B2B accounting reconciliation SaaS. ~280 customers. $140K MRR. 6 engineers.

LedgerLink, here's how your tools talk to each other today and what to fix first.

Three sections below. Pantry = your raw data. Prep table = the weekly briefs Claude assembles. Plate = the decisions you actually serve up. Click any card or recipe for plain English.

7
Tools mapped
4
Recipes ready to build
3
Quick wins flagged
Legend:
✓ You've got it
? Missing or not set up yet
! Quick win, money or hours on the table
Where to start: click any card with the orange ! badge first, those are the 3 unlocks Claude can earn its keep on.
Pantry
Your raw ingredients, the tools your data already lives in.
!
Pendo
CSV weekly
High volume
!
Intercom
CSV real-time
High volume
✓
GitHub Issues
CSV real-time
Steady flow
✓
Linear
API real-time
Steady flow
✓
Stripe
CSV weekly
Steady flow
✓
Sentry
API real-time
Steady flow
✓
Otter / Fireflies
DOCX/TXT ad-hoc
Steady flow
Prep Table
Where the night's mise-en-place gets done, weekly summaries Claude writes for you.
✓
Feedback themes
Customer feedback Mon 6am
In: Pendo, Intercom, Otter / Fireflies, GitHub Issues
✓
Churn root cause
Churn Mon 6am
In: Stripe, Intercom, Sentry
✓
Product roadmap signal
Product / roadmap Mon 6am
In: Linear
✓
Support volume
Operations Mon 6am
In: Intercom
Plate
What gets served to a human, the briefs and drafts you actually act on.
!
Codex fix PROPOSAL.md (in worktree)
Written by the codex fix drafter For: Jordan, Sai
🔒 Engineer appends [engineer_action] approved line in the audit log
✓
Weekly product report
Written by your triage chief of staff For: Devon, Maya
🔒 Devon reviews Monday before standup

The mental model in one picture. Your data moves left to right: raw sources feed weekly silver platters, platters feed the briefs and drafts Claude writes for you, and those land in the channels you actually use (Slack, your inbox, your iPad). The AI layer sits on top of the data prep, never the other way around. Hover any line to trace it. Click a node to open the details.

Pantry (raw data)
Prep (silver platters)
Plate (briefs you read)
Where you read it
Opportunity / suggested

If a Pantry box has no line leaving it, that data is sitting unused, those are usually the loudest quick wins. Dashed boxes on the right are channels we suggest adding.

LedgerLink, these are the briefs and drafts your team can stop building manually. Each one shows the ingredients, the time saved, and the back-and-forth Claude does on your behalf.

Monday inbound triage

Devon stops being the human router. Monday morning, 47 inbound tickets are already classified, clustered, prioritized, and routed.

Intercom tickets → GitHub issues → Customer calls → feedback_themes.md → Triage Orchestrator
Before: Devon spent 3+ hours every Monday reading Slack, opening Intercom, scrolling Pendo, scrubbing GitHub. Half the tickets were dupes. Pickup latency 3 weeks.
After: Triage runs Sunday night. Monday standup walks the report. Pickup latency drops to 3-5 days.
Click for the full walkthrough →

Codex fix in a worktree (when a cluster matures)

When a cluster reaches 3+ tickets and is mechanical, the fix gets drafted in a git worktree with PROPOSAL.md + the actual code change. Engineer reviews. Never touches main without approval.

feedback_themes.md → GitHub issues → Recent commits → Codex Fix Drafter
Before: Bug fix takes 3 weeks from 'we found it' to 'shipped'. By then customer churned.
After: Mechanical fixes draft within a session of identification. Engineer review = 30 min. Ship same day.
Click for the full walkthrough →

Weekly product report (Monday standup)

Devon and Maya start Monday with a single report: what customers said, what churned, what's drifting on roadmap. Engineering knows what to work on.

feedback_themes.md → churn_root_cause.md → product_roadmap_signal.md → Triage Orchestrator
Before: Monday standup ran on Devon's gut feel. Engineering chose work based on whoever shouted loudest.
After: Standup runs on data. Engineering picks the highest-impact cluster. Roadmap drift surfaces before it costs MRR.
Click for the full walkthrough →

Churn investigation (when a customer cancels)

When a customer cancels, root-cause-analyst investigates within hours, not days. Surfaces the support history + error timing + roadmap miss that drove the churn.

Stripe (cancellation) → Intercom history → Sentry errors → churn_root_cause.md → Root-Cause Analyst
Before: Churn investigations took 2 days each because the data lived in 5 places. Often skipped entirely.
After: Investigation drafts within hours. Devon reviews in 15 min. Patterns surface.
Click for the full walkthrough →

LedgerLink, here's the order to build this in. Each step builds on the last, total: A focused weekend. Step 1 unblocks everything. Steps 2-3 are the bulk of the value.. If a step looks like a cliff, hand it to @claude-code-guide.

