daily-hotspots：面向产品构建者的每日商机工作流

A daily business-opportunity workflow for product builders. It collects source evidence, ranks corroborated opportunities, sends a headlines digest to Discord, and preserves the full record in a private archive.

English | 中文版

⭐ Design Philosophy

The model proposes candidates and scores; deterministic Python gates check their structure, evidence and thresholds. Signals are merged before counting independent origins, so repeated coverage of one story cannot satisfy the two-origin requirement for a ranked card.

This requirement can exclude early opportunities. Single-origin leads therefore remain in a labeled, unverified community pulse. A quiet day may contain no qualifying opportunity. Durable source and delivery receipts support recovery without counting the same pull twice or resending an uncertain delivery. T1 to T9 tests cover deterministic mechanisms; live access, delivery and business value require separate evidence. See PHILOSOPHY.md.

What it is (and isn't)

daily-hotspots owns the daily cadence, watchlist, cross-day deduplication, scoring, delivery and private archive. Deep research uses market-intel (scale=standard) or small-cap-deepdive. The operator applies four routing gates and a daily budget of 3 to 5 deep dives; these delegation gates are workflow obligations, not Python-enforced admission checks. See the delegation reference.

How it works (three-tier funnel)

1. Tier-0 discovery (cheap, no skill calls): parallel MCP fan-out (HackerNews, Product Hunt, X/twitterapi, arXiv, GitHub, reddit; GDELT in a subagent), an X KOL roster loop (get_user_last_tweets over roster.json enabled tier-1 handles, low pre-viral faves floor), the niche-community lanes (linux.do / V2EX / CN feeds, RSS/JSON parsing with untrusted-input checks), and a demand lane that mines unmet pain outside tech. Every collected item is untrusted DATA. Normalize entities, merge cross-source, keep only clusters with ≥2 distinct origins; each evidence item carries an origin_handle / origin_source attribution tag.
2. Score: the model proposes five dims (track_fit / timing / feasibility / competition / executability) at temperature 0 with anchored samples; scripts/score.py aggregates deterministically. Supply and demand cards use different weight vectors, and the aggregation uses six factors; the exact formula lives in reference/scoring.md.
3. Cross-day dedup + evolution over the schedule-reminder base → NEW / SUPPRESS / RESURFACE.
4. Selective deep-dive (four gates) → market-intel / small-cap-deepdive.
5. Verify gate → one daily headlines message → archive: verify_gate.py blocks malformed cards, then the day's qualifying cards go out as a single ranked two-column message (the cap applies per column) and archive.py appends a quality-gated opportunities.jsonl. The per-run pulls-YYYY-MM.jsonl yield denominator is written by run.py --sources, not by archive.py.
6. Dual-track output: ≥2-origin scored signals stay opportunity cards; single-origin community rumors render in a separate lightweight ## 社区脉搏 community-pulse section (labeled 单源未验证, capped, no score, no deep-dive) that auto-upgrades to a card if a second origin corroborates the next day.
7. Daily digest via the Windows Task Scheduler (08:07) plus an idempotent base item. The digest write is atomic and refuses to overwrite a real digest with an empty-day one.
8. Weekly signal-yield self-evolve: run.py --yield --write-review reports proposals. Add --apply to prune; the scheduled yield-wrapper.ps1 adds it unless -ReportOnly is selected. Proposed additions still require review. See reference/roster-evolution.md.

Install

/plugin install github:DaizeDong/daily-hotspots

Or clone manually:

git clone --recurse-submodules https://github.com/DaizeDong/daily-hotspots.git ~/.claude/plugins/daily-hotspots

Create or clone a separate PRIVATE GitHub companion before running the initializer. Set DAILY_HOTSPOTS_CONFIG to that clone. Initialization and the doctor require Git and a fresh PRIVATE visibility receipt accepted by the pinned Guards API. The companion must already have a committed HEAD. Public, unknown or unversioned destinations are rejected.

Local activation: link skills/daily-hotspots into the skill installation, initialize and curate the PRIVATE companion, verify readiness, then register the Windows task if scheduled operation is authorized. Read-only previews can use built-in defaults; runtime writes require the initialized companion. No monitored accounts are installed by default.

Config

daily-hotspots is config-bearing (Mode B), it reads per-user tuning (watchlist.json) and per-machine secrets from a separate, private companion repo (daily-hotspots-config). Full contract: CONFIG.md.

