OutClaw — Research
FIRST ACTION RULES — read before anything else
1. Discovery intent (user said "get me leads / pull leads / prospects / outreach plan")?
Plans are delivered in daily batches of 10–15 fresh targets, not 1–3.
Three is noise for a salesperson at a 5% reply rate. Use the batch flow:
# a) Pull + save the Leadbay response, then let the batch script do the
# bulk work (tier-1 persist for every lead, fresh-pick top N, emit
# manifest).
# Agent: call leadbay_pull_leads (no args) → save JSON → /tmp/leadbay-pull.json
bash ~/.openclaw/skills/outclaw/shared/scripts/outclaw_daily_batch.sh --n 15 --stale-after-days 7
# b) DEFAULT: skip Tier-2 for the full batch. Tier-1 bodies now carry the
# Leadbay AI qualification_summary + tags + recommended contact (see
# leadbay_tier1_persist.py's richer body format) — enough for Day-1
# drafts. Tier-2 research is OPT-IN for the ~top 3 the user wants to
# go deeper on (costs ~3 min / lead via leadbay_research_lead).
# c) Build the Day-1 scaffold for all 15 targets. Scaffolder pulls angle
# hooks from each Tier-1 body (no Tier-2 required) + on-brand copy +
# sender identity.
python3 ~/.openclaw/skills/outclaw/shared/scripts/plan_scaffolder.py --max 15
# Output: /tmp/outclaw-plan-draft.md
After the scaffold is written, the agent's remaining work is only
writing the 15 Day-1 email bodies + Provenance blocks (no research, no
bookkeeping), then running channel_validator.py + draft_checker.py, then
presenting. See outclaw-plan SKILL.
Tier-2 is opt-in. When the user says "go deeper on <name>" or
"research <name> more", THEN call leadbay_research_lead +
leadbay_tier2_persist.py for that specific target. Do NOT Tier-2 the
full batch by default — the turn budget won't cover 15 × 3-min research
calls.
2. Targeted research intent (user named a specific person/company)?
Start at "Step 1 — Parse the input into a target" below.
3. Never fabricate. Every fact in a KB page body, plan angle, or
draft email MUST be traceable to a raw source file under
~/.openclaw/outclaw/kb/raw/ or kb/me/. If you can't cite the file, you
can't include the fact. (See § Zero-hallucination rule.)
4. No promise-then-silent turn closes. Do NOT end a turn with "I'll
spin on that now", "surface shortly", "standby", "more to come". That
pattern is a silent failure — the agent yields control and never returns.
Either finish the work in this turn and present the results, or stop
cleanly at the last completed step with a short "I've done X / Y is next
— want me to continue?". The draft_checker.py promises-regex rejects
plans containing these phrases.
Your job: given a person (name, LinkedIn, email, company) or a discovery
intent ("find me someone at Acme worth talking to"), use every available tool
to build a rich, persistent profile and store it in the KB.
Output quality bar: ≥3 concrete, recent (≤30 day) signals per target +
a section cross-referencing the target with kb/me/ (mutual orgs, schools,
topics, connections). If you can't hit that bar, explicitly say what you were
blocked on — never paper over thin research with fluff.
Use the tools you actually have
Check your own tool list at session start. Your ordering preference:
leadbay_research_lead / leadbay_research_company — if available, use first
web_search + web_fetch — baseline every OpenClaw agent has these
read (for local caches under kb/raw/)
- User-provided text pasted into the conversation
If leadbay_* isn't in your tool list, that's fine — use web_search to
find the target's LinkedIn/Twitter/company page, web_fetch the specific
pages, and write what you actually learn into raw/<slug>-<ts>.md.
Never insist on a tool that isn't in your active tool list. Stating
"Leadbay isn't connected" is correct; pretending you used it is a
hallucination failure.
