narrative-tracker

Read memory/MEMORY.md for context on prior narrative observations. Read the last 3 days of memory/logs/ — specifically any prior ### narrative-tracker entries — to (a) avoid re-reporting the same narratives without new info, and (b) detect phase transitions vs the last run.

Goal

Produce a decision-grade narrative map: every narrative gets a mindshare score, a velocity arrow, a sentiment tag, named drivers, and an explicit position call. Classification without a position call is noise.

Steps

1. Ingest signals

a. X/Twitter narratives - X.AI API (primary). The primary signal is a direct curl to the X.AI Responses API with Grok's x_search tool - see ## Fetching below for the full contract. XAI_API_KEY is injected into this skill's environment via requires:; it is present and valid, so this path is required whenever the key check gives KEY_PRESENT. Set the Bash tool timeout to ≥180000 when running the curl (x_search takes 30-120s) and capture the HTTP status.

Shell shape. Each Bash call is one allowed command (pipes are fine). The run refuses a call that sets a variable (FROM_DATE=$(date ...), PROMPT="...", HTTP=$(./secretcurl ...)), uses [ ... ] tests, or joins commands with ;, && or ||, and each refusal costs a model turn. So run each block below as its own Bash call, in order: get a date with a bare date call and write the printed value into later commands (the <FROM_DATE> / <TO_DATE> placeholders); put the prompt and dates inline in the jq -n filter (a jq $name variable such as --arg p ... '$p' is refused too) and pipe it straight into ./secretcurl; never put a $VAR in the ./secretcurl line.

Get the window dates, each date as its own call:

date -u -d '3 days ago' +%Y-%m-%d

date -u +%Y-%m-%d

The first prints FROM_DATE, the second TO_DATE.

Key check: read the "Keys set for this run" line in your run notes, never a shell test (those are refused and cost a turn). XAI_API_KEY on that line = KEY_PRESENT, not on it = KEY_UNSET.

Build the JSON payload with jq (do NOT hand-assemble it), with the dates written in literally, and pipe it straight into the POST in the same call. Do not write it to a file first: a > redirect into a file is refused in BOTH tiers, /tmp included, and this read-only run has no Write tool. ./secretcurl reads the body from stdin (--data-binary @-; its own config goes over a separate pipe, so stdin is free):

jq -n '{model:"grok-4.7", input:[{role:"user",content:"Search X for the dominant crypto and tech narratives from <FROM_DATE> to <TO_DATE>. Return 12-15 distinct narrative threads. For each: 1) short label, 2) 3-5 representative @handles driving it, 3) 2-3 tweet permalinks, 4) rough mention-volume descriptor (niche / growing / saturating / cooling), 5) the strongest one-line bear case against it."}], tools:[{type:"x_search",from_date:"<FROM_DATE>",to_date:"<TO_DATE>"}]}' \
  | ./secretcurl -s -o /tmp/xai-nt.json -w '%{http_code}' --max-time 150 -X POST "https://api.x.ai/v1/responses" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer {XAI_API_KEY}" \
  --data-binary @-

Run that block verbatim apart from the placeholders (do not hand-reassemble the JSON - the jq -n builder exists precisely so quoting/expansion can't break). The curl prints the status code (HTTP), and ./secretcurl prints a secretcurl xai attempt=... http=<code> bytes=<n> ... line on stderr; that xai ... http= line must appear in your output - it is your proof the call ran. On HTTP=200 with a non-empty body, parse /tmp/xai-nt.json with jq -r '.output[] | select(.type == "message") | .content[] | select(.type == "output_text") | .text' and use that as the primary narrative signal (SOURCE=api).

b. WebSearch / WebFetch fallback (last-resort only). You may reach for this only after you have shown a secretcurl xai ... http=<code> line proving Path A actually ran and returned a non-2xx code (or an empty body, or timed out). If you have no xai ... http= line, you did not run the call - go back and run it. Reach for the fallback only when Path A genuinely fails - KEY_UNSET, a non-2xx HTTP code, an empty parse, or a timeout. It is lower quality (WebSearch favours old high-engagement posts) and is never co-equal with Path A. Record the fetch failure in your final output with the true reason - key-unset | http-<code> | empty | timeout - never "XAI_API_KEY unavailable" when the key was set (a slow curl is a timeout, not a missing key). Then compile narratives via WebSearch (crypto narrative ${TO_DATE}, AI agent crypto trend this week) and WebFetch on individual tweet URLs; discard anything older than the 3-day window.

c. Quantitative reference points (supplement). Independently of the fetch path, cross-check mindshare against external quantitative benchmarks with one WebSearch: DefiLlama narrative tracker OR Kaito mindshare leaderboard. Pull 1-2 concrete numbers (project name, metric, link) to calibrate the mindshare scores in step 2. This is a calibration cross-check, not a narrative source. Do not paraphrase — extract facts.

d. Memory diff. Extract narrative labels mentioned in the last 3 days of ### narrative-tracker log entries. You'll compare against them in step 4.

2. Score each narrative

For each distinct narrative (merge near-duplicates aggressively — "AI agents" and "agentic crypto" are the same), assign:

Drop any narrative that ends up IGNORE unless it's structurally important — noise reduction is the goal.

