rohunvora/trade

/trade

0 - Intro

Think through trades live. The user is watching the work, not just the final card. Narrate what changed your mind, what has no clean expression, and why one instrument beats another.

Supporting docs: skill/references/ (skill index, ASCII map, CLI cheatsheet, routing decision rules, event types, trade data index, Hyperliquid thematic universe, prediction markets, search API).

context

When the user is thinking through a trade idea -- not pasting a URL -- query paste.trade for context first:

curl -s -H "Authorization: Bearer $PASTE_TRADE_KEY" "https://paste.trade/api/search?ticker=GOLD&limit=5"

See skill/references/search-api.md for all query modes, response shape, and how to present results.

If the query returns 0 results, suggest the user add data -- every gap is a contribution prompt. "Nobody's tracking AAPL on paste.trade yet. Want to be the first?"

update

If the input is exactly "update": pull the latest version of the skill repo, show what changed, and stop.

1. Find the skill repo root (the directory containing this SKILL.md).
2. Run git -C <repo_root> pull origin main.
3. Run git -C <repo_root> log --oneline -5 to show recent changes.
4. Tell the user what updated (SKILL.md, scripts, references, etc.) and stop. Do not continue to the trade pipeline.

add to source

If the input contains a paste.trade/s/ URL plus a thesis or trade idea: load references/add-to-source.md and follow that flow. Do not run the normal extraction pipeline.

1 - Defaults

• $100K risk capital, max upside
• Supported venues: Hyperliquid, Robinhood, Polymarket
• Prefer Hyperliquid when available
• Best single trade per thesis
• No em dashes in output
• End every response with: Expressions, not advice. Do your own research.

2 - Chat UX

• Keep chat updates operational and brief.
• First status line should set expectation: Running /trade now. I will send a live link shortly.
• For transcript sources, next status line should set duration expectation: On it. Pulling transcript now. Longer videos can take a few minutes.
• After create-source.ts, send Watch live: {source_url} immediately unless your runtime already delivers it.
• Continue the pipeline after the live link. Never treat the live link as the end of the /trade run.
• Do not wait for user input before continuing.
• If your runtime surfaces progress in chat, send updates when the state changes. Do not hold them and dump them at the end.

Preflight

Before running any script, verify bun is available (command -v bun). If missing, install it: curl -fsSL https://bun.sh/install | bash, then restart the shell.

Run before every /trade:

bun run skill/scripts/onboard.ts
# Returns: { status, env_path, keys, handle?, profile_url? }

If status is "ready", continue to §3.

If status is "onboarding":

• Greet the user with their handle and paste.trade profile link.
• For each key with status: "missing", share the hint text.
• For missing items with an install_command, offer to run it (e.g., "Want me to install yt-dlp?").
• Do not gate on optional keys. Tell the user they can add them later and continue.
• Offer to save keys now: "Want to add any of these? You can paste each key and I'll save it to your .env."
• If the user pastes a key in reply, append KEY=value to the env_path from onboard output. Confirm what was saved and where. Never echo the full key value back in chat.
• Re-run onboard.ts after saving to verify the key is detected.
• Then continue to §3 with the original /trade input.

If status is "failed", stop and show the error. Do not proceed without a working PASTE_TRADE_KEY.

Core Loop

3 - Classify

• URL source: extract first.
• PDF source: read the PDF yourself (do not pass to extract.ts — it cannot parse PDFs). Use the text you read as the source artifact for thesis extraction.
• User-typed thesis: skip extraction. Save the user's exact words in quotes[]. Choose headline_quote as the sharpest exact substring of the user's words. If it's <120 characters, can include whole thing. Put the cleaned interpretation in thesis.
• If URL is paste.trade/s/:id or paste.trade/t/:id, treat as normal source input.

4 - Extract

Primary extraction:

bun run skill/scripts/extract.ts "URL"
# Returns: { source, word_count, saved_to, title?, published_at?, channel_handle?, description?, duration_seconds?, image_files? }
# YouTube: transcript omitted from output; read the file at saved_to.
# Tweet images: downloaded to local files listed in image_files[]. Read them for visual context.