Resolver mandate (non-negotiable)
Before creating or modifying any page under ~/.openclaw/outclaw/kb/ or
any entry in memory/<tenant>/memory.jsonl, read
shared/references/RESOLVER.md and file by primary subject, not by source
format or skill name. Use shared/scripts/kb_ingest.py + kb_page.py —
do not hand-craft file paths. Research facts go in kb/{people,orgs,topics,places}/;
runtime observations ("email bounced") go in per-tenant memory, not the KB.
Zero-hallucination rule (non-negotiable)
Every signal, hook, career fact, org fact, and connection point MUST be
traceable to a real source — either a URL you fetched or a file you wrote
under ~/.openclaw/outclaw/kb/raw/. If you cannot point at a real source,
do NOT write the fact. Leave the field empty and list it under gaps[]
in your return.
Specifically forbidden:
- Inventing a date ("March 2026 post", "€5M Series A in April 2026") without
a URL or raw/ file that says so.
- Inventing a connection point ("HEC Paris alum", "both at Stripe") unless
you can cite a line in
kb/me/self.md or kb/me/org.md.
- Claiming you wrote a KB page without running the write command and
reading the file back.
Producing plausible-sounding content you can't source is a hard failure
caught by the verification gate below.
Preamble (skip if called from orchestrator)
SHARED="$(dirname "$(dirname "$(cd "$(dirname "$0")" && pwd)")")/shared"
bash "$SHARED/scripts/memory_search.sh" --type tool_inventory --limit 1
cat ~/.openclaw/outclaw/kb/me/self.md 2>/dev/null
cat ~/.openclaw/outclaw/kb/me/org.md 2>/dev/null
Flow
Step 1 — Parse the input into a target
The user gave you some combination of: name, LinkedIn URL, email, company,
title, company URL. From these, compute:
slug — kb_page.py style: lower-case name, non-alnum → -, e.g.
alice-chen. Disambiguate with company if needed (alice-chen-stripe).
display_name
company (if known) → compute an org slug the same way (acme-corp)
Discovery intent (e.g. "get me some leadbay leads", "pull a promising
lead", "purchase contacts for the best ones"):
Leadbay is a sales inbox, not a queryable database. See
shared/references/leadbay-integration.md for the full mental model.
Key rules:
- NEVER ask the user for "targeting criteria" (industry, company
size, geography, lead count, job titles). The user's ICP is already
configured inside Leadbay. Asking for criteria = you're thinking of
Leadbay as a DB, which it isn't.
- NEVER ask which titles to enrich when the user says "recommended
contacts" or "the best contacts" —
leadbay_enrich_titles auto-picks
based on the ICP. Recommended means Leadbay chooses.
- Verify your tool list FIRST. Before attempting any Leadbay
operation, check that
leadbay_pull_leads (or at minimum
leadbay_account_status) is in YOUR active tool list — not just
plugins.entries.leadclaw. If the composite tools aren't bound to
your agent, refuse per the template in leadbay-integration.md
§"What to do when the composite tools aren't in your tool list".
The canonical flow — four tool-call blocks, no improvisation:
The flow is written so the agent only decides which lead_ids to deep-research.
Everything else is driven by helper scripts. Execute each block; paste the
literal stdout to the user. Do not paraphrase.
# --- Block 1: pre-counts + pull ---
BEFORE_ORGS=$(ls ~/.openclaw/outclaw/kb/orgs/ 2>/dev/null | wc -l | tr -d ' ')
BEFORE_PEOPLE=$(ls ~/.openclaw/outclaw/kb/people/ 2>/dev/null | wc -l | tr -d ' ')
echo "before: orgs=$BEFORE_ORGS people=$BEFORE_PEOPLE"
Then call the MCP tool leadbay_account_status, followed by leadbay_pull_leads
(no args). Save the pull response JSON to /tmp/leadbay-pull.json using your
Write tool.
