hailneed Complaint radar (scan)
Turn scattered public complaints into a ranked, checkable report. Only public data is fetched; nothing about the user's repo, persona or chat history is sent anywhere.
Language rule: write every user-facing output in the language the user is speaking with you. Evidence rule: a theme needs ≥ 2 linked items; a claim without a link does not go in the report.
Step 0 — Configuration
Take the topic from the argument or the conversation. If it is missing, ask exactly one
question (topic + who the user serves), then proceed. Infer 3–6 HN queries and 1–3
GitHub Issue queries in the way people actually write (e.g. "<tool> frustrating",
"<job> any tool that", "<competitor> alternative"). Reddit stays off unless the
user explicitly asks for it — it is licence-constrained and rate-limited (say so once).
Read ${CLAUDE_PLUGIN_ROOT}/skills/scan/references/rubric.md before scoring.
Step 1 — Scan
Create ~/.claude/radar/ if missing (expand the home directory yourself; never pass a
literal ~ to the script), then run:
node "${CLAUDE_PLUGIN_ROOT}/scripts/radar.mjs" --topic "<topic>" --days 30 \
--hn "<q1>;<q2>;<q3>" --github "<q1>;<q2>" --out "<home>/.claude/radar/<slug>-raw.json"
Optional flags: --apps <appstore ids> (1–2★ reviews of competitor apps), --reddit <subs> (only on explicit request), --trustmrr "<Category>,<Category>" (needs
TRUSTMRR_API_KEY; used by painradar:fit), --days 7 for a daily run.
If node is missing, stop and say Node.js 18+ is required. Read the JSON. Check
sources: any blocked: true or error must appear in the report's "Kaynak notu" —
never hide a failed source and never invent items for it. If totals.complaints < 8,
widen: add 2–3 broader queries or --days 90, say you widened, rerun once.
Step 2 — Filter
Keep an item only if it expresses a complaint, frustration, unmet need, workaround or
buying intent and is about the topic (complaint: true is a hint, you decide).
Drop promotion, generic news and duplicates. Paraphrase; quote at most a few words per
item; never include usernames, real names or personal details.
Step 3 — Cluster and score
Group items into 3–7 themes by job-to-be-done or failure mode, not by source. Score each
theme's Signal (1–5) with the rubric and write one line of reasoning per theme. If
~/.claude/radar/state.json exists, tag themes Yeni / Devam eden by matching
theme keywords to themes in the state (reuse the old slug when it is clearly the same
theme).
Step 4 — Solutions and content ideas
Per theme propose 2–4 solution paths of mixed type — product fix / workaround the reader can apply today / offer someone could sell / education-content — each with effort (D/O/Y), impact (D/O/Y) and the evidence items that motivated it. Be specific. Add 2–3 content one-liners per theme tagged with a format (blog, linkedin, x_thread, short_video).
Step 5 — Write
Save ~/.claude/radar/<slug>-<YYYY-MM-DD>.md with exactly this order:
- Title + Kaynak notu (timeframe, item count, which sources worked / were blocked)
- Three-sentence summary
- Themes table (#, theme, Signal, Yeni/Devam eden, evidence count, one line)
- One block per theme: why this score · what people say · who it hurts · evidence links · solutions table · content ideas
- Raw evidence table (source, title, date, engagement, theme, link) — keep it complete
- Configuration & assumptions + one suggestion for the next run (a query to add/drop)
Then update ~/.claude/radar/state.json (runs, themes{slug: first_seen, last_seen, runs_seen, peak_signal, keywords}, seen_urls capped at 500). If it cannot be written,
say so in one line and continue.
Step 6 — Reply
5–8 lines in chat: file path; top 3 themes with Signal and Yeni/Devam eden; the single
best content idea; which sources were blocked; and one line: /painradar:fit ranks
what is worth building from this report, /painradar:launch turns it into a launch plan.
If evidence was thin, say "güçlü bir sinyal yok" plainly instead of padding.