Create the source page as soon as you know the source metadata:

bun run skill/scripts/create-source.ts '{ "url": "...", "title": "...", "platform": "...", "author_handle": "...", "author_avatar_url": "...", "source_date": "...", "source_images": [...], "body_text": "...", "word_count": N, "duration_seconds": N, "speakers_count": N }'
# Returns: { source_id, source_url, status: "processing", run_id }

Execution sequence

1. Run extract.ts.
2. If image_files are present in the output, read them now. Charts, screenshots, and diagrams are critical source context — use them to inform thesis extraction, ticker identification, and derivation reasoning. Describe what you see.
3. If YouTube with multiple speakers with competing or independent market views (panels, debates, co-hosted roundtables — not single-guest interviews) and GEMINI_API_KEY is missing: ask the user now — before creating the source. Offer to paste a key or continue without speaker attribution.
4. Run create-source.ts. Send the Watch live link immediately unless your runtime handles delivery, then continue the pipeline in the same run. Do not stop or wait for user input. (See §2 Chat UX.)
5. Do NOT read the saved_to file before this point.
6. Only after source creation, run enrichment, transcript reads, and uploads.

Notes

• author_handle here means the source publisher/channel handle.
• YouTube uses channel_handle, not a guest speaker.
• author_avatar_url is the author's profile picture URL from extract output. Always include it — the source page and feed cards display it immediately.
• body_text is the full original source text (e.g., the complete tweet, article body). Always include it — the source page displays it verbatim.
• word_count, duration_seconds, speakers_count are optional extraction metadata for the live stats bar.
• Save run_id and thread it through every later adapter call for this source run.
• If the prompt includes internal tracing metadata (run_id=...), pass that value as run_id in the create-source.ts payload.
• Use the canonical live-link line from Chat UX.

Status update payload shape:

bun run skill/scripts/status.ts <source_id> '{ "event_type": "status", "data": { "message": "..." } }'

5 - Enrich

Timing

Runs after the source page exists and the user has a live link. Runs before thesis extraction.

Metadata

• Check extraction output for missing author_handle, source_date, title.
• If author missing: scan extracted text for byline patterns, then web search URL/title to find author and X handle.
• If source_date missing: scan text for date indicators, web search, or "now" as last resort. For user-typed theses, always use "now". Scripts resolve it to actual current time. Never guess a time like noon UTC.
• Enriched metadata is used in trade posts (source page author stays as-is).

Dense source enrichment

→ Read skill/references/dense.md for diarization, speaker identity, and transcript handling. Sparse sources skip to §6.

• Avatars not in scope: backend auto-resolves via ensureAuthor + enqueueAssetJob.

Push enriched metadata:

If enrichment resolved new metadata (author handle, source date, speakers, or thumbnail), push it now so the source page updates before thesis extraction:

bun run skill/scripts/update-source.ts <source_id> --run-id <run_id> '{ "author_handle": "...", "source_date": "...", "thumbnail_url": "...", "speakers": [...] }'

6 - Theses

Core

Read the canonical source artifact and find every tradeable thesis.

A thesis is a directional belief about what changes and what that means for price.

Extraction

• Dense source (podcast, article, PDF): → read skill/references/dense.md for three-pass extraction, thesis map, parallelization, and chunking.
• Sparse source (tweet, user thesis, screenshot): → read skill/references/sparse.md. Handles extraction through routing (§6-§9). Resume at §10 Post.

Both paths use the thesis schema and save commands below.

{
  "thesis": "author's directional belief in one sentence, in your words not theirs",
  "horizon": "author's timing language, if any",
  "route_status": "unrouted",
  "unrouted_reason": "pending_route_check",
  "who": [
    { "ticker": "NVDA", "direction": "long" },
    { "ticker": "AI infrastructure companies", "direction": "long" }
  ],
  "why": ["reasoning step from author", { "text": "researched fact", "url": "...", "origin": "research" }],
  "quotes": ["exact words from source that anchor the thesis"],
  "headline_quote": "the single best card-ready exact quote or exact substring from quotes[]. Choose the sharpest self-contained claim. Must read cleanly out of context, avoid dangling setup, target <=100 chars when possible, hard cap 120. Frozen at extraction. post.ts validates exact match or exact substring.",
  "source_date": "ISO 8601 datetime with time (e.g. 2026-03-10T14:30:00Z), or \"now\" for user-typed theses. Scripts resolve \"now\" to actual current time. Date-only resolves to midnight UTC → wrong price.",
}

Who field

who captures 1-3 trade ideas per thesis. These are starting points for routing, not final selections. Can range from specific tickers to broad descriptions. During routing, who is overwritten with the final selected expression.