# --- Block 2: Tier-1 persist + verification gate (MANDATORY) ---
python3 ~/.openclaw/skills/outclaw/shared/scripts/leadbay_tier1_persist.py \
--from-file /tmp/leadbay-pull.json
AFTER_ORGS=$(ls ~/.openclaw/outclaw/kb/orgs/ | wc -l | tr -d ' ')
AFTER_PEOPLE=$(ls ~/.openclaw/outclaw/kb/people/ | wc -l | tr -d ' ')
N_LEADS=$(python3 -c 'import json; print(len(json.load(open("/tmp/leadbay-pull.json")).get("leads",[])))')
echo "after: orgs=$AFTER_ORGS (+$((AFTER_ORGS-BEFORE_ORGS))) people=$AFTER_PEOPLE (+$((AFTER_PEOPLE-BEFORE_PEOPLE))) / leads_pulled=$N_LEADS"
# GATE: the orgs delta MUST be > 0 while N_LEADS > 0. If zero, STOP — the
# persist step was skipped. Do not continue to block 3.
# --- Block 3: pick top N to deep-research (freshness-aware) ---
python3 ~/.openclaw/skills/outclaw/shared/scripts/leadbay_top_picks.py \
--from-file /tmp/leadbay-pull.json --n 15 --stale-after-days 7
This returns {ids: [...], summaries: [...], skipped_recent: [...]}.
The skipped_recent array names any orgs we drafted a plan for in the
last 7 days — they are INTENTIONALLY deprioritised so the user sees
fresh targets day-to-day. Show the skipped list to the user with a note:
"Still in flight from <date> — include anyway?" Only re-add to the plan
if the user says yes.
Tier-2 research is OPT-IN — do NOT run it for the full batch by default.
The Tier-1 body already carries the Leadbay AI qualification excerpt +
tags + contact, which is enough for a Day-1 draft. Only call
leadbay_research_lead when the user says "go deeper on <X>" for a
specific target.
# --- Block 4: Tier-2 persist (per researched lead) ---
for f in /tmp/leadbay-research-*.json; do
python3 ~/.openclaw/skills/outclaw/shared/scripts/leadbay_tier2_persist.py --from-file "$f"
done
# Final audit
ls ~/.openclaw/outclaw/kb/orgs/ | wc -l
ls ~/.openclaw/outclaw/kb/people/ | wc -l
tail -6 ~/.openclaw/outclaw/kb/log.md
After block 4, hand off to outclaw-plan with the list of slugs for the
top picks. Plan draws EVERY concrete fact from the Tier-2 org/person
pages — no improvised claims, no [Your Name] / [Your Company]
placeholders, no fabricated "recent Series C" unless it's on the page.
Non-negotiable rules:
- Block 2's gate must pass before you touch block 3. Orgs delta zero = STOP.
- Never paraphrase a script's stdout. Paste it as-is.
- Never claim a KB write that isn't visible in the post-counts or log tail.
- Use
enrich_titles only when the user asks for richer contact coverage —
leadbay_pull_leads already returns a recommended contact per lead.
If you find yourself writing the words "what industry", "what size",
"which titles", STOP. You're asking the wrong question. Read
shared/references/leadbay-integration.md again.
Step 2 — Check KB first
python3 "$SHARED/scripts/kb_search.py" --slug person:$SLUG
If an entry exists and its last_updated is ≤30 days, return it and ask the
user if they want a refresh. If yes, proceed. If no entry, proceed.
Step 2.25 — Email-domain discipline
If you have an email address for the target, classify the domain BEFORE
deriving a company from it:
python3 "$SHARED/scripts/domain_classifier.py" "<email>"
private (gmail, hotmail, yahoo, proton, yandex, icloud, …): the email
is a personal mailbox. Do NOT set company from this domain. Use
LinkedIn / Leadbay / the target's own statements instead.
isp (comcast, bt, verizon, cox, …): residential internet provider —
same rule. Do NOT infer employer.
school (.edu, .ac.uk, hec.fr, mit.edu, stanford.edu, …): academic
affiliation. Note it on the page as ## Affiliations (school, not
current employer) but do NOT treat as employer unless the target is
confirmed faculty/staff.
corporate: the domain IS the company. Use company_slug_from_domain
result as the kb/orgs/<slug>.md target. Confirm by cross-checking
LinkedIn / company URL when available.
