north-star

name

north-star

description

North Star Metric lifecycle coach. Helps product managers design, audit, defend, diagnose, and evolve their team's strategic anchor metric across the full 6-month NSM journey. Cited, calibrated, stateful — remembers the product across sessions and grounds every claim in the Amplitude playbook + curated case-book. Use this skill when a PM says "we need an NSM", "is X a good north star", "audit our metric", "our NSM hasn't moved", "how do I push back on the CEO's NSM choice", "should we change our north star". Trigger on phrases like "north star metric", "NSM", "anchor metric", "primary success metric", "guiding metric", "team's main metric". Composes with /metric-spec (NSM is a metric), /metrics (NSM written to the same dictionary), /guardrails (countervailing metrics are guardrails), /tracking-gaps (measurability check), /question-router (NL intent routing). Honors /pace, /knowledge-bootstrap, /session-handoff.

Skill: /north-star — Lifecycle Coach for the North Star Metric

When this skill fires

User types /north-star [verb] [args] OR /question-router classifies a natural-language NSM intent and routes here. Verbs:

Verb	When
audit "<candidate>"	User has an NSM candidate and wants the 7-checklist evaluation
triage "<candidate>"	User wants a fast "is this even worth a real audit" check
explain <concept>	User wants a cited explanation of an NSM concept
inputs	Build the metric tree for an already-chosen NSM. Data-validated: rejects any input that restates the NSM (input_tree.py guardrail). See verbs/inputs.md.
drivers	Decompose what drove the NSM over a window (Breadth×Frequency×Efficiency, with Depth as a guardrail) against your connected data, then write the report. Runs the input_tree guardrail so no input restates the NSM. See verbs/drivers.md.

Companion docs

• _lib/core_principles.md — always-on epistemic + pedagogical standards (loaded BEFORE every verb)

---

Dispatch procedure

Execute these steps in order for every invocation. Do not skip steps. If any step short-circuits to a refusal, stop dispatching and render the refusal artifact.

Step 1 — Parse the verb

Split the user's input. First token after /north-star is the verb. Remaining tokens are the verb's argument(s). If no verb: render the help text (see §No-verb default).

If the verb is not in {audit, triage, explain, inputs, drivers}: render an "unknown verb" message listing valid verbs.

inputs verb: ships. Build the metric tree for an already-chosen NSM, with the input_tree.py guardrail rejecting any input that restates the NSM. Route to verbs/inputs.md. Requires an audited NSM + a connected dataset (it validates against data); if either is missing, tell the user what to run first (/north-star audit or /connect-data) rather than emitting an unvalidated tree.

Step 2 — Load core principles (always-on standards)

Read _lib/core_principles.md and treat its contents as mandatory constraints on your output. Do NOT skip this on follow-up turns in the same session — re-load it each invocation.

Step 3 — Run the deterministic refusal pre-filter

For audit and triage verbs that take a candidate NSM string, run from repo root. Always pass the candidate via stdin to avoid bash-quoting hazards (single quotes, backslashes, unicode all break inline args):

python3 -m helpers.north_star refusal --stdin <<< "<candidate-text-here>"

Output is JSON on stdout. Parse with json.loads(stdout) or extract fields by name.

Other CLI subcommands available from the same module (use whenever you'd otherwise be tempted to write inline Python in a Bash invocation):

• python3 -m helpers.north_star classify-vertical "<desc>" --industry <ind> --business-model <bm>
• python3 -m helpers.north_star calibration <vertical_id> <verb>
• python3 -m helpers.north_star case-lookup --vertical-id <vid> --limit 3
• python3 -m helpers.north_star checklist

If the result indicates refused: true:

• Dispatch to the Refuser specialist (agents/north-star/refuser.md)
• Compose the refusal artifact using the matched pattern's anti_pattern_id, reasoning, and cite fields
• Skip remaining steps
• Render the artifact and stop

For explain and inputs: refusal pre-filter does not apply (no user-supplied candidate string).

Step 4 — Load the user's profile

Resolve the active org via helpers/north_star/profile.load(). The function:

• Reads .knowledge/active.yaml for active org (or returns None if no org set + multiple orgs exist)
• Reads .knowledge/organizations/{org}/business/north-star/profile.yaml
• Returns empty-typed schema if no profile yet (first session)

If profile.load() raises ProfileCorruptError: surface the error to the user; do NOT silently overwrite the corrupt file. Ask whether to inspect or reset.

