research process builder

→ For the scheduled pipeline (Product Hunt monitor, Supabase setup, SQL, deploy): trigger/README.md

---

build validated web research processes through self-annealing loops. takes any research goal, generates search patterns, tests them against real companies, scores accuracy, and iterates until 90%+ reliability.

this is the factory that produces research agent prompts.

the output is a portable .md file with step-by-step search instructions that any agent (Claude Code, Clay/Claygent, custom GPT, browser agent, OpenAI Agents) can follow to reliably surface specific intelligence about companies.

what's inside

├── SKILL.md                          # the methodology — how to build research processes
└── processes/
    ├── find-profiles.md              # 6 steps · 100% accuracy
    ├── find-competitors.md           # 7 steps · 93% accuracy
    ├── find-reviews.md               # 6 steps · 95% accuracy
    ├── find-news.md                  # 7 steps · 90% accuracy
    ├── find-pr-releases.md           # 5 steps · 90% accuracy
    ├── find-hiring.md                # 5 steps · 93% accuracy
    ├── find-job-role-insights.md     # 5 steps · 90% accuracy (companion to find-hiring)
    ├── find-growth-signals.md        # 7 steps · 93% accuracy
    └── find-negativity.md            # 6 steps · 90% accuracy

how it works

the methodology has 6 phases:

1. define the goal — state what you're looking for and what a "good result" looks like
2. generate 15-20 candidate search patterns — parameterized queries like {{company_name}} competitors
3. test patterns against real companies — 6-10 sample companies across 3 size tiers (enterprise to micro startup)
4. score and classify — quality (1-5) x consistency (1-5). classify as PRIMARY / ENRICHMENT / FALLBACK / KILL
5. iterate until 90%+ — identify failure modes, generate fix patterns, retest
6. assemble the process file — ordered steps with extract instructions, early stopping, kill list, output template

full methodology is in SKILL.md.

the 9 processes

built using the methodology above. 200+ patterns tested across 11 companies ranging from SpaceX ($400B+) to micro bootstrapped agencies.

process	what it finds	steps	accuracy
find-profiles	company fact sheet from zoominfo, crunchbase, linkedin, rocketreach, pitchbook, tracxn	6	100%
find-competitors	direct competitors with positioning and justification	7	93%
find-reviews	individual reviews tagged positive/negative with three-sentence summaries	6	95%
find-news	partnerships, acquisitions, funding, launches, expansions, leadership changes	7	90%
find-pr-releases	official announcements, press releases, blog posts, wire distributions	5	90%
find-hiring	open roles, departments hiring, ATS platform, hiring velocity	5	93%
find-job-role-insights	tech stack, pain points, strategic signals from specific job descriptions	5	90%
find-growth-signals	blog activity, lead magnets, social presence, community, newsletters, pricing maturity	7	93%
find-negativity	customer complaints, negative reviews, controversy, churn signals	6	90%

each process file includes:

• clear {{input}} placeholders you fill in before running
• step-by-step search patterns with exact queries
• exact extraction specs (what to pull from each search, three-sentence summaries)
• stop if conditions so the workflow stops when it has enough
• a kill list of patterns that look promising but waste searches
• a casual, structured output template

how to use the processes

with any AI agent (ChatGPT, Claude, etc.)

paste the process file content as the system prompt or instructions. fill in the {{inputs}}. the agent follows the steps, stops when it has enough, and outputs in the specified format.

with Clay / Claygent

each step maps to a Clay enrichment column. use the search query as your Claygent prompt. the extract instructions tell you what to pull from results.

with Claude Code

drop the process files into your .claude/skills/ directory. reference them when researching companies:

"research [company] using the find-competitors process"

how to build your own processes

use SKILL.md to build processes for any research goal:

• tech stack detection
• hiring signal monitoring
• pricing intelligence
• market sizing
• content gap analysis
• anything you can search the web for

key discoveries

things we learned testing 220+ search patterns:

