop and tell the user — no PR needed for a skill that already meets the floor. Offer to find a different skill or a different target.
Phase 3 — Harvest metrics for the PR
Read two files produced by the auto-improver:
.asm-improver/baseline.json — the before snapshot
- The latest
.asm-improver/iter-N.json — the after snapshot
Extract for both:
overallScore, grade
- Every
categories[].score (7 categories)
topSuggestions summary (for context, not quoted verbatim)
Compute deltas. If the overall score did not improve by at least 3 points or no category moved from below 8 to at least 8, stop and tell the user — the change isn't meaningful enough to justify a PR. Offer the auto-improver report as feedback they can share informally instead.
Phase 4 — Build the PR body
Read references/pr-template.md and fill it in. The template enforces friendly, suggestion-style tone. Key sections:
- What changed — one-sentence summary
- Before/after metrics — table with overallScore, grade, and all 7 categories
- Files touched — list every modified path under
$SKILL_PATH
- Iterations taken — N of 8 from the auto-improver loop
- How to verify — the
asm eval command the maintainer can run locally
Tone rules (read references/tone-guide.md):
- Lead with "Hi — noticed X, wanted to share Y"
- Frame as a suggestion, not a fix: "happy to adjust if the direction doesn't fit"
- Never imply the maintainer did something wrong
- Thank them for open-sourcing the skill
- Close with "no obligation to merge — totally fine to close if this isn't the right direction"
Phase 5 — Last-mile confirmation (mandatory)
Opening a PR on someone else's repo is public and hard to reverse. Before pushing anything, show the user:
- The PR title
- The full PR body (rendered markdown)
- The exact diff (
git diff upstream/<default-branch>)
- The commands you are about to run
Ask the user to confirm. If they want edits, apply them and re-preview. Do not push until the user says "go" or equivalent.
Phase 6 — Push and open the PR
After approval:
git add -A
git commit -m "$(cat <<'EOF'
improve SKILL.md: clarify triggers, add acceptance criteria, fix frontmatter
Suggested via skill-upstream-pr. See PR body for before/after metrics.
EOF
)"
git push -u origin "$(git rev-parse --abbrev-ref HEAD)"
gh pr create \
--repo "$OWNER/$REPO" \
--title "$PR_TITLE" \
--body-file .asm-improver/pr-body.md
Never push to upstream. Never use --no-verify or other hook-skip flags. If gh pr create fails, stop and report — do not retry in a loop.
Print the PR URL returned by gh pr create so the user can open it.
Step Completion Reports (mandatory)
Emit a compact status block after each phase:
◆ Phase N — [phase name] ([N of 6])
··································································
[check 1]: √ pass
[check 2]: √ pass (note if relevant)
[check 3]: × fail — [reason]
Criteria: √ M/K met
______________________________
Result: PASS | FAIL | PARTIAL
Checks per phase:
- Phase 0 —
Fork created, Clone succeeded, Branch created, Remotes correct
- Phase 1 —
SKILL.md located, Path unambiguous
- Phase 2 —
Baseline captured, Auto-improver ran, Final score >= 85, All categories >= 8
- Phase 3 —
Before/after delta >= 3 points OR category promoted
- Phase 4 —
PR body rendered, Tone checks passed
- Phase 5 —
User approved
- Phase 6 —
Push succeeded, PR opened, URL printed
Acceptance Criteria
- Fork + clone done via
gh repo fork --clone --remote — never a direct clone of upstream
- Dedicated feature branch created before any edits
skill-auto-improver workflow run on the target; baseline + final JSON captured under .asm-improver/
- Overall score improved by ≥ 3 points, OR at least one category moved from below 8 to ≥ 8
- PR body built from
references/pr-template.md with the full before/after table
- User explicitly approved the PR preview before push
git push targets origin (the fork), never upstream
gh pr create --repo $OWNER/$REPO used — the PR targets the upstream repo
- PR URL printed to the user at the end
Expected output
- A new branch on the user's fork of
$OWNER/$REPO
- A public PR on
$OWNER/$REPO with a before/after metrics table and friendly tone
- A local
.asm-improver/ directory with baseline, iterations, report, and pr-body.md
Edge Cases
- Skill already passes 85/8 — stop at Phase 2; do not open a PR for a skill that already meets the floor
- Improvement too small (< 3 point overall delta AND no category promoted) — stop at Phase 3; offer the auto-improver report as informal feedback instead
- Multiple SKILL.md files in the repo — ask the user which one; never batch
- Fork already exists from a prior run —
gh repo fork --clone --remote reuses it; rebase on upstream's default branch before editing
- Upstream force-pushed or rewrote history — stop and ask the user; do not force-push the fork to "catch up"
asm eval --fix writes changes you disagree with — revert SKILL.md.bak and do targeted edits only; don't ship the auto-fix if it harms the skill
- Maintainer has a CONTRIBUTING.md or PR template — read it before Phase 4 and align the PR body with their conventions; their template wins over ours
References
references/pr-template.md — PR title + body template with before/after table
references/tone-guide.md — wording patterns for friendly, suggestion-style contributions
skills/skill-auto-improver/SKILL.md — the improvement loop this skill delegates to
asm eval --help — flag reference for the evaluator