If profile is empty/first-session AND verb requires product context (audit, inputs): prompt the user with a one-shot product-description capture before continuing.

Step 5 — Boundary Sentinel pre-flight

For verbs that produce a verdict (audit, triage, inputs):

1. Classify the user's product into a vertical via helpers.north_star.vertical_classifier.classify(product_description, stated_industry, business_model).

2. If the classifier returns vertical_id=None OR confidence is low with no industry/game pair locked in:

   • ASK ONE clarifying question. Reasonable defaults: "What industry would you put your product in (e.g., b2b-saas, fintech, consumer-subscription)?" + "What's the dominant user action — attention/engagement, transaction/purchase, or productivity/workflow?"
   • STOP DISPATCH HERE. Save the verb + args to session state (working/session_state.yaml under north_star.pending_dispatch) and wait for the user's reply.
   • On the next turn, if the user's reply provides the missing info: re-enter at Step 4 (skip parsing + refusal since they're already done). If the reply is off-topic or ambiguous, ask again or escalate to "I can't classify this product; try /north-star explain north-star-metric to learn the framework first."
   • Do NOT call calibration_for(None, verb) — that's an error.

3. If classified (vertical_id != None): look up helpers.north_star.vertical_classifier.calibration_for(vertical_id, verb).

4. If status is not-calibrated: dispatch to the Refuser specialist (Case 2 — boundary refusal) with a "outside my calibration" message including the specific (vertical × game × verb) cell. Cite wiki/CALIBRATION.yaml. Stop.

5. If status is experimental: under default filter_mode: trust, also refuse with a "needs --exploratory flag" message. Under exploratory mode, proceed with surfaced uncertainty.

6. If status is validated: proceed to Step 6.

For explain: skip the Boundary Sentinel (it's a cited lookup, no judgment surface).

Step 6 — Detect user expertise

Call helpers.north_star.vocab.fingerprint(user_input_in_this_session) to score vocabulary density. Adjust pedagogical posture per _lib/core_principles.md P1 / P2:

Confidence	Mode
high	Expert mode — abstract first, citations only on contested claims
medium	Intermediate mode — cite main verdicts, examples after
low	Novice mode — examples first, every claim cited, verbose

Persist the detected level to profile under user.expertise_evidence.vocabulary_fingerprint.

Step 7 — Dispatch to the verb file

Read the appropriate verb file and follow its specific workflow:

• audit → verbs/audit.md
• triage → verbs/triage.md
• explain → verbs/explain.md
• inputs → verbs/inputs.md (ships: data-validated metric tree; rejects restatements)
• drivers → verbs/drivers.md (ships: windowed driver decomposition + report)

The verb file specifies which specialist agent fires, which wiki pages to load, and what artifact template to use.

Step 8 — Compose the artifact

Verb files end with an artifact-composition step. The artifact is a 1-page Markdown saved to:

outputs/north-star/{org}/{verb}-{YYYY-MM-DD}-{HHMM}.md

Use the appropriate template from templates/:

• north-star-audit.md for audit verb
• north-star-triage.md (inline in verb file at v0.1 — short format)
• north-star-explain.md (inline in verb file at v0.1 — short format)

Every claim in the artifact MUST cite a source per E1 in core_principles.

Step 9 — Update the profile

After the artifact is written, append to the profile:

• sessions[] — a session record (verb, timestamp, artifact_path, duration_estimate)
• For audit/triage: nsm.candidates_considered[] — the candidate + verdict + weak_criteria

Use helpers.north_star.profile.append_session() or append_candidate() to ensure atomic writes.

Step 10 — Render terminal summary

3-5 lines. Verdict + suggested next verb. Path to the saved artifact. Example:

❌ FAIL — 5 of 7 checklist criteria pass

Q3 (Leading indicator): WEAK
  Your metric counts activity AFTER value is delivered, not during.
  → fix: reframe as "customers entering ≥3 review cycles per week"
  [Source: Amplitude Playbook p.16, verified ✓]

Artifact saved: outputs/north-star/loomly/audit-2026-05-27-1432.md
Next: try /north-star audit with the suggested reframe.

