Vertical Forecast Brief

Vertical Forecast Brief

A weather-report framing for a vertical: each run re-pulls the corpus aggregates and, if a prior snapshot exists in the workspace, reports the movement as "conditions" — what's warming, cooling, or flat. The honest constraint baked into this skill: the MCP has no historical time-series tool, so deltas are computed client-side by storing each run's snapshot in the harness workspace and diffing against it. It reports what the active_since-windowed corpus comparison actually shows — it does not fabricate trend predictions. For a one-shot portrait, use vertical-briefing.

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

Inputs to collect

• Vertical (required) — e.g. "Fitness". Reuse on recurring runs.
• Comparison window (default: trailing period since last run) — sets the active_since cuts used to read current activity. If no prior snapshot, this is the baseline run (no deltas — say so).
• Snapshot store location — the workspace path where snapshots live (default a skill-owned file, e.g. forecast-snapshots/<vertical>.json). If the harness has no durable workspace, the skill runs as a baseline-only portrait and states it can't compute deltas without persistence.
• Mode — default thorough; thrifty on request.

Flow

All market-intel calls wrapped in Refusal Recovery (market floor 5/25, rate floor 10/50). 5 credits each incl. retries.

1. Load prior snapshot — read the stored snapshot for this vertical from the workspace (no credits). If none, this is a baseline run.

2. Pull current aggregates — the same call set each run so snapshots are comparable:

   • { mode: "market", vertical: <vertical> } — volume, deal-type mix.
   • { mode: "market", vertical: <vertical>, active_since: "<window start>" } — recent activity (the "recency signal").
   • { mode: "rate", vertical: <vertical> } — pricing band. Optionally scope the pricing band to a tier by adding creator_tier (emerging <1k / nano 1k-10k / micro 10k-100k / mid 100k-500k / macro 500k-1M / mega 1M+); omit it (the default) for the vertical-wide band. Comparability across snapshots requires holding tier constant — pick tiered or vertical-wide once and keep it the same every run, or the delta is not meaningful. On any refusal, ladder and record the clearance level; a metric that cleared at a different level than last run is NOT comparable — flag it as "scope changed, delta not meaningful" rather than diffing across scopes. Thrifty: the two market calls only; max 2 ladder rungs.

3. Diff vs prior (no tool calls) — for each metric measured at the SAME clearance level both runs, compute the delta and the direction. Label each as: measured delta (same scope both runs) or not comparable (scope/floor changed). Direction words ("warming") are a reading of the recency-window numbers — labeled as a directional read, never a forecast of the future.

4. Store the new snapshot — write this run's aggregates + their clearance levels + run date to the workspace for next time (no credits).

5. Compose the conditions report. Pre-run estimate fires only if extra cuts push over ~30 credits.

Deliverable

# Conditions report — <Vertical>
*Run <date> · vs prior snapshot <prior date or "baseline — no prior">*

## Conditions
<Per metric, one line: current value, prior value, delta, direction word.
Only for metrics measured at the SAME clearance level both runs.>
- Deal volume: <cur> (was <prior>, <+/−X%> — <warming/cooling/flat>)
- Recency signal (<window>): <cur> active deals (was <prior>, <delta>)
- Median rate: $<cur> (was $<prior>, <delta>) · band p25–p75 $<x>–$<z>

## Not comparable this run
<Metrics whose clearance level changed between runs — reported, not diffed,
with the reason (e.g. "the active_since slice cleared at vertical level this
run vs sub_category last run").>

## Reading
<2–3 sentences interpreting the conditions. Every sentence labeled as a
directional read of the current data, NOT a prediction. No forecast of where
the vertical "will" go — the corpus doesn't support that claim.>

---
**Method note:** the MCP has no time-series endpoint; deltas are computed by
this skill diffing stored snapshots — directional reads describe movement
between two measured points, not a forecast.
**Benchmark basis:** <clearance level per metric; floor-disclosure note
where broadened — "a privacy feature of the data source, not missing data">
**Provenance:** <provenance line per response, verbatim> · windows as noted.
Credits used this run: ~N (breakdown: <calls>×5)

Honesty rules

• No fabricated forecasts. The name says "forecast" but the data is retrospective. The skill reports measured deltas between snapshots and labels any directional language ("warming") as a read of current numbers, never a prediction of future ones. Never write "will rise / is trending toward" as fact.
• Only diff like-for-like. A metric is comparable only if it cleared the privacy floor at the SAME scope both runs; otherwise it goes in "not comparable," not a misleading delta.
• Deltas need two snapshots. A baseline run has no deltas and says so — never invent a prior to manufacture movement.
• Bands are vertical-level market bands (convention 2), never a creator's.
• Provenance + method note survive every run (convention 1).
• No contact info (convention 7).

Credit footprint

thorough: ~15–25 credits (3 base calls + ladder retries, ×5) · thrifty: ~10 credits (2 market calls, capped ladder). Snapshot read/write is free (workspace I/O, 0 credits).