A thesis is one belief. If the same belief could be traded through different instruments, those are who entries, not separate theses.

Include the most direct expression of the thesis alongside any specific ticker names.

For unresolved candidates, do not drop them. Save them as:

{
  "thesis": "...",
  "route_status": "unrouted",
  "unrouted_reason": "no clean liquid instrument / weak directional expression / evidence gap",
  "who": [],
  "why": ["..."],
  "quotes": ["..."],
  "headline_quote": "..."
}

Save and parallel

Save all theses from extraction in one batch call (pass --total on first save if using individual saves instead):

# Preferred: batch save all theses at once:
echo '[{...}, {...}]' | bun run skill/scripts/batch-save.ts --run-id <run_id> --total 5
# Returns: [{ id, index }, ...]

# Individual save (when extracting one at a time):
bun run skill/scripts/save.ts --run-id <run_id> --total 5 '<thesis JSON>'
# Returns: { id, file, count }

# Update a saved thesis (used during routing):
echo '<partial JSON>' | bun run skill/scripts/save.ts --run-id <run_id> --update <id>

Track the returned thesis IDs. You need every one for finalization.

Before starting research, narrate the transition so the live page stays active: bun run skill/scripts/stream-thought.ts --run-id <run_id> "Researching market context..."

Save and post return {"ok": false, "error": "..."} on validation errors (exit 0), so parallel calls are safe -- one failure does not cancel siblings. Always check the ok field (or presence of error) in tool output before proceeding. Do not use routing difficulty as a filter at extraction time. Capture first, then route or explicitly mark unrouted.

7 - Research

Sparse sources: §7-§9 are handled in skill/references/sparse.md. Skip to §10.

For each thesis, determine the best executable expression on supported venues. On adapter error, retry the failed step once. If it fails again, try an alternative ticker or skip the thesis.

Venues

Supported venues:

• Hyperliquid
• Robinhood
• Polymarket

Parallel steps

1. Research (run in parallel):
   • Web search: verify the thesis holds today, find developments, and research tradeable instruments for the ideas in who. Your training data is stale for tickers and listings. Search to find what's actually available. Cite findings in why as { "text": "...", "url": "...", "origin": "research" }.
   • Instrument discovery (skill/scripts/discover.ts --query "<keywords>"): search available instruments across all venues (Hyperliquid + Polymarket) using terms from who. Works best with single concrete terms, not multi-word abstractions. Use --catalog for a full listing of non-crypto HL instruments. For HIP-3/non-crypto results, prefer entries whose reference_symbols and routing_note show the same ETF, benchmark, commodity, or private company the thesis is really about.
   • Source context (skill/scripts/source-excerpt.ts --run-id <run_id> --file <saved_to> --query "<thesis keywords>"): retrieve surrounding context from the original source for this thesis. After extraction splits a source into theses, adjacent details get lost. Use this to find what the author said around each claim: qualifications, supporting numbers, competitive landscape, or nuance that strengthens the derivation. Also use --around "<exact quote>" to expand a specific quote.
2. Route (skill/scripts/route.ts): validate the best candidates from both sources against supported venues and get pricing. Takes ticker symbols only.
3. Select and save: pick the expression with the tightest link between the source quote and the instrument. The trade ideas in who are starting context, not decisions. Routing may confirm them, improve on them, or find something better entirely. Prefer sector-level instruments over single equities for broad theses. Persist via save.ts --update.

Venue upgrades

• ETFs and broad-sector stocks: Hyperliquid often has a thematic index or commodity perp tracking the same underlying with leverage and no ETF overhead. Run discover.ts --query "<theme>" to check. See skill/references/hl-universe.md.
• Event-driven theses: Polymarket may have a binary contract that directly prices the catalyst. Run discover.ts --query "<event keywords>" to check. See skill/references/prediction-markets.md. If discover.ts returns zero PM results for a thesis, do not route to Polymarket.
• If a better venue exists, route there and present the original as an alternative.

Requirements

• If a thesis is executable on both Hyperliquid and Robinhood, prefer Hyperliquid.
• If best trade is not one of the initially considered direct tickers, update thesis with explicit proxy reasoning and citations.
• Before final route, check quote-to-trade logic: if original author would not recognize the link, reroute.

Directness

• direct: original author would recognize this as their trade.
• derived: author did not name it, but market link is immediate and defensible.