Example wrong inference to avoid: "alice@gmail.com" → org slug "gmail".
The classifier catches that; the skill must ACTUALLY call the classifier,
not eyeball the domain.
Step 2.5 — Cross-reference: past conversations + same-company people
Before running fresh external research, sweep local context you already
have:
a. Past conversations with THIS person. Query every connected
communication-capable tool for threads where the target appears as
sender or recipient. Priority order:
| Tool in inventory | Query |
|---|
gog (Gmail) | gog gmail messages search "from:<email> OR to:<email>" --max 50 -j |
email-mcp | same search via the MCP's search endpoint |
slack-mcp-server / slack | slack search --user <slack_id_or_email> --max 50 |
whatsapp-mcp-ts / wacli | search conversation history by phone |
telegram-mcp | search chats where counterparty matches handle |
mac_messages_mcp / imsg | iMessage/SMS history by phone or contact |
discord-mcp | DM history by user id |
linkedin-cli | InMail / message history with the profile URL |
apple-notes / things-mac | search for user's own prior notes about the target |
For every match, record:
- Date of last interaction
- Direction (sent-by-user | received-from-target)
- Channel
- A one-line summary (don't paste the body — PII)
Drop a raw/ snippet if the thread is substantive:
kb_ingest.py begin person <slug> --note "prior-thread-<channel>-<date>: <summary>".
Under ## Prior conversations on the target's page, list each match:
- 2025-11-14 · Gmail · user → target · "Intro from David, re: ICP pilot"
- 2026-02-03 · LinkedIn DM · target → user · "Circling back after SaaStr Annual"
b. Signature extraction. When a past email from the target exists,
parse the signature block. Useful fields to harvest INTO the KB
frontmatter (not prose):
phone (if present)
linkedin_url (if present — updates a missing field)
company (only if the domain check said corporate AND the signature's
stated company matches the domain; otherwise flag the conflict)
address
- Confirmed current
title
If the signature is an IMAGE (.png / .jpg attachment inline), transcribe
it via an available OCR-capable tool in your inventory:
| Tool | How |
|---|
gemini | gemini describe --image <path> --prompt "Transcribe this email signature verbatim. Return fields: name, title, company, phone, email, address, socials." |
image_generate / multimodal fallback | pass the image bytes + the same prompt |
openai-whisper-api | audio only — not applicable |
Write extracted fields onto the KB frontmatter via kb_page.py upsert.
Never invent fields — if the OCR result is unclear, log a gap, don't guess.
c. People from the same company (if known). If company_slug is
known (from email or LinkedIn), pull every page in the KB for that org:
# List all people with this org in their connections
grep -l "orgs: .*\b<company-slug>\b" ~/.openclaw/outclaw/kb/people/*.md
# Or the org's own backlinks:
python3 "$SHARED/scripts/kb_search.py" --slug org:<company-slug>
For each of those people, surface a one-line summary on the target's
page under ## Same-company context:
- Bob Smith (CTO, kb/people/bob-smith.md) — active thread Mar 2026, warm
- Carol Wong (VP Sales, kb/people/carol-wong.md) — opted out 2025-10
If ANY same-company person has contact_status: opt_out, surface it
prominently — opt-outs propagate to company-level wariness but NOT to
automatic blocking (one person's opt-out doesn't ban outreach to their
whole org).
d. Prior interactions with the ORG. Same idea but at org-level —
reach into kb/orgs/<slug>.md's existing body for any ## Update or
## Prior interactions sections.
The summary of Step 2.5 goes into the target's page body BEFORE the
external-research sections you'll fill from Step 3. Past conversations +
same-company context are often the most valuable signals — they beat a
fresh LinkedIn scrape every time.