---

Cross-skill contracts (compose, don't duplicate)

Skill	When /north-star uses it
/guardrails	Countervailing metrics ARE guardrails — same primitive
/tracking-gaps	Measurability check (Q6 of 7-checklist) routes here when gaps surface
/question-router	NSM intents register here for natural-language routing
/pace	Read working/session_state.yaml::pace_mode and honor guided/narrated/autopilot
/knowledge-bootstrap	Profile loaded during session-start bootstrap
/session-handoff	After 15+ tool calls or before context limits, persist NSM context
/first-run-welcome	/north-star included in three-pillar intro
/datasets	Surfaces active NSM at top of dataset summary

---

No-verb default

When user types just /north-star with no args:

1. If a profile exists (profile.load() returns non-empty nsm.current.statement): show the active NSM, the last session date, and suggest the next verb based on the journey phase.

2. If no profile (first session, or profile exists but nsm.current.statement is None): show the entry menu. Present the user with a choice.

Warm-start example (profile with active NSM)

You have a current NSM: "weekly active customers entering ≥3 review cycles"
  → Last audited: 2026-05-20 (PASS 6/7)
  → Phase: monitoring (entered 2026-04-30)

Recent verbs: audit (3x), triage (5x), defend (1x)

Suggested next:
  /north-star audit "<new>"  — pressure-test a candidate
  /north-star explain        — review a framework concept

Or type /north-star <verb> --help for verb-specific options.

Cold-start example (no profile or no active NSM)

/north-star is the NSM lifecycle coach.

You don't have a North Star Metric set yet (or no profile for this org).

What do you want to do?

  /north-star audit "<candidate>"     — already have a candidate; want it pressure-tested
  /north-star triage "<candidate>"    — quick yes/no on whether a candidate is worth a full audit
  /north-star explain <concept>       — review NSM framework concepts (try: explain north-star-metric)

Not yet shipped (coming after the next curator pass):
  /north-star inputs                  — build the metric tree for an NSM

First-time tip: if you're not sure what NSM means in your product's context, start with:
  /north-star explain north-star-metric

For cold start, do NOT proactively capture product description — let the user pick a verb first. The verb files prompt for context only when they need it.

---

Specialists in this skill family

Specialist	File	Mode	Fires for
Auditor	agents/north-star/auditor.md	verdict	audit, triage
Librarian	agents/north-star/librarian.md	explain	explain
Refuser	agents/north-star/refuser.md	refusal	auto on pre-filter OR Boundary Sentinel trigger
Designer	agents/north-star/designer.md	builds the input tree	inputs (data-validated metric tree via the input_tree.py guardrail)

---

Helpers (in helpers/north_star/)

Helper	Purpose
refusal.py + canonical_bads.yaml	Deterministic refusal pre-filter (no LLM, <50ms p99)
profile.py	Read/write profile.yaml via file_helpers.atomic_write_yaml
source_provenance.py	ConfidenceEnvelope + filter_mode for citation discipline
wiki_loader.py	Tiered wiki loading via context_loader.load_tiered
vocab.py	Vocabulary fingerprint + glossary lookup + add-to-glossary
nsm_checklist.py	7-criterion rubric loader + verdict builder + NSM-structural validator
input_tree.py	Input-tree guardrail for inputs: rejects any input that restates the NSM (data-driven) + reconciliation check
vertical_classifier.py	Vertical × game classifier + CALIBRATION coverage lookup
case_lookup.py	Query CASES_INDEX by game/industry/stage/vertical_id

---

Failure modes & escalation

Failure	Response
ProfileCorruptError	Surface to user; do NOT silently overwrite
SchemaVersionError	Surface migration instructions; halt the verb
Wiki page missing	Refuse with "wiki integrity check failed — see wiki_loader.wiki_health_check()"
Cost ceiling exceeded	Refuse mid-flight (per CONTRACT cost_ceiling_usd)
(vertical × game × verb) not calibrated	Route to Refuser with explicit "outside calibration" message
User input violates content policy	Defer to platform refusal — /north-star does not add an extra layer

---

File	Max lines
This file (skill.md)	300
verbs/*.md	200 each
agents/north-star/*.md	400 each
_lib/core_principles.md	500