Route evidence

{
  "route_status": "routed",
  "who": [{ "ticker": "SMR", "direction": "short" }],
  "route_evidence": {
    "subjects": [{ "label": "NuScale Power", "subject_kind": "company" }],
    "direct_checks": [
      {
        "subject_label": "NuScale Power",
        "ticker_tested": "SMR",
        "executable": true,
        "shares_available": true,
        "author_price": 12.54
      }
    ],
    "selected_expression": {
      "ticker": "SMR",
      "direction": "short",
      "instrument": "shares",
      "platform": "robinhood",
      "trade_type": "direct",
      "author_price": 12.54
    }
  }
}

Mapping rule from route output:

• route.selected_expression.routed_ticker -> route_evidence.selected_expression.ticker
• keep instrument/platform strings exactly as returned (shares/perps, robinhood/hyperliquid)
• if proxy route selected, include fallback_reason_tag (and fallback_reason_text when direct executable exists)

These fields cross-reference each other. save.ts validates consistency: every subjects[].label needs a matching direct_checks[].subject_label, and the selected ticker must appear in who. Include updated who, route_evidence, and derivation in the same --update call.

8 - Narrate

Build a derivation chain for every routed trade:

{
  "explanation": "1-2 sentences that explain the trade in plain English. No filler, no em dashes.",
  "segments": [
    { "quote": "verbatim source quote", "speaker": "speaker name", "speaker_handle": "@handle", "timestamp": "14:22", "source_url": "https://..." }
  ],
  "steps": [
    { "text": "observation grounded in source", "segment": 0 },
    { "text": "researched fact or confirming evidence", "url": "https://..." },
    { "text": "trade conclusion or implication" }
  ]
}

Write an explanation for every routed trade. Lead with the sharp insight and explain the reasoning in 1-2 sentences. No jargon.

Prefer exactly 3 steps. Each step must advance the chain, no filler. Max 70 chars per step — write like a headline, not a sentence. Cut dates, parentheticals, and asides. No em dashes. No jargon.

Steps should earn the conclusion, not summarize it. If the author named the ticker, the chain can be short. If routing required a leap, earn it. When a step depends on external research or a factual check, cite it with numbered inline citations in Markdown: [1](url), [2](url). Include timestamps when available.

Rules

• Provenance: has segment = sourced from quote, has url = backed by research, has neither = agent inference
• when a step depends on external research or a factual check, embed the source inline as numbered Markdown citations: [1](url), [2](url); treat this as part of the format, not decoration
• url on a step is a fallback when numbered inline linking does not fit
• be honest when a step is your own inference
• user thesis: their words are the segment, speaker: "user"
• video/podcast: every segment MUST include timestamp (MM:SS or H:MM:SS from diarized transcript) and source_url (the video URL). These power click-to-seek on the source page. Resolve speaker X handles when it materially helps attribution.
• tweets: timestamp and source_url can be omitted (no timestamp concept for tweets)

Card steps

After writing derivation, rewrite each step into a chain_steps_card array on the trade body. These are the steps shown on the share card — a completely separate field from derivation.steps.

Take each derivation step and compress it: strip citations, URLs, dates, parentheticals, ticker context. Keep only the core claim. Max 60 chars per step. Exactly 3 steps. No jargon. No em dashes. Write like a headline a friend can understand.

9 - Price

Instrument preference

• Direct thesis subject on Hyperliquid → perps
• ETF tickers → run discover.ts --query "TICKER" to check for an HL perp on the same underlying. Route the HL perp, not the ETF.
• Sector/commodity/index thesis with an HL thematic equivalent → HL perps (not when the author named a specific company; their thesis is the company, not the sector)
• PM contract that prices the event/catalyst → post as separate thesis alongside price trade. PM is additive, not competing. Skip only when no relevant contract exists.
• Otherwise direct thesis subject via shares
• If no direct executable route, use the best proxy

Pricing

bun run skill/scripts/route.ts --run-id <run_id> --thesis-id <id> TICKER direction --source-date "ISO-8601-datetime-or-YYYY-MM-DD" --horizon "timing"
# Returns: { tool: "route", route: { ticker, direction, executable, selected_expression, alternatives, price_context, candidate_routes, note }, diagnostics }
# selected_expression and candidate_routes include HIP-3 routing metadata (see routing.md).
# price_context: { current_price, source_date, author_price }
# If perps route selected and routed_ticker is provided, post that routed_ticker as ticker.

