Instagram Trend Scout

Stage: S1 - Research (first skill in the Instagram Trend Post Builder flywheel)

This skill answers one question: what is working on Instagram for this niche, right now? It gathers public signals through web search - it does not log into Instagram or call any private API - and returns a structured, machine-readable scouting report that the next skill (trend-educator) turns into a plain-language briefing.

Be honest about confidence. Web search surfaces secondhand signals (roundup articles, creator breakdowns, public engagement screenshots), not the Instagram Graph API. Label every number as an estimate and never invent precise metrics.

Input Schema

Workflow

1. Validate input. If keyword is missing or under 2 characters, stop and ask the user for a niche. If time_range is not one of the allowed values, default to "7d" and note the substitution.
2. Build the search plan. Construct 4-6 web_search queries that triangulate the niche from different angles. Always include recency terms derived from time_range (e.g. for "7d" use "this week" / "2026"). Template queries:

• best performing Instagram Reels {keyword} {recency}
• viral Instagram carousel {keyword} {recency}
• Instagram hook ideas {keyword} that get views
• {keyword} Instagram engagement trends {recency}
• top {keyword} creators Instagram {recency}
• (if format_focus is set, replace the format word accordingly)

1. Run searches and collect raw signals. Execute each query with web_search. For each promising result, capture: the content format, the topic/angle, the hook (first line or on-screen text if quoted), any caption-length cues, any hashtags mentioned, and any engagement numbers stated (views, likes, saves, shares).
2. Normalize each piece of content into a top_content row. Where a hard number is unavailable, set engagement_signal to a qualitative band ("low", "medium", "high", "viral") inferred from how the source describes it, and set source_url so the claim is traceable.
3. Cluster formats. Group the collected content by format + structural pattern (e.g. "talking-head Reel with text hook", "5-slide problem→solution carousel"). Rank clusters by how often they recur across sources. Emit the top 3-5 as winning_formats.
4. Extract hook patterns. Pull recurring opening-line structures (e.g. "POV:", "Nobody tells you…", "3 mistakes…"). Generalize each into a reusable pattern template plus a concrete example. Emit as winning_hooks.
5. Find the gaps. Identify angles, audience questions, or sub-topics that the niche audience clearly cares about but that few creators are covering well. Emit as content_gaps with the evidence that points to each gap.
6. Set the benchmark. Synthesize an engagement_benchmark describing what "good" looks like in this niche this period, as bands (not false precision), plus a confidence rating reflecting how much corroborating data was found.
7. Assemble and return the full output object below. Pass it forward to trend-educator.

Output Schema

{
  "keyword": "string - echoes the input niche",
  "time_range": "string - the effective time range used",
  "scanned_at": "string - ISO date the scout was run",
  "top_content": [
    {
      "rank": "number - 1-based, most relevant first",
      "format": "string - one of reel | carousel | single_image | story",
      "topic": "string - the angle/subject of the piece",
      "hook_text": "string - the opening line or on-screen hook, quoted if available else paraphrased",
      "hook_style": "string - the structural pattern, e.g. 'POV', 'listicle', 'contrarian', 'before-after'",
      "caption_length": "string - one of short | medium | long (short <125 chars, medium 125-500, long >500)",
      "hashtags": ["string - observed or inferred hashtags for this piece"],
      "engagement_signal": "string - one of low | medium | high | viral (qualitative band)",
      "source_url": "string - where this signal was found"
    }
  ],
  "winning_formats": [
    {
      "format": "string - the recurring structural format",
      "frequency": "string - one of rare | common | dominant (how often it recurred across sources)",
      "why_now": "string - short note on why this format appears to be surfacing this period"
    }
  ],
  "winning_hooks": [
    {
      "hook_style": "string - the pattern name",
      "pattern": "string - a reusable fill-in-the-blank template",
      "example": "string - a concrete example seen in the niche"
    }
  ],
  "content_gaps": [
    {
      "gap": "string - an underserved angle or question",
      "evidence": "string - why this looks like a gap (demand without good supply)"
    }
  ],
  "engagement_benchmark": {
    "median_engagement_band": "string - e.g. '3-6% engagement rate' or 'unknown'",
    "saves_signal": "string - qualitative note on what drives saves in this niche",
    "comments_signal": "string - qualitative note on what drives comments",
    "confidence": "string - one of low | medium | high, based on corroboration"
  }
}

Output Format

Present the report to the user as a scannable markdown brief, then hand the full JSON object to trend-educator:

• Header line: 🔎 Trend Scout - {keyword} (last {time_range})
• Top Content as a table with columns: Rank · Format · Topic · Hook · Signal · Source.
• Winning Formats, Winning Hooks, Content Gaps as bulleted lists.
• Engagement Benchmark as a short callout block, leading with the confidence level.
• Close with: → Piping this to trend-educator to explain why these are working.

Error Handling

• No keyword / too vague: Do not guess a niche. Ask: "What niche or audience should I scout? (e.g. 'home barista', 'B2B SaaS founders')" and wait.
• No data found: If searches return nothing usable, return top_content: [], set engagement_benchmark.confidence to "low", and tell the user plainly: "I couldn't find recent public signals for this niche. Try a broader keyword or a longer time_range." Offer 2-3 adjacent keywords to retry.
• Search blocked / rate-limited: If web_search is unavailable or throttled, do not fabricate results. Report the failure, return whatever partial signals were gathered with confidence: "low", and suggest the user re-run later or supply example posts manually.
• Thin / single-source data: If only one source supports a claim, still include it but mark engagement_signal conservatively and lower confidence. Never upgrade a qualitative band into a fake precise metric.
• Stale results: If the freshest results predate time_range, surface that mismatch ("best available data is ~6 weeks old") rather than silently presenting it as current.

Flywheel Connections

• Fed by: _(entry point - this is where the weekly cycle begins; the user supplies a niche)_
• Feeds into: trend-educator - consumes this entire output object as its input and translates the data into a plain-language briefing.
• Part of pack: Instagram Trend Post Builder (S1 → S2 → S3).