Deal Flow

<!-- autoresearch: variation B — sharper output via leverage scoring, valuation context, and banned-phrase quality gates -->

Read memory/MEMORY.md for context (tracked verticals, active theses). Read the last 14 days of memory/logs/ and extract company names from prior ### deal-flow entries to build the dedup set.

Goal

Produce a weekly funding-round digest that tells the reader what the capital reveals, not just what closed. A flat list of rounds is noise; a ranked, contextualized read is signal.

Steps

1. Gather candidates from tiered sources

Run searches in parallel. Use the current month + year in queries. Aim for ≥25 candidate deals before filtering. If a source fails, record source=fail for the footer; do not abort.

Source fetch rule (applies to every WebFetch below): if the fetch returns 404, 403, is rate-limited, or times out, log DEAL_FLOW_SOURCE_MISS: <url> (<reason>) and continue with remaining sources. Never abort the run on a single source miss — the tiered design is specifically so that any one source can fail.

Tier 1 — broad aggregators (always run):

• WebSearch: "biggest funding rounds" week ${month} ${year} site:news.crunchbase.com
• WebSearch: funding round ${month} ${year} site:techcrunch.com
• WebSearch: "Series A" OR "Series B" OR "seed" raised ${month} ${year} startup
• WebFetch https://news.crunchbase.com/sections/venture/ — extract this week's roundup if present (apply source-miss rule above)

Tier 2 — vertical-specific (run for verticals in MEMORY.md; defaults: AI, crypto, infra/devtools):

• AI: WebFetch https://aifundingtracker.com/; WebSearch "AI startup" "raised" Series ${month} ${year}
• Crypto: WebFetch https://crypto-fundraising.info/ and https://cryptorank.io/funding-rounds (apply source-miss rule); WebSearch crypto funding round ${month} ${year}
• Infra/devtools: WebSearch developer tools OR infrastructure OR compute startup raised ${month} ${year}

Tier 3 — signal flags (always include if found, regardless of size):

• Prediction markets, coordination markets, agentic payments, x402, autonomous-agent infra, on-chain identity
• WebSearch: prediction market OR agentic payment OR x402 funding ${month} ${year}

X/Twitter fallback signal path (always run, especially when Tier 1/2 sources miss):

• Use fetch-tweets outputs or direct X search for the following queries — these surface deals that aggregators miss or haven't indexed yet:
  • funding-announcement (literal hashtag and phrase)
  • "announcing our Series" OR "we raised" ${month} ${year}
  • "led by @<tier1_fund>" raised (substitute a16z, paradigm, sequoia, etc.)
  • "thrilled to announce" funding ${month} ${year}
• Treat tweets from verified founder/investor accounts as primary-source candidates; still dedup by company name and verify amount/lead from a second source before including.

2. Dedup and enrich

• Drop any deal whose company name (case-insensitive) appears in the last-14-days dedup set, unless the new round is materially different (later stage, >2× larger, or new lead).
• For each surviving candidate, capture:
  • Company + ≤12-word description
  • Round (pre-seed / seed / A / B / C / strategic)
  • Amount in USD
  • Valuation (post-money) if disclosed
  • Prior valuation if known → flag UP / FLAT / DOWN vs prior round
  • Lead investor(s) + notable co-leads
  • Source URL

If valuation or prior is unknown, write n/d. Never hallucinate numbers.

3. Score for leverage (1–5 each)

Dimension	5 (high)	3 (mid)	1 (low)
Magnitude	Round size unusual for stage (Series A >$50M; seed >$15M)	Standard for stage	Below median
Investor signal	Lead is a16z, Paradigm, Founders Fund, Sequoia, Benchmark, Pantera, USV, Lightspeed	Reputable but not top-tier	Unknown / pure syndicate
Thesis fit	Validates a tracked vertical/thesis directly	Adjacent	Off-thesis
Narrative weight	First-of-kind, contrarian, or category-defining	Confirmation of existing trend	Me-too
Valuation signal	UP >2× prior, OR DOWN (down rounds are signal)	FLAT or UP <2×	n/d

Sum the five. Keep the top 8 deals by total. Ties broken by recency.

4. Quality gates (drop if any fail)

• Acqui-hire, grant <$1M, or token sale → drop (token-pick covers tokens)
• Round older than 14 days → drop
• Amount or lead unverifiable → drop (no rumors)
• Total leverage score <12 → drop unless it carries a Tier-3 signal flag
• Cannot write a non-vague "Why it matters" line (see banned phrases below) → drop

5. Write the digest

*Deal Flow — ${today}*

**Read:** [1 sentence: what this week's capital signals — where smart money clustered, what's overfunded/underfunded, which thesis got validated or broken]

1. **Company** — what they do (≤12 words)
   $XXM ${round} @ $YYM post (${UP|FLAT|DOWN|n/d} vs prior) | Lead: ${investor}
   *Why it matters:* [≤20 words; must name the specific signal]

2. ...

*Sources:* crunchbase=ok|fail, techcrunch=ok|fail, cryptorank=ok|fail, aift=ok|fail | candidates=${N} → kept=${K}

Hard rules for "Why it matters":

• Banned phrases (drop the deal if you can't avoid them): "interesting", "growing", "exciting", "potential", "could be", "watch this space", "important space", "validates the trend"
• Must reference at least one of: a named thesis, a competitive dynamic, a valuation anomaly, an investor-pattern shift, a regulatory/macro trigger, or a first-of-kind structural detail.

Cap total at 4000 chars. If everything gets dropped, send: *Deal Flow — ${today}* — No deals cleared signal threshold this week. Sources: ...

6. Log to memory/logs/${today}.md

### deal-flow
- Candidates: ${N} → Kept: ${K}
- Top deal: ${company} ($${amount} ${round}, lead ${investor}) — score ${score}/25
- Themes: [1–2 sentence cluster summary]
- Source health: [any source=fail with reason]
- Dedup: ${count} candidates dropped as already-covered
- Deals (for next-week dedup):
  - ${company1}, ${company2}, ...

7. Send via ./notify

Run ./notify "<digest>". The fan-out handles channels.

Sandbox note

The sandbox may block outbound curl. Use WebFetch (built-in Claude tool) for any URL fetch — it bypasses the sandbox. WebSearch is always available.

Constraints

• Never invent numbers. n/d is acceptable; fabrication is not.
• Never include the same deal as last week unless materially different (later stage, >2× bigger, or new lead).
• Compare week-over-week implicitly via the Read: lead — is capital accelerating or decelerating in each vertical?
• Quality over completeness: if only 3 deals clear the threshold, ship 3.

Environment Variables Required

• None (uses WebSearch + WebFetch)