Skip to content

citation-queue.md

Latest commit

 

History

History
116 lines (98 loc) · 5.5 KB

citation-queue.md

File metadata and controls

116 lines (98 loc) · 5.5 KB
Raw

Spec: citation-queue (next template, not built)

Status: spec only. No code exists for this template. This document is the contract a first implementation must satisfy before it joins the catalog.

The bet

fresh-intent-reply-queue answers fresh intent: someone asked a buying question on Reddit or HN an hour ago. Its known weakness is the other half of the surface: the threads the answer engines already cite when a buyer asks ChatGPT, Gemini or Perplexity the same question. Those citations are durable - a thread an engine cites today it will probably keep citing for months - and a helpful, human reply inside one of them pays out on every future answer that quotes the thread. Fresh intent is a race; durable citations are real estate.

The data source already exists: am-i-cited asks the answer engines the buying questions in a niche, parses the links they cite, and tracks share of voice per domain over weekly re-samples. This template turns that citation feed into a work queue.

Known limit, priced in: a durable-citations-only queue drains in roughly six weeks for a niche - the cited set turns over slowly by definition. That is why this is a companion template, not a replacement: fresh-intent keeps volume, citation-queue keeps compounding value. The merged design is fresh + durable signals in one catalog, ranked by citation value where that data exists.

Pipeline (shared runner contract)

source -> normalize evidence -> qualify/extract -> draft -> approval queue -> outcome
  1. Source. An am-i-cited run export (JSON): the buying questions asked, the engines sampled, and for each answer the cited URLs with rank and sample date. v0 reads the export file; the direct connector is a later adapter behind the same template.yaml connector scope.
  2. Normalize evidence. Each cited URL becomes an item: canonical URL, surface type (reddit thread / HN thread / blog / forum / docs page), which questions cite it, which engines cite it, first-seen and last-seen sample dates. Dedupe on canonical URL across samples so a durable citation is one queue item with a citation history, not a weekly duplicate.
  3. Qualify. Score 0-100 on two axes, BYOK LLM with the ICP blurb:
    • citation value - how many buying questions and engines cite this URL, weighted by how durable it has proven across samples;
    • participability - can a human still add value here? Open Reddit/HN thread: yes. Archived thread, closed forum, a competitor's own docs: no. Threads fetchable for context go through the existing Exa / Firecrawl connector stubs rather than a new scraping path.
  4. Draft. A value-first contribution in the operator's voice for that specific thread: answer the original question properly, mention the product only where it genuinely helps. Few-shot from prior approved edits, same learning loop as the other templates.
  5. Approval queue. Unchanged and non-negotiable: every draft waits for a human. The operator posts from their own account. Nothing in this template ever posts, votes, or touches a social account.
  6. Outcome. Posted permalinks are recorded and watched for removal (72h, same best-effort RSS re-read as fresh-intent). The next am-i-cited re-sample closes the loop: did share of voice move, and is the thread we contributed to still cited?

Why the human gate matters more here, not less

Reddit's rules on automated posting do not bend because the thread is old - if anything, drive-by promotional comments on high-traffic cited threads are the fastest way to get a domain's citations removed. The approval step is the compliance layer, same as in fresh-intent, and the helpful:promotional guardrail ratio carries over unchanged.

Template contract sketch

name: citation-queue
version: 0.1.0
inputs:
  - kind: file
    format: am-i-cited export JSON
connectors:
  - am_i_cited_export   # v0: file import
  - exa                 # stub: thread/page context fetch
  - firecrawl           # stub: thread/page context fetch
trigger: manual (weekly, after each am-i-cited re-sample)
dag: [normalize, qualify, draft, approval]
approval_gates:
  - every draft            # human approval before anything leaves
  - promo_ratio_guardrail  # carried over from fresh-intent-reply-queue
budgets:
  weekly_llm_usd: 1.00     # qualification + drafting on a weekly cited set is small
  max_runs_per_day: 1
state: sqlite, append-only runs/edits/outcomes, as all templates

Cost shape

Weekly cadence on a niche-scale cited set (tens to low hundreds of URLs): one qualification call per new or re-surfaced URL, one draft call per approved-for-drafting item. On free OpenRouter models: $0. On cheap paid models: cents per week. The expensive input (the LLM sampling in am-i-cited) is already paid for by that tool's own budget.

Open questions for v0

  • Exact am-i-cited export schema version this consumes (pin one; the export is the interface).
  • Whether qualification fetches thread bodies in v0 or qualifies on title+question context only (cheaper; fetching is the upgrade).
  • How citation value should decay for URLs that drop out of the cited set between samples.

What done looks like for v0

A folder in templates/ that passes the native template contract: readable template.yaml, versioned prompts, offline evals over a fixture export, mock mode with no keys, and a README that says verified vs stubbed honestly - same bar as meeting-to-content.