Revise decision-guide-agent spec per auto-researcher review (#22)
The auto-researcher project (~/Documents/auto-researcher/) already built ~70% of this agent — the full research engine, with 4 shipping-ready posts. Its review flagged the build directive + 8 empirically-grounded
The auto-researcher project (~/Documents/auto-researcher/) already built ~70% of this agent — the full research engine, with 4 shipping-ready posts. Its review flagged the build directive + 8 empirically-grounded methodology corrections. Folded all of it in:
Build directive:
- Build ON auto-researcher (~21-24 hrs), not fresh (~80-100 hrs). Added
the reuse map + read-first file order. New agent = marketing_bot/ agents/decision-guide/ calling into auto-researcher modules.
Methodology corrections (battle-tested across 4 domains, override the thinner earlier draft):
- Topic intake: narrow "first X for buyer Y use case Z", not "best X";
niche discovery surfaces 4-5 real posts where naive read writes 1
- Harvest whole niche once (~100k threads), narrow per topic to 15-50;
A/B from full corpus not per-topic slice
- Pareto + table-stakes BEFORE weighting; 7-9 criteria not 3-5
- Pareto-first for fragmented fields, conjoint-first only when a T1
source covers >70% of candidates
- mention-frequency != decision-weight (documented haircut)
- 3-tier table: ranked / dominated / disqualified
- structured per-brand quote pull; comparison-ratio signal
- engagement (48h comment replies) is part of the strategy
GDA config: answered the reviewer's 5 open questions from this session's work (data API = /api/recent-deals/<slug>; analytics = GA4 527676777 + ai-referrals + GSC; Flask/Heroku /blog/<slug> pages; voice primitives from auto-researcher farmout spec; edges.jsonl schema = cross-property spec §4, MP5K post edge_gda_011 as template). Plus the SKU classifier reuse (dealClassifier.js) and the UTM convention.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>