capx-ai/casa/casa-start

casa-start

The front door. Takes a founder to a confirmed, personalized build map that starts at the right level. It adapts to where it is run: a clean folder gets the full interview; a folder that already holds a project gets read first, so the founder confirms and fills gaps instead of answering everything cold.

Steps

1. Read the directory. Detect whether this is an existing project:

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/scan.mjs . ``

It returns is_existing_project and deterministic signals (repo, deployed app, payments, auth, analytics dependencies, a type hint). The company-brain/ directory is excluded, so an initialized-but-empty folder still reads as greenfield.

1. Initialize the workspace if needed. If company-brain/ does not exist, create it:

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/brain.mjs init company-brain ``

Confirm before writing. This does not touch the founder's own files.

1. Build the profile, by branch. The interview bank

${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/skills/casa-start/questions.json has four passes: Define (what and who), Locate (stage tier), Core (north star, do-or-die constraint, horizon, anti-priorities, funding), and Backfill (skipped foundations). Skip any question whose ask_when is not met.

A. Empty folder (is_existing_project is false). Run all four passes (Define + Locate + Core + Backfill) conversationally, in about 6 to 9 interactions, never as a cold questionnaire: i. Ask the one_liner and the stage tier first. These two answers carry most of the signal. ii. Infer, do not interrogate. From the one_liner, INFER every low-yield Define answer you can: type (exactly one of saas, marketplace, ecommerce, b2b-service, crypto, consumer, content, hardware), the trait questions (software, local_service, sends_email, collects_data, technical_audience, runs_paid_media, runs_agents), monetization if the one-liner states it, and audience. Present ALL inferences in ONE batch confirmation the founder can correct in a single reply ("Here is what I took from that; correct anything wrong"). Never ask these as individual questions. Ask individually only what genuinely cannot be inferred. iii. Draft-map preview. After Define + Locate are confirmed, write the partial answers so far to company-brain/answers.json and derive a draft (the Core fields may still be absent; stage.mjs derive accepts that):

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/stage.mjs derive company-brain/answers.json ``

Show a DRAFT plan preview: the top 4 or 5 plays in plain words and the starting level, with one line: "Here is the shape of your plan. A few more questions to tune it." iv. Then run the Core pass as today (north star, do-or-die constraint, win definition, horizon, anti-priorities, and funding where it applies), plus any Backfill question whose ask_when holds.

B. Existing project (is_existing_project is true). Do not interview cold. Let the scan do the Define and Locate work, and spend the budget on the Core pass plus a gap confirmation, targeting about 4 to 6 interactions. i. Deep read. Spawn the project-scanner agent with this directory and the scan.mjs output. It reads the README, root CLAUDE.md, docs, manifests, and source, and returns an inferred profile (type, audience, monetization, traits, stage tier, gaps) with evidence and a confidence on each field, plus a tight ask_founder list (revenue-truth, funding, north star, do-or-die constraint, and done/gap reconciliation for the four high-value unblockers). ii. Confirm Define and Locate in one batch. Use the scan-aware keys in questions.json: a question marked mode_when_scanned: "skip" whose satisfied_by_signal is present is taken silently (the code proved it); a question marked mode_when_scanned: "confirm" is pre-filled from its prefilled_by field and shown for the founder to eyeball in a single confirmation, never asked cold. Show, in plain language, what you learned about their business and its stage and what already looks done. The tier question carries revenue_truth: never set launched, revenue, or scaling from the scan alone (code cannot prove paying customers); the founder must confirm it. iii. Run the Core pass and reconcile gaps. The Core pass is never scannable, so always ask it (north star, constraint, win and horizon, anti-priorities, and funding only if not already known). For Backfill, reconcile only the gaps the scanner flagged, focused on the four high-value unblockers (analytics, payments, legal/ToS, deploy); do not ask the whole battery cold. Read the founder's own files for context. Never edit or overwrite them. Casa's state lives only in company-brain/.

In the Core pass, the north-star question offers business-type chips (mirroring the north-star families in scripts/northstar.mjs via chips_by_type and chip_overrides), then you map the founder's free text to exactly one canonical north_star_archetype id. The constraint question maps free text to one constraint_archetype id. Map anti-priorities to playbook ids only where you can name them confidently; otherwise leave the array empty.