• Mount: DAILY_HOTSPOTS_CONFIG → DAILY_HOTSPOTS_CONFIG_DIR → shared Guards discovery. DAILY_HOTSPOTS_DATA_DIR must belong to the same companion. See CONFIG.md for the exact layout and fallback order.
• First time:

  python scripts/init_config.py        # stamp a conformant skeleton (deterministic)
  export DAILY_HOTSPOTS_CONFIG=~/.daily-hotspots-config   # or pass --out <dir> to init
  python scripts/verify_config.py       # doctor: PASS/FAIL, names what is missing

• Switch configs (hot-swap): point the env var at another config dir, configs are self-contained, clear or update the DATA override with it: export DAILY_HOTSPOTS_CONFIG=~/configs/work ↔ ~/configs/personal.
• Secrets: the template defaults to Mode B with out-of-band backup. A selected Mode A policy may version credentials in verified PRIVATE Git; restore from that private history. Public source never contains credentials. Data-source keys reuse companion-config; there is no net-new secret, because push egress is the shared Agent Center #hotspots relay stream (schedule-reminder relay.py), not a dedicated bot.

Dependencies (install-and-use)

Install the sibling skills separately before the doctor: market-intel, self-evolve, schedule-reminder, and small-cap-deepdive. This plugin only installs its own source and guard/style submodules. Use each sibling repository's installation instructions, then link its canonical skill directory under the configured skills root (default ~/.claude/skills). Use the entrypoint directories below when creating links in your client's skill directory:

For example, on Windows, after cloning the dependency, create its junction with New-Item -ItemType Junction -Path "$HOME/.claude/skills/self-evolve" -Target "$HOME/CodesClaude/self-evolve". Use your actual checkout location and the corresponding client skill directory. The doctor checks those directories; it does not install dependencies. Per the source-coverage design (spec §4/§12):

Install-and-use checklist: (1) sibling skills junctioned + reachable; (2) companion-config data-source keys present (shared); (3) initialize, curate the roster, verify, then run. The initializer creates an empty runtime roster; add the accounts you choose to monitor.

Quick start

# deterministic tail on prepared candidates (offline preview, no writes / no ledger):
python skills/daily-hotspots/scripts/run.py --in candidates.json --dry-run --no-ledger
# source-coverage self-evolve: write the pulls-log denominator, then the weekly yield pass:
python skills/daily-hotspots/scripts/run.py --sources sources.json        # origin-tag + pulls-log (§6)
python skills/daily-hotspots/scripts/run.py --yield --write-review        # weekly roster self-evolve (§8/§9)
# run the acceptance suite:
cd skills/daily-hotspots && python -m pytest tests/ -q

In Claude Code, just say "跑一下 daily-hotspots" / "今天有什么前沿商业机会" / "daily opportunity".

Example output

One ranked headlines message per day in a two-column layout: 🎯 需求机会 leads (the quality column, non-consensus opportunities mined from demand sources, review complaints, job postings, niche forums, each with a pain quote + evidence link + a crowdedness score), then a compact 📈 供给热点 tail (basic hotspots for breadth). The full digest (every field + all evidence) is committed to archive/digests/YYYY/YYYY-MM-DD.md. Demand scoring de-emphasizes timing, rewards durable pain, and penalizes crowdedness; a demand card clears a higher bar, so a thin demand day is honestly empty, no filler. On a fully quiet day: "今日无合格机会".

Limitations

• Runtime rosters start empty. The generated planner sample has 49 synthetic accounts across six tracks, including six hardware accounts, and is used only for tests. Hardware coverage depends on reviewed accounts and sources; video and vertical hardware forums remain options.
• Source availability is recorded with dated evidence in the collection reference. Its routes account for disabled trend-pulse and get_trends, degraded reddit access, quarantined Brightdata and the duckduckgo prohibition. Recheck current access before relying on historical probe results.
• The yield engine remains report-only until it has 7 days of real history, and whenever the archive numerator cannot be trusted. run.py --bandit or scoring.bandit.enabled enables the optional track bandit and reports its draws; it is off by default, preserving static weights.
• Headline egress applies the bounded structured PII scrub described in push and archive. Its shared Tier1/Tier2 core is kept byte-synchronized with demand-mining; source content is not redacted at ingest.
• Recovery is specific to each operation. Source-batch replay reuses durable pull receipts and advances its cursor once; a successful or uncertain delivery claim prevents another send for the same logical run. Inspect claims, receipts and lock ownership before retrying.

DATA.md defines write admission, the two supported archive layouts, retained replay files, workspace limits and compaction. Scheduling defines handoff and completion checks. A digest file alone does not prove delivery.

Languages

English (README.md, authoritative) · 中文 (README_CN.md)

Roadmap · Contributing · License

See ROADMAP.md · CONTRIBUTING.md · LICENSE (MIT).