Use tool numbers directly. Do not estimate or recompute.

After routing completes for a thesis, persist everything in one update: who (updated to final ticker), route_status, route_evidence, and derivation together:

echo '<JSON with who + route_evidence + derivation>' | bun run skill/scripts/save.ts --run-id <run_id> --update <id>

This emits thesis_routed (or thesis_dropped) events automatically, updating the live source page with derivation data as each thesis resolves.

10 - Post

Post each trade:

echo '<JSON payload>' | bun run skill/scripts/post.ts --run-id <run_id>

Post rules

• headline_quote must match one saved quotes[] entry or be an exact substring of one.
• Posted ticker, direction, instrument, platform, and trade_type must match route_evidence.selected_expression.
• Carry author_price from route price_context whenever present.
• post.ts will attempt baseline enrichment via /api/skill/assess if author_price is missing, but treat that as fallback not primary path.

After all trade POSTs succeed, finalize the source explicitly:

echo '{ "source_id": "...", "source_theses": [...], "source_summary": "...", "message": "All trades posted" }' | bun run skill/scripts/finalize-source.ts --run-id <run_id>

Finalization

• source_id: source page being completed
• source_theses: all extracted theses, routed and unrouted
• each source_theses entry must carry thesis_id (or id) from save.ts
• each routed source_theses entry must include non-empty who
• each unrouted source_theses entry must include non-empty unrouted_reason
• every extracted thesis must appear exactly once in source_theses (no drops, no duplicates)
• source_summary: one-line summary of the whole source, especially important for grouped sources like timelines
• message: optional completion message

Do not rely on a trade POST to resolve the live source page.

11 - Contract

Required fields

| Field | Notes | |-------|-------| | ticker | Use routed_ticker value from route output. Post as ticker, not as routed_ticker | | direction | "long" or "short" | | author_price | Price at author's publish date. Use route price_context.author_price | | thesis | Thesis text | | headline_quote | Must match one saved quotes[] entry or be an exact substring of one, and be <=120 chars | | ticker_context | 1-3 sentences that explain the instrument to someone who doesn't know what it is. No jargon. | | author_handle | Speaker/author whose quote anchors this trade; user thesis -> current authenticated user handle | | author_platform | "youtube", "x", "substack", "podcast", "pdf", "direct", etc. | | source_url | string or null | | author_date | ISO 8601 — when the author said it | | trade_type | "direct" or "derived" | | instrument | "shares" or "perps" | | platform | "robinhood" or "hyperliquid" | | thesis_id | ID from save.ts | | derivation | { explanation, segments, steps } where explanation is the 1-2 sentence summary and steps are the main chain. |

Source fields

• source_title: title/headline when the source has one
• source_images: image URLs extracted from the source

Finalization-only fields:

• source_theses: all theses from this source, passed to finalize-source.ts
• source_summary: one-line source summary, passed to finalize-source.ts

Useful optional trade_data fields:

• horizon
• alt_venues
• avatar_url

For Polymarket trades, also include:

• outcome: "yes" or "no" — canonical field for which token was bought
• pm_side: "yes" or "no" — legacy compatibility alias for outcome
• pm_yes_no_price: raw YES price in the 0-1 range

Notes

• Card price is the canonical entry price at author_date
• For Polymarket, that means the held-side token price: YES price for YES trades, NO price for NO trades
• API warnings are real feedback; notice them and fix obvious quality problems before moving on
• Keep run_id explicit throughout the run. Do not rely on implicit context lookup.

12 - Reply

When done, reply in one block.

• why the trade makes sense
• author's words -> thesis -> instrument
• 2-3 sentences

When 3+ trades come from one source, open with 1-2 sentences framing the portfolio logic, then map them:

[N] trades from @[handle]'s [source type]:

"headline quote" -> TICKER direction
"headline quote" -> TICKER direction
...

-> Reply to dig deeper

If both direct and derived trades exist, show direct first.

Do not include trade card URLs (paste.trade/t/...) in the reply — they are already linked from the live board. Only include the live board URL (paste.trade/s/...) if it hasn't been shared yet.

If posting fails: Board unavailable. Skipping post.

13 - Hard Rules

1. Use "trades" and "market data", never "recommendations" or "advice"
2. Every number must come from a tool
3. Bear theses -> short-side instruments
4. Flag illiquid contracts