Skip to content

CLAUDE.md

Latest commit

 

History

History
50 lines (36 loc) · 4.02 KB

CLAUDE.md

File metadata and controls

50 lines (36 loc) · 4.02 KB
Raw

TrendingAI — project rules for Claude Code

Auto-loaded when Claude Code starts in the cloned bundle directory. Read these once; they govern every /trendingai run.

First time using this bundle?

Read README.md cover-to-cover. The /trendingai skill walks you through env-var setup and the pipeline. Everything you need is there.

Canonical entry point

/trendingai slash command → reads .claude/commands/trendingai.md → invokes the TrendingAI skill at .claude/skills/TrendingAI/SKILL.md.

Trigger phrases that route the same way: run trendingai, let's do our AI trending report, AI trending report, fresh trendingai run, full harvest.

Hard rules for this project

  1. All synthesis happens INLINE in this Claude Code session. Opus 4.7 (the model running this conversation) IS the synthesis layer. NO Task subagents. NO claude --print subprocess. NO @anthropic-ai/sdk imports. NO ANTHROPIC_API_KEY reads (the orchestrating session is the model — it authenticates via OAuth).
  2. The renderer is FROZEN. Do not edit src/render-features-first.ts for cosmetic changes. Layout/visual changes are out of scope.
  3. The deterministic surfaces are FROZEN. Data sources, harvesters, extractors (src/features-first.ts), clustering math, scoring formulas, scripts/classify-one-to-watch.ts, scripts/fetch-captions-and-match.ts — do not edit. They produce byte-identical output against unchanged input by design.
  4. Each run produces a NEW numbered file. Never overwrite the existing out/features-confirmed.html (immutable baseline included with bundle) or any prior out/features-confirmed-runNNN-*.html. New runs land at out/features-confirmed-run{NNN+1}-{anchor}.html. Update out/latest.html symlink to point at the freshest.
  5. JSON-schema validation is mandatory. After EVERY synthesis stage writes its output JSON, call bun scripts/validate-synthesis.ts --stage <name> --file <path>. On failure (exit 1), retry the synthesis ONCE with the same prompt/data; on second failure, mark the affected row with error: "schema-incomplete" and continue.
  6. Hard $ cost ceiling per run: ~$0.30. Data-source providers + Whisper transcription. Anthropic is $0 marginal (subscription via session OAuth). Announce estimated spend at Stage 1 before harvest fires.
  7. Wall-clock target: <25 min end-to-end for fresh-harvest. If a stage exceeds 5 min, narrate the delay; if any stage exceeds 10 min OR the pipeline exceeds 25 min total, surface the overrun and ask the user whether to continue.

Don't do these things

  • Don't run bun run features:sentiment:legacy or any *:legacy script. They exist as reference only — they use claude --print which lost subscription billing post-2026-06-15.
  • Don't suggest installing @anthropic-ai/sdk, anthropic, or any other SDK package. The bundle has none and shouldn't.
  • Don't suggest a workflow that bypasses the per-stage JSON validation. The validator catches schema drift before the renderer breaks.
  • Don't overwrite out/features-confirmed.html (the baseline reference). Each run is archival-additive.
  • Don't fabricate citations, quotes, or feature URLs. Synthesis stages cite real URLs from harvested data only; honest empty arrays beat fake content.

Cost transparency

Stage Hard cost
Data harvesting + reply threads + transcription ~$0.10-$0.20
Whisper audio transcription (for video reels) ~$0.05
Anthropic Opus (synthesis) $0 marginal (subscription)
Total fresh-harvest run ~$0.15-$0.30
Render-only path $0

What "done" looks like per run

  1. Pipeline narrated stage-by-stage with timestamp + count + status.
  2. JSON validated after every synthesis stage.
  3. out/features-confirmed-run{NNN}-{anchor}.html written with run-incremented N.
  4. out/latest.html symlink updated.
  5. Report opens in user's browser via interceptor open (or the absolute path is printed for manual open if Interceptor is unavailable).
  6. Final narration summarizes: run number, output path, size, content counts, wall-clock, hard $ spend.