1. Assemble the answers and write company-brain/answers.json. Produce every field,

greenfield or existing:

`` { "type": "...", "secondary_type": "", "company_name": "...", "one_liner": "...", "icp": "...", "monetization": "...", "traits": [...], "tier": "...", "gaps": [...], "north_star_archetype": "...", "constraint_archetype": "...", "win_definition": "...", "horizon": "...", "anti_priorities": [...] } ``

north_star_archetype is one of activation, engagement_retention, revenue_mrr, acquisition_growth, conversion, gmv_liquidity, efficiency_unit_econ, local_reputation. constraint_archetype is one of no_users, no_revenue, runway_burn, regulatory_legal, tech_scale, hiring_capacity. horizon is one of this_week, 30d, 90d, quarter. anti_priorities is an array of playbook ids and may be empty. Use only canonical values; stage.mjs validates the type, traits, tier, and gaps and errors on drift.

1. Branch on the stage. Idea stage (tier is "idea") runs the validation gate via the

casa-validate skill (GO or KILL). Any other stage skips validation; the business already exists and is not sent back to Level 0. An existing project is almost never idea stage.

Dated 2026-08-21: idea-stage companies run Phase 0 distribution foundation first. The engine headlines phase0-website at level 0 (existential_at: idea). Do not seed those four as already-done. An existing business keeps its stage floor; incomplete Phase 0 is catch-up, never a regression to Level 0, and never a silent publish.

1. Derive the profile, level, seed, and the initial pulse, then render the brain:

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/stage.mjs apply company-brain/answers.json company-brain node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/brain.mjs sync company-brain ``

stage.mjs apply writes profile.json and state.json (the seed and the level floor) AND seeds the initial pulse into company-brain/pulse.json from the Core answers: north_star_archetype sets the department weights, constraint_archetype adds a department delta, anti_priorities become demote ids, and win_definition is carried for display. This makes the very FIRST build map business-aware (a retention-focused founder sees retention work lead) before any manual pulse cascade. It will not clobber a richer hand-authored pulse.json if one already exists.

Then derive the department roster and seed the autonomy dials:

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/roster.mjs company-brain --write ``

This reads the company type and traits (profile.json) and the binding constraint (state.json), derives only the departments this business needs (with the constraint's lead departments marked), and writes company-brain/roster.json plus company-brain/dials.json (per-department autonomy, defaulting to approve-first, with the always-ask line: spend money, go public, merge to main, destructive). Show the founder the derived departments and operators and let them override before continuing.

1. Deepen the founder's pulse (optional). The Core pass already captured the north star,

the do-or-die constraint, the win and horizon, and the anti-priorities, and stage.mjs apply already seeded a business-aware pulse.json from them. This step only DEEPENS that seed into the full hand-authored pulse; it does not re-ask what Core covered. Build on the Core answers and the project read, drilling into the still-empty checklist slots in short adaptive rounds, never asking beyond what you need:

• Reflect the Core answers back ("your north star is retention, the constraint is

runway") so the founder sees they were heard, then fill what is missing: the ranked focus areas, the priority order, risk appetite, resourcing, and the one thing.

• Each later round is generated from the prior answers and the project read, asking

only about the still-empty checklist slots. Guardrails: this whole step is skippable ("good enough, build my plan" leaves the seeded pulse in place); cap at five rounds; reflect their last answers back so the questions feel earned, not like a form.

The pulse checklist to fill: focus (ranked areas), win (definition and horizon), constraint, anti_priorities, priority_order, risk_appetite, resourcing, one_thing.

Write two files:

• company-brain/pulse.md: the human-readable pulse.
• company-brain/pulse.json: the structured pulse PLUS the priority weights it

implies, which the engine consumes:

`` { "focus": [...], "win": { "definition": "...", "horizon": "..." }, "constraint": "...", "anti_priorities": [...], "risk_appetite": "...", "resourcing": "...", "one_thing": "...", "weights": { "byDepartment": { "Growth": 0.4, "Finance": 1.6 }, "byLevel": { "4": 0.4 }, "promote_ids": [...], "demote_ids": [...], "default": 1.0 } } ``

Use the canonical 11 department names: Strategy, Brand, Product, Engineering, Data, Growth, Sales, Finance, Legal, Success, Operations. These are the values authored on every playbook. The engine normalizes the common aliases (Marketing to Growth, Design to Product, Support to Success) and warns on any unrecognized name, but write the canonical names directly. Down-weight what they are not doing yet, up-weight their focus. Be decisive but conservative: the weights nudge the order, your judgment in casa-next makes the final call.

Then re-render so the weights take effect:

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/brain.mjs sync company-brain ``

1. Show the build map with casa-map and get approval. Call out the current level, the

catch-up items, and the critical path.

1. Hand off to /casa (the front door). It opens the first working session: a plain

briefing on where the company is and the single move to make, asked before anything runs.

Rules