The JSONCRM agent cookbook

Recipe 1 — Triage an intake list

You exported 200 LinkedIn post likers. Get the clean ones into leads, scored:

# 1. normalize the MCP export into flat CRM records (in place)
jsoncrm parse-from-linkedin-mcp likes.json

# 2. preview, then drop records already in the pipeline
jsoncrm deduplicate likes.json --dry-run
jsoncrm deduplicate likes.json

# 3. preview, then drop competitors / blocklist matches
jsoncrm filter-competitors likes.json --dry-run
jsoncrm filter-competitors likes.json

# 4. agent scores each record (external data, judgment) — then merge
jsoncrm merge likes.json

# 5. sanity + audit trail
jsoncrm validate
git add -A && git commit -m "intake: N new leads from LinkedIn likers"

What's left in likes.json after step 4 is the merge report — the unmerged are duplicates or invalid; check validate output if anything looks off.

Recipe 2 — Score unscored leads, one at a time

For careful scoring with a human in the loop, use the intake/pending cycle:

jsoncrm intake intake.json     # writes .pending_update.json with the next unscored record

The agent researches the record, fills in score (must be a value from .crm.json's scores.order) plus a one-line notes rationale, then:

jsoncrm apply_update # applies the pending file, one record per cycle

For bulk scoring, update with --dry-run per record is faster — but the pending cycle forces you to look at each one.

Recipe 3 — Bulk promote qualified leads

The highest-leverage agent loop. Preview everything, review, then commit:

# who qualifies?
jsoncrm list leads --score "⭐⭐⭐⭐" -n 50

# preview each promotion — nothing written
jsoncrm promote "https://linkedin.com/in/alice" --lead --dry-run
jsoncrm promote "https://linkedin.com/in/bob"   --lead --dry-run

# apply for real
jsoncrm promote "https://linkedin.com/in/alice" --lead --apply
jsoncrm promote "https://linkedin.com/in/bob"   --lead --apply

git diff --stat
git add -A && git commit -m "pipeline: promote 2 qualified leads"

Over MCP, the same prompt-shaped version: "promote everyone at four stars or above — dry run first and show me each record." Invalid transitions (customer → lead) simply don't exist — they're not configured in .crm.json.

Recipe 4 — The Monday review (for humans)

jsoncrm stats                  # stage counts, scored/contacted, paid customers
jsoncrm top -n 10 --min ⭐⭐⭐     # best leads not yet contacted
jsoncrm recent -n 10             # freshest additions
jsoncrm search "Acme" --company  # the deal you're wondering about

Or just ask your agent: "give me this week's pipeline summary — new leads, promotions, and the top 10 by score that we haven't contacted."

Recipe 5 — Hygiene day

jsoncrm validate               # structural integrity of every pipeline file
jsoncrm backup                 # ZIP snapshot outside git → ~/.jsoncrm/backups/
git log --oneline --since="2 weeks ago" -- leads.json prospects.json

Git history won't save you from damage to .git itself — the backup is the independent copy.

Teach your agent the house rules

All of this fits in a skill. Install the first-party one and every Claude Code session in your CRM repo knows the rules, the command map, and the recipes:

mkdir -p .claude/skills/jsoncrm
curl -fsSL https://raw.githubusercontent.com/pauling-ai/jsoncrm/main/skills/jsoncrm/SKILL.md \
  -o .claude/skills/jsoncrm/SKILL.md

For Codex/Cursor, append skills/AGENTS-snippet.md to your AGENTS.md.

The agent cookbook

⚖️ The three house rules