/northstar is how I work, written down. It lands in your repo, senses the environment, and executes in small visible loops. Free, because I trust it enough to let your team run it without me.
Give it a problem area and it owns it end to end. It reads your repo, drafts a plan you can redirect, then works in loops. Specialised problems get delegated to the sub-skills below, each of which also runs on its own.
Senses the environment, states its read, proposes a plan you confirm.
Warm, hedged PR descriptions. Disagreement with evidence, never corporate.
Stakeholder work, executive packs, greenfield setups, funnel instincts.
Proposes a PR to its own skill file whenever it hits a gap.
It starts the way a good consultant starts: read, then ask.
Reads your repo, conventions and access first. Never asks what the code already answers.
Five to eight questions that encode a decade of scar tissue, not "what is your BI tool".
Writes a SKILL.md with your real paths and rules. Versioned like any other code.
Hits a gap, asks a human, proposes a PR to itself with what it just learned.
One job each. /northstar delegates to them, but every one also runs on its own.
Traces the break to the dbt model, checks for fanout, drafts the PR in your review style.
Asks which tools and what access it has, then builds conversion rules around that pair.
Issues verdicts from your real usage data. Nobody misses dashboards they can't name.
Learns your conventions from your best models, then checks a duplicate does not exist.
Traces two figures to where their definitions diverge, in language a CFO accepts.
A named owner, traceable metrics, a lifecycle stage, before anything gets published.
Session one ends with a merged PR, not a slide deck.
Built from the questions your team actually asks, with a staleness check on every page.
Classifies questions, delegates to your other skills, routes to the right owner.
Assembles the pack from governed metrics only. No ad-hoc number reaches the deck.
Makes every request state the decision it serves before it enters the queue.
Draft, Verified, Drifted, Archived. The audit is an event. Rot is a process.
The questions are the product. Anyone can ask what your BI tool is. A builder asks which models must never be auto-fixed.
❯ /build-fix-chart
Inspecting repository… found 2 dbt projects, Lightdash meta blocks,
CODEOWNERS on /models/finance. 3 of 8 sampled models lack unique
tests on primary keys, fanout risk noted.
Q1. When a chart breaks, where does the report land first?
❯ #analytics-support
Q6. Which models must never be auto-fixed?
❯ anything under models/finance/
Writing .claude/skills/fix-chart/SKILL.md, with models/finance/
on the blocklist and diagnoses formatted for #analytics-support.
Plenty of people will tell you their AI is safe to trust. I would rather you read mine. If the free version makes your team faster, you have just previewed how I work. Here is what that preview does not include.
When your team wants the version built from weeks inside your workflows, that starts with a 30-minute call.
/plugin marketplace add NorthStar-Analytics-LTD/skill-foundry
/plugin install skill-foundry@northstar-analytics
Prefer a one-off copy? Download the .zip instead.
The free edge of my BI Platform Migration work: the same Claude Code skills I build inside engagements.