gh-quick-wins

/gh-quick-wins

Two-step recipe. The script applies a deterministic exclusion filter to all open issues and emits the surviving candidates as JSON. The LLM ranking pass then reads those signals and produces a top-10 markdown table ordered by impact / blast_radius. Step 1 is zero-LLM; step 2 is the only place LLM tokens are spent.

Blast radius is the load-bearing concept — not effort. A one-line CLAUDE.md typo fix is low effort but high blast radius (every future session is affected). A full Excalidraw reference guide is high effort but low blast radius (new file, isolated). This skill ranks by blast radius first, impact second.

Step 1 — Run the deterministic script

Run the script and capture its stdout. The interpreter path differs between Windows (Git Bash: Scripts/python.exe) and POSIX (bin/python); pick the one that exists:

PY="${CLAUDE_PLUGIN_DATA}/venv/Scripts/python.exe"
[ -f "$PY" ] || PY="${CLAUDE_PLUGIN_DATA}/venv/bin/python"
"$PY" "${CLAUDE_PLUGIN_ROOT}/scripts/gh-quick-wins.py"

The script exits 0 and prints a JSON array. Each surviving issue object includes a signals sub-object with:

• touches_load_bearing (bool) — body or title references CLAUDE.md, agents/, skills/, hooks/, or deploy.py.
• body_appears_drafted (bool) — body is long (>800 chars) AND contains a fenced code block, markdown table, or >3 ### headers. True means the issue body already contains the deliverable text.
• comment_count (int) — length of the comments array.
• days_since_update (int) — days since updatedAt.
• label_set (list[str]) — label names on the issue.

If the script exits non-zero, surface its stderr and stop.

The script supports --repo owner/repo and --max-ac N flags. Pass them through if the user specifies a different repo or checkbox threshold.

Step 2 — LLM ranking pass

Read the JSON from Step 1. Rank the surviving issues using the rubric below. Render a top-10 markdown table; if zero issues survived, render the zero- survivor message.

Scoring rubric

Blast radius (lower is better for quick-wins):

• low: New file under docs/, .tmp/, or a single isolated doc file. touches_load_bearing is False and the change scope is self-contained.
• med: touches_load_bearing is True — references CLAUDE.md, agents/, skills/, hooks/, or deploy.py. These affect every future session; downweight them even if they look small.

There is no high blast-radius tier in this output — items with clearly unbound blast should have been excluded by label (epic, umbrella) or prose (blocked by #N). If a load-bearing item slips through, label it med and let impact be the tiebreaker.

Impact (higher is better):

• high: body_appears_drafted is True (issue body already contains the deliverable — a new file's full content, a design already written out); OR comment_count >= 3 (active external interest); OR documentation label on a user-facing surface (README, getting-started guide, public reference doc).
• med: Isolated bug fix, small scoped feature, minor UX improvement.
• low: Internal-only cleanup, single-line typo, no user-facing effect.

Rank by: blast=low before blast=med; within same blast tier, rank impact=high before impact=med before impact=low; within same blast+ impact, rank by days_since_update ascending (fresher first).

Output format

Render as a markdown table with these columns:

| # | Title | Labels | Blast | Impact | Why |

• # — issue number as a link: [#N](url).
• Title — issue title, truncated to ~60 chars if needed.
• Labels — label_set joined by , . Empty if none.
• Blast — low or med per rubric above.
• Impact — low, med, or high per rubric above.
• Why — one sentence, max ~12 words, explaining why this is a quick win.

Cap the table at 10 rows. If more than 10 survived the filter, add a footer line:

Showing top 10 of N filtered candidates.

If zero issues survived the filter, render:

No low-blast-radius candidates in the current backlog.

Do not paraphrase or re-render the script's JSON — use its signals fields as the scoring inputs and produce the table directly.

Notes

• Step 1's JSON is the structural source of truth for issue data. Do not fetch issues a second time or call gh issue list again in step 2.
• The touches_load_bearing signal is a strong downweight, not a hard exclude. An issue touching CLAUDE.md can still appear as blast=med if its impact is high — it just ranks below equivalent blast=low items.
• body_appears_drafted=True is the strongest positive impact signal because it means the deliverable is already written. Shipping it is mostly a copy-paste from the issue to a new file.
• This skill is read-only. It does not create, label, assign, or close issues. The user always acts on the output themselves.

Validation

Two reference issues calibrate the scoring. Use these to verify the rubric stays calibrated after any changes to the skill body:

#789 — Excalidraw MCP server reference guide (PARAGON: HIGH quick-win)

• touches_load_bearing: False (new standalone file, no harness paths in body or title).
• body_appears_drafted: True (body contained the full reference guide text — a fenced code block, multiple headers, table of contents).
• Labels: documentation.
• Expected scoring: Blast=low, Impact=high. This is the archetype the skill is optimised for — content already exists in the issue, blast is zero, the documentation label confirms user-facing value.

#800 — bash-as-default migration (PARAGON: LOW / filtered)

• This issue would have been excluded at the deterministic filter stage by either the umbrella or epic label (it was a multi-step migration touching many harness files).
• If it somehow survived filtering (e.g. in a repo without those labels): touches_load_bearing: True (CLAUDE.md, agents/, hooks/ all in scope). body_appears_drafted: True (detailed implementation plan in body). Expected scoring: Blast=med, Impact=high. Ranks below any equivalent Blast=low item despite high impact — the harness-wide blast radius overrides the drafted-body signal.

If #789 does not score Blast=low, Impact=high in your output, the rubric has drifted. Re-check touches_load_bearing and body_appears_drafted against the signals the script emitted.