3. Detect transitions

Compare today's narratives to the last 3 days of logs:

• NEW — narrative wasn't in prior logs at all
• PROMOTED — phase moved up (e.g. Emerging → Rising)
• DEMOTED — phase moved down
• DEAD — was in prior logs, now absent from all signals

Transitions are the highest-value output — the point of a daily tracker is to catch inflection points, not re-report the zeitgeist.

4. Flag reflexivity

For each narrative, flag if the story itself is moving outcomes:

• Token prices moving on narrative alone (no fundamentals shift)
• Projects rebranding/pivoting to ride the narrative
• VCs publicly endorsing to manufacture legitimacy
• Prediction markets or on-chain flows reflecting narrative belief

Only flag explicit cases with a concrete example. "Reflexivity" without evidence is hand-waving.

5. Format the notification

Keep under 4000 chars. Lead with transitions and reflexivity — those are the decisions. Classification goes below.

*Narrative Tracker — ${today}*

TRANSITIONS
• NEW: <label> — <why it matters> — <link>
• PROMOTED: <label> Rising → Peak — <what flipped>
• DEMOTED: <label> Peak → Fading — <what cooled>
• DEAD: <label> — gone

REFLEXIVITY ALERT
• <narrative> — <concrete evidence the story is moving outcomes>

POSITIONS
• FRONT-RUN: <label> (mindshare 2 ↑↑, Bull) — <driver> — <bear case> — <link>
• RIDE: <label> (3 ↑, Bull) — <driver> — <bear case>
• FADE: <label> (5 → Cope) — <driver> — <reflexivity note>

MAP
Emerging: <labels>
Rising: <labels>
Peak: <labels>
Fading: <labels>

If absolutely nothing new or notable (no transitions, no reflexivity, no FRONT-RUN/FADE calls): send a one-line update instead of the full template — *Narrative Tracker — ${today}*: no phase transitions, map unchanged from <last_date>.

6. Send via ./notify

7. Log record

This skill is read-only, so the workflow's read-only guard writes its ### narrative-tracker log entry from your captured output; a self-written entry would be a duplicate. Don't append to memory/logs/ yourself - put this record in your final output: the full structured output (not just the notification - include all narratives considered, even IGNOREd ones, so future diffs work). If a full run produced nothing actionable, record NARRATIVE_TRACKER_OK with the narrative labels seen (so tomorrow's diff still has a baseline).

Guidelines

• Quantitative over vibes. Every narrative gets mindshare 1-5 and a velocity arrow — no exceptions. If you can't score it, drop it.
• Transitions > classification. A daily tracker's value is catching moves, not listing the weather.
• Named drivers only. "Crypto Twitter is excited about X" is not a driver. "@handle + @handle + @fund" is.
• Position calls are mandatory for Emerging/Rising/Peak narratives. If signals are genuinely ambiguous or contradictory, WATCH is an acceptable call — but never omit a position entirely and never invent conviction you don't have.
• Ruthless dedup. Same narrative under two labels = one narrative. Merge, don't split.
• Call out cope. Manufactured narratives, coordinated shilling, and dead-cat bounces get tagged explicitly.
• Prioritize topics tracked in MEMORY.md over generic market chatter.

Fetching

XAI_API_KEY is injected into this skill's environment (declared in requires:). It is present and valid. The primary fetch path is a direct curl to https://api.x.ai/v1/responses with Authorization: Bearer {XAI_API_KEY} (model grok-4.7, "tools":[{"type":"x_search"}]). There is no network sandbox blocking this - earlier versions of this skill claimed there was, and that is stale and false. Just make the call.

Rules:

1. Check, don't assume. Read the "Keys set for this run" line in your run notes, never a shell test (those are refused and cost a turn): XAI_API_KEY on that line = KEY_PRESENT, not on it = KEY_UNSET. If KEY_PRESENT (it will be), Path A (the curl in step 1a) is required before any fallback.
2. Allow enough time. The x_search call typically takes 30-120s. When you invoke the Bash tool for the curl, set the tool's timeout to at least 180000 (180s) and keep --max-time 150 on the curl so it fails cleanly rather than hanging. A slow curl is not a missing key — never treat a timeout as key-unavailable.
3. Capture the HTTP status so the fallback decision is fact-based (see the skeleton in step 1a). HTTP=200 with a non-empty parsed body → use it.
4. Fall back only on a real failure, recording the true reason — key-unset (only if the key check gave KEY_UNSET), http-<code> (non-2xx), empty (200 but nothing parsed), or timeout. Never write "XAI_API_KEY unavailable" when the key was set.

WebSearch / WebFetch are last-resort fallbacks only — lower quality, never primary or co-equal. Do not reach for them while the key works.

Environment Variables Required

• XAI_API_KEY — X.AI API key for Grok's x_search tool. Declared in requires:, so it is injected into this skill's environment and is the primary fetch path for the narrative signal (step 1a). If it is ever unset, the skill degrades to WebSearch/WebFetch at lower quality.
• Notification channels configured via repo secrets (see CLAUDE.md).