Skip to content
Agent Orchestration
Skill

/lamina-init

Use only when explicitly invoked as lamina-init. Establish business evidence and canonical Product, Actor, and Persona resources in Lamina's transactional graph.

BOOST
From plugin
lamina
12010 skills
Install
$ npx -y skills add aryaniyaps/lamina --skill lamina-init --agent claude-code

How it fires

How this skill gets triggered: by you, by Claude, or both.

  • Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
  • You can call itInvoke it directly when you want it.
  • Slash command/lamina-init

Context preview

The summary Claude sees to decide when to auto-load this skill.

Use only when explicitly invoked as lamina-init. Establish business evidence and canonical Product, Actor, and Persona resources in Lamina's transactional graph.

SKILL.md

lamina-init.SKILL.md
name: lamina-init
description: "Use only when explicitly invoked as lamina-init. Establish business evidence and canonical Product, Actor, and Persona resources in Lamina's transactional graph."

/lamina-init

Before any evidence or graph mutation, read and apply `../lamina/orchestrator/prerequisites/cli-required.md`.

Read `../lamina/orchestrator/audit-profiles.yaml`. Load the `init.always` references, then only those `init.conditional` references whose `when` signal is present. For each selected entry, open the named capability `SKILL.md` and then its exact reference; do not load sibling topics.

For a directory that is not yet a Git project, initialize Git metadata as specified by that prerequisite before writing evidence. Do not ask for or create an initial commit.

Init writes evidence files only under `.lamina/` and canonical knowledge through graphd. It never edits application source.

Evidence artifacts

Write `.lamina/business-context.md` with frontmatter containing `lamina.maturity`, `platform`, and `last_updated`, followed by exactly these non-placeholder sections: Problem statement, Business goals, Success metrics, Scope, Users & market, Product posture, Constraints, Stakeholders, Risks & unknowns, Research posture, Triad check.

Write `.lamina/personas.json` as evidence-source JSON with evidence-grounded personas. Goals, constraints, and evidence are arrays. Do not invent demographics. These files are indexable evidence, not canonical graph state.

Run the shipped init/persona validators when available, then `lamina graph observe` so CocoIndex produces explicit source Observation envelopes. The CLI automatically replaces an incompatible graphd and retries observation once; do not repeatedly rebuild observations to repair daemon compatibility.

If observation exits nonzero after that recovery, treat observation as degraded and optional. Core graph initialization may continue, but do not create Evidence Resources or attach evidence claims for the unavailable snapshot. Do not describe aliases, observation view names, or generation labels as evidence.

When observation succeeds, use `observed.resource_ids` from its output (or the exact `Resource.id` values returned by querying that active observation view) as graph evidence. Never substitute a path alias, source key, view name, or generation for an Observation Resource id.

Canonical graph

Start one explicit session. Propose:

  • one inferred Product Resource proposal grounded in the user's explicit product intent;
  • every evidence-grounded Persona as a Persona Resource;
  • corresponding Actor Resources when authority/ownership is known;
  • `lamina:canAssume` Statements between Personas and Actors;
  • Evidence Resources referencing the relevant source observations.

Publish atomically. Agents must not submit epistemic class or approval. Never cap Personas. Record the actual returned Resource ids for Product, Personas, Actors, and Evidence; generated canonical Resources normally use `res_*` ids. Do not report input aliases as ids.

All agent-accessible proposal methods use inferred ingress, including `claim.add`. Never select an epistemic class by choosing a method name. Intended knowledge requires a trusted engine-owned intent ingress; until that ingress supplies it, preserve the user's words as provenance and keep the proposal inferred.

Update mode

Merge changed business evidence, append a dated changelog, rerun observations, and propose new Statements or aliases without replacing stable Resource identity. Preserve conflicting valid facts as Contradictions.

Completion

Report canonical graph initialization and observation as separate outcomes. Always report the GraphVersion, source revision, actual Product/Persona/Actor Resource ids, contradictions, and evidence gaps. Install or refresh passive provider rules with `lamina setup --agent <provider>`, then state that future ordinary product requests automatically prepare graph-backed implementation context, complete design gaps, implement, and verify. Do not recommend another slash command.

Only report observation coverage and Observation Resource ids when `lamina graph observe` exited zero and its completion checks passed. After a nonzero observation command, report observation as degraded/unavailable and explicitly state that the published GraphVersion has no current observation-backed evidence. Never claim complete observation coverage on that path.

Read more
Ships withlamina

Headless product design for AI coding agents, backed by a transactional product graph | Design how it works, verify what you ship.

Get the whole plugin

Other skills on lamina.