• site:reddit.com is completely broken — zero results universally. use [name] reddit discussion without the site: operator.
• year modifiers are the highest-leverage search modifier. [name] review 2026 outperforms [name] review by a wide margin.
• zoominfo + linkedin are the only platforms that cover ALL company sizes, including 6-month-old startups.
• generic company names (Clay, Keep, Cursor, Harvey) need mandatory disambiguation. add category qualifier or use domain.
• kill lists save more time than pattern lists. knowing which searches to NOT run prevents wasting 30-40% of your search budget.
• ATS board searches are gold for hiring data. site:boards.greenhouse.io [name] and site:jobs.ashbyhq.com [name] return actual role listings with titles and descriptions.
• [name] social media twitter youtube is a trap. returns product feature content, not the company's actual social accounts. use site:twitter.com OR site:x.com with company name instead.
• OR operators in a single query are powerful. [name] alternatives OR competitors OR "vs" catches 3 result types in one search, tested Q4.75/C4.75.
• wellfound (formerly angellist) is the T3 lifeline for hiring data. small startups without greenhouse/lever/ashby pages still have wellfound profiles with employee count, funding, and industry tags.
• site:[domain] with OR operators is the most efficient growth signal detector. a single query like site:[domain] blog OR pricing OR newsletter OR demo catches 4+ signal types in one search.
• rocketreach is at rocketreach.co, NOT .com. site:rocketreach.com returns zero results universally. site:rocketreach.co returns rich org chart data including employee titles, department breakdown, and key people even for T3 companies.
• combined platform OR queries are a cheat code. site:zoominfo.com OR site:rocketreach.co OR site:crunchbase.com [name] pulls from 3 ungated platforms in one search. tested with Lovable: returned $200M funding, $1.8B valuation, ARR, founder, and org chart in a single query.
• churn-signal searches are a trap. [name] "switched from" OR "left" OR "cancelled" returns marketing content about people switching TO the tool, not FROM it. tested Q2/C1.
• "do not recommend" and "waste of money" searches return nothing. people don't use these exact phrases in searchable contexts. use [name] complaints OR problems instead.
• never hardcode the year in process files. use {{current_year}} as an input variable so processes stay valid across years. in Clay, populate from YEAR({Created At}).
• ATS-specific JD searches are the best path to full job descriptions. site:jobs.ashbyhq.com/[company] [role] returns exact JD links. combined ATS OR query (site:jobs.ashbyhq.com OR site:boards.greenhouse.io OR site:jobs.lever.co [company] [role]) catches roles across all three platforms in one search.
• site:linkedin.com/jobs is broken for web search. returns generic LinkedIn job search pages, not company-specific listings. for LinkedIn jobs data, use Claygent to visit linkedin.com/company/[slug]/jobs/ directly.
• companion processes unlock depth. find-hiring gives you breadth (all open roles). find-job-role-insights gives you depth (what a specific JD reveals about strategy, tech stack, and pain points). chain them in Clay for the full picture.
• site:{{domain}}/blog beats [name] blog for finding owned content. [name] [category] blog 2026 returns third-party blogs ABOUT the company (Q3). site:[domain]/blog returns the company's own posts (Q5). use the former for buzz, the latter for owned content.
• [name] [category] newsletter is a trap. returns product feature content about newsletters, not the company's actual newsletter. tested Q2.25 across all tiers. site:[domain] "subscribe" OR "newsletter" finds actual signup mechanisms (Q4).
• community platforms (discord/slack) are the most underrated growth signal. [name] discord OR slack OR community tested Q5 across Clay (Slack 15K+), Lovable (Discord 162K+), Cursor (Discord 15K+). an active community signals product-market fit and organic advocacy.
• site:youtube.com [name] returns zero results. YouTube is not searchable via site: operator. avoid entirely.

validation

• 220+ patterns tested via live web search
• 11 companies: SpaceX, Cohere, Harvey AI, Cursor, Clay, Lovable, Keep, Cluely, ClickUp, LeadGrow, The Kiln
• companies ranged from $400B+ to micro bootstrapped
• each pattern tested against 3-4 companies minimum
• three iteration rounds with 30+ fix patterns targeting identified failure modes

license

MIT