Step 3 — Enrich, category-by-category
Read references/research-playbook.md for the full matrix. Summary:
- For each outreach-relevant category present in the inventory, query
at least once. Don't skip a category that's
ready — if you do, plan
quality suffers later and the evaluation will catch it.
- Signals you're hunting for, in order of value:
- Recent public posts / talks / podcasts (≤30 days)
- Career trajectory (current role tenure, prior companies)
- Topic interests (inferred from what they share / write)
- Mutual connections with the user (from
kb/me/self.md)
- Contact info (for later planning only — never use it here)
- Never fabricate. If a source doesn't yield a signal, move on.
Categories → tools (from inventory):
crm → LeadClaw for full enrichment + ICP + relationship graph
professional_network → LinkedIn for posts, tenure, connections
social → Twitter/Bluesky for recent posts
research_web → browse, summarize for company blog, podcasts, news
research_contacts → apple-notes, things-mac if the user kept prior notes
Step 4 — Write the KB
Delegate the file-writing to shared/scripts/kb_ingest.py + kb_page.py.
Every target MUST:
- Begin with
kb_ingest.py begin person <slug> --name "<Display>" — creates
a raw/ entry tagged with the first source.
- Have a body containing sections:
## Role & trajectory, ## Recent activity (≤30 days), ## Topics & interests, ## Connection points with <user>,
## Sources. Write via kb_page.py upsert person <slug> --body <path>.
- Touch the org:
kb_ingest.py touch org <org-slug> --name "<Org>" --connection person:<slug>.
Also append a ## Update <date> section to the org page with whatever you
learned about it while researching the person.
- Touch topics: for each interest,
kb_ingest.py touch topic <topic-slug> --connection person:<slug>. Create stubs if missing.
- Rebuild index:
python3 "$SHARED/scripts/kb_index_rebuild.py".
- Log:
kb_ingest.py log ingest "<slug> research — <one-line summary>" --pages people/<slug>.md orgs/<org>.md topics/<topic>.md.
Step 4.5 — Verification gate (MANDATORY)
After every kb_page.py upsert and kb_ingest.py touch, cat the file back
and include the first 40 lines under a ## Verified writes section in
your response. Example:
for p in ~/.openclaw/outclaw/kb/people/<slug>.md \
~/.openclaw/outclaw/kb/orgs/<orgslug>.md \
~/.openclaw/outclaw/kb/index.md \
~/.openclaw/outclaw/kb/log.md; do
echo "--- $p ---"
/usr/bin/head -40 "$p" 2>/dev/null || echo "MISSING: $p"
done
If any file is missing, STOP. Do not proceed to return. Report the failure
in plain language — "I couldn't write the KB page because X" — and exit.
Do NOT continue to Step 5 with fake content.
Step 5 — Return
Return to caller:
- The KB page paths (relative to
~/.openclaw/outclaw/kb/)
- An executive paragraph: role / why-now / 2-3 specific hooks / 1-2 connection
points with the user
- What you couldn't find and why (blocked on LinkedIn auth? LeadClaw has no
match? Their Twitter is private?)
Delegate deep multi-tool work to the agents/person-researcher.md sub-agent
when the target is rich enough to justify it.
Degraded mode (no LeadClaw)
When the inventory has crm=[leadclaw:missing] or :needs_setup]:
- Skip ICP scoring + relationship-graph
- Rely on web search + LinkedIn + social for everything
- Quality score #3 (research breadth) will drop — that's expected and the
user has been warned at setup time. Mention in the returned summary that
Leadbay would unlock richer signals.
What this skill does NOT do
- Does not plan outreach — that's
outclaw-plan.
- Does not contact anyone.
- Does not fabricate signals. When in doubt, say "I couldn't find …".