Data Storyline Miner

Data Storyline Miner

The corpus is the content engine. This internal marketing skill sweeps the market-intelligence surface across verticals, deal types, and recency windows looking for contrasts a reader would stop scrolling for — "Beauty deal volume is up but median value is down", "Affiliate is eating flat-fee in Fitness" — and drafts publishable newsletter and LinkedIn copy from them, every claim shipping with its provenance line so the post is defensible the moment it goes out. It is top-of-funnel sourced from our own data; the artifact is a ranked shortlist of candidate stories plus ready-to-paste drafts. For a single vertical's leadership memo use vertical-briefing; for movement-over-time use vertical-forecast-brief.

Read first: ${CLAUDE_PLUGIN_ROOT}/shared/conventions.md (tool schemas, credit prices, the conventions). This skill honors thrifty/thorough credit modes (${CLAUDE_PLUGIN_ROOT}/shared/credit-modes.md) and Refusal Recovery (${CLAUDE_PLUGIN_ROOT}/shared/refusal-recovery.md). No creator lists are produced, so the Freshness Gate does not fire; recency comes from the market-intel responses themselves.

The "trend" constraint (read this before drafting)

There is no historical time-series tool. A genuine "up / down / rising" story therefore needs two measured points. This skill gets them one of two ways only:

• Cross-sectional contrast (default, no persistence needed): the contrast is between two slices measured in the SAME run — vertical A vs vertical B, deal type X vs deal type Y, or a recency window vs the all-time slice of the same vertical. This is an honest "as of now" comparison, never a claim about change over time.
• Snapshot delta (only if a prior snapshot exists): if the workspace holds a prior storyline snapshot for the same slice, a true over-time delta can be computed and stated as such. With no prior snapshot, over-time language is banned — the story is framed cross-sectionally or not at all.

Never write "up since last quarter / trending / on the rise" unless a stored prior snapshot backs it. A within-run contrast is phrased as a contrast ("Beauty's median sits below Fitness's", "this window is quieter than the vertical's all-time rate"), not as motion.

Inputs to collect

• Verticals to sweep (default: a standard set of the densest corpus verticals, e.g. Beauty, Fitness, Fashion, Food, Tech/Gaming — confirm once). More verticals = more calls; warn on cost in thorough mode.
• Angle (optional) — "pricing", "deal-type mix", "where's the momentum". Shapes which cuts to pull; default is a broad sweep across all three.
• Channel (default: both) — newsletter, linkedin, or both. Sets draft length and voice, not the data.
• Snapshot store location (optional) — workspace path where storyline snapshots live (default storyline-snapshots/<slice>.json). Only needed if the user wants over-time deltas; absent it, the skill runs cross-sectional.
• Mode — default thorough; thrifty on request.

If this is a recurring content run, reuse prior inputs without re-asking and offer to schedule it via the harness's scheduling facility if one exists.

Flow

All market-intel calls are wrapped in Refusal Recovery (market floor 5 brands / 25 deals; rate floor 10 brands / 50 deals). 5 credits each, including each ladder retry. A thin slice that refuses is itself a finding ("we don't yet have floor-clearing volume in ") — it never becomes a fabricated number, and a refused slice is simply not eligible to anchor a story.

1. Per-vertical landscape — for each vertical in the sweep: query_market_intelligence { mode: "market", vertical: <v> } -> deal counts, deal-type mix, company-type spread.

2. Per-vertical pricing — for each vertical: { mode: "rate", vertical: <v> } -> p25 / median / p75 band. Optionally add creator_tier (emerging <1k / nano 1k-10k / micro 10k-100k / mid 100k-500k / macro 500k-1M / mega 1M+) to mine a tier-specific contrast (e.g. "micro Beauty median up, macro flat"); omit it (the default) for the vertical-wide band, and never diff a tiered band against a vertical-wide one in the same story.

3. Recency slice (thorough) — for each vertical: { mode: "market", vertical: <v>, active_since: "<recent window start>" } -> how active the vertical has been lately, for "where's the momentum" angles. Ladder by widening the window if it refuses; label the section with the window that actually cleared.

4. Snapshot I/O (only if over-time deltas requested) — read any prior snapshot for each slice before drafting; write this run's aggregates + clearance levels + run date after. Free (workspace I/O, 0 credits). A metric that cleared at a different level than last run is not comparable — flag it, don't diff across scopes.

5. Mine contrasts (no tool calls) — scan the assembled aggregates for the sharpest legitimate contrasts: volume vs value divergence, deal-type shift, one vertical against another, a recency window against the all-time slice. Rank candidate stories by how surprising AND how floor-clean they are.

6. Draft (no tool calls) — write the top stories into channel-appropriate copy with provenance attached.

Pre-run estimate fires before step 1 when the sweep's planned calls exceed ~30 credits (e.g. 5 verticals x 3 calls x 5 = 75 credits -> say so first).

Thrifty: steps 1 + 2 only, fewer verticals, no recency slice, max 2 ladder rungs anywhere — cross-sectional stories only.

Deliverable

# Data storylines — <date>
*Source: Creatorland deal corpus · sweep: <verticals / cuts covered>*

## Ranked story candidates
1. **<headline contrast>** — <one line: the two slices and the gap>.
   Strength: <surprising + floor-clean>. Framing: <cross-sectional / over-time>.
2. ...
<Each candidate names the exact slices compared and whether it is a within-run
contrast or a snapshot-backed delta.>

## Drafts

### Newsletter — "<working title>"
<2-4 short paragraphs. Every stat inline-cited: stat + the provenance line the
tool returned + the recency window. No over-time verb unless snapshot-backed.>

### LinkedIn — "<hook>"
<Hook line + 3-5 short lines + soft CTA. Same citation discipline; provenance
can sit in a closing "source" line rather than inline, but never omitted.>

---
**Benchmark basis:** <clearance level per slice used; floor-disclosure note
where any query was broadened — "a privacy feature of the data source, not
missing data">. Slices that refused are listed here as "below floor — not used
as a story anchor".
**Provenance:** <provenance lines from each response used, verbatim>
· recency windows as noted.
**Framing note:** <if any draft uses over-time language, the prior-snapshot date
that backs it; otherwise "all contrasts are point-in-time / cross-sectional">.
Credits used this run: ~N (breakdown: <calls>x5)

Honesty rules

• Contrast, not motion, by default. With no prior snapshot the story is a within-run comparison; never dress a cross-sectional gap as a trend. "Up / rising / since last quarter" requires a stored prior snapshot and says which.
• Refused slices are findings, never anchors. A slice below the privacy floor is reportable as such; it cannot anchor a story and is never padded to a number. Disclose it in the benchmark basis.
• Bands are vertical-level market bands (convention 2) — never attached to a creator, and never implied to be one creator's rate even in a punchy hook.
• Provenance ships with the post. The whole point is defensible content: every published stat carries its citation, even in LinkedIn copy. Never strip it for polish (convention 1).
• Clearance level is part of the claim. If a contrast only holds at vertical level (not sub-category), the copy says so rather than implying sub-category precision (convention 2).
• No contact info anywhere in drafts (convention 7).
• Interpretation is labeled. A "what this might mean" line is opinion and is marked as such; the data carries the headline, the read carries the asterisk.

Credit footprint

thorough: ~45-75 credits for a 5-vertical sweep (3 calls/vertical x 5; ladder retries add 5 each) — pre-run estimate fires · thrifty: ~20 credits (2 calls x fewer verticals, capped ladder). Snapshot read/write is free (0 credits).