  1. 0
    Pre-step (skip if you already have Claude Code)

    Get Claude Code on your machine

    What

    Install the Claude Code CLI. It's the tool that runs everything below. If you've never opened Terminal, that's fine, this is the only command-line moment until you're set up.

    Install
    npm install -g @anthropic-ai/claude-code
claude --version  # should print a version number
    You'll know it's working when: typing claude --version in Terminal prints a version number instead of an error.

    Setup time: 5 minutes.

  2. 1

    Wire the conversion hook + ingest_intercom_tickets

    What

    SessionStart hook for raw_dropzone. Then write ingest_intercom_tickets skill that pulls last 7 days of conversations.

    Why this first

    Intercom is your highest-volume signal. Without ingestion, the triage orchestrator has nothing to read.

    Install
    # 1. Hook setup (5 min)
brew install pandoc poppler && pip install xlsx2csv
# add SessionStart entry to .claude/settings.json

# 2. Skill
# write under .claude/skills/ingest_intercom_tickets/
# uses INTERCOM_ACCESS_TOKEN in ~/.env
# writes to data/feedback/intercom_tickets.csv
    Before: Tom dumps tickets in Slack each Monday. Devon reads them by hand.
    After: Tickets land in data/feedback/ overnight. Triage is ready Monday 7am.

    Setup time: Half a day for the skill

  3. 2

    Build feedback_themes silver platter (cluster_themes skill)

    What

    Skill that reads Intercom + Pendo + customer calls + GitHub. Clusters by root cause. Writes silver_platters/feedback_themes_.md.

    Why this first

    Replaces 200+ tickets/responses with 5 themes + verbatim quotes. Devon reads 5 things, not 200.

    Install
    # write under .claude/skills/cluster_themes/
# reads from data/feedback/* and data/product/github_issues.csv
# writes silver_platters/feedback_themes_.md
# scheduled via cron Sunday 11pm
    Before: Pendo dumps unread. Devon never sees the cluster pattern.
    After: Mon 6am brief shows top 5 clusters with churn-risk flags. Engineering picks highest-impact.

    Setup time: Half a day

  4. 3

    Triage Orchestrator + Root-Cause Analyst + Codex Fix Drafter

    What

    Three subagents under .claude/agents/. Orchestrator routes. Root-Cause analyzes. Codex drafts the fix in a worktree (never on main).

    Why this first

    The orchestrator hierarchy makes the system usable. Devon talks to Triage; specialists report up. Worktree pattern keeps Devon from worrying about AI pushing bad code.

    Install
    # write under .claude/agents/:
#   triage-orchestrator.md (model: opus, tools: [Read, Glob, Grep, Task])
#   root-cause-analyst.md (model: opus, tools include Bash for git log)
#   codex-fix-drafter.md (model: opus, tools include Edit, Write, Bash; write only in worktrees/)
# Hand to @claude-code-guide for templates.
    Before: Devon is the human router. Pickup latency 3 weeks.
    After: Mechanical fixes draft same-session. Engineer reviews PROPOSAL.md, approves via audit log.

    Setup time: 2-3 hours

  5. 4

    Audit log + approval-gate hooks (Stop warns on unsigned PROPOSAL)

    What

    PostToolUse hook for every Edit/Write. Stop hook that warns if a worktree's PROPOSAL.md is unacknowledged at session end. Permissions deny: never push to main.

    Why this first

    Trust layer. Engineer review is non-negotiable. Permissions config is the second line of defense.

    Install
    # permissions in .claude/settings.json:
# "deny": ["Bash(git push*)", "Bash(git merge*)"]
# hooks per business_os_hooks.md
    Before: AI wrote code? You don't know what. AI pushed code? Yikes.
    After: Every action recorded. AI cannot push. Engineer's [approved] line is the merge trigger.

    Setup time: 10 minutes

  6. 5

    Add /weekly_product_report and /churn_investigation slash commands

    What

    Two slash commands under .claude/commands/. Each triggers the right agent chain.

    Why this first

    Turns Devon's recurring motions into one-keystroke triggers. Standup-ready report Monday morning. Churn investigation drafts within hours.

    Install
    # write under .claude/commands/:
#   weekly_product_report.md
#   churn_investigation.md (takes  as arg)
    Before: Each Monday is a fresh prompting session.
    After: Devon runs /weekly_product_report at 7am, walks the team through it at 9am.

    Setup time: 30 minutes

  7. You've shipped your back-of-house.

    Next: hand @claude-code-guide a one-line ask for your next skill, or join the Early AI Dopters community to see how other operators evolved theirs.