empiricaai/empirica/ewm-interview

/ewm-interview — Epistemic Workflow Manager Interview

Produce a workflow-protocol.yaml describing how this person wants to work with AI — then actually set up what it implies.

The rule that governs the whole interview

Ask with options, never with a blank. Every question goes through AskUserQuestion with 2–4 concrete, pickable options. The tool supplies its own "Other" path, so a free-form answer is always available without you offering one — do not add an "Other"/"Something else" option yourself, it duplicates the built-in.

This matters more than it looks. "What are your main constraints?" asked into a void gets a shrug or a paragraph you cannot map to a field. The same question as four options gets an answer in one click, and the options themselves TEACH what the field means. Prose questions are what made v0.1 a form-fill wearing a conversation's clothes.

Three rules for the options you write:

• Make every option genuinely pickable. An option nobody could choose is

filler; it makes the list look considered while narrowing the real choice.

• Put a recommendation first when there is a sane default, and say so:

"Collaborative — check in on approach (Recommended)".

• `multiSelect: true` when answers are not exclusive (domains, tools,

non-negotiables). Most of this interview is multi-select.

Batch related questions into ONE AskUserQuestion call — it takes up to 4. Five calls of 3–4 questions each covers the whole interview.

Phase 1 — Goals

One call, up to 4 questions:

Capture into goals.primary[] (description, success_criteria, timeline) and goals.secondary[].

Phase 2 — Domains & expertise

Infer the option list — do not ask people to type their own field. Read the repo first: languages present, docs/ topics, the project's own .empirica/project.yaml. Offer what you found. This is the single biggest friction reduction in the interview.

Capture domains.expert[], domains.learning[], domains.novice[].

Phase 3 — Tools & connections

Detect before asking. Read ~/.claude/mcp.json (or the harness equivalent) and offer what is configured rather than a generic menu. Mark detected entries so the user is confirming, not recalling.

The last two questions are load-bearing — they feed Phase 6.

Phase 4 — Work preferences

Capture work_preferences.* and the three task_splitting lists.

Note the asymmetry: "act without asking" is a floor, "always wait" is a ceiling. If an item appears in both, the ceiling wins — say so rather than silently resolving it.

Phase 5 — Trust & non-negotiables

Capture trust_building.*.

Phase 6 — Provision what the answers imply

This is the part v0.1 was missing. An interview that produces a YAML file and stops has described a setup rather than performed one.

empirica provision-practice already does the whole chain, idempotently: mkdir → project-init → patch .empirica/project.yaml (ai_id / tenant / org / substrate) → project-register with Cortex → optional Forgejo backup remote. Safe to re-run; every step no-ops if already done.

So: from Phases 1–3, propose the practices their answers imply — a separate practice per distinct domain of work, which is the model Empirica is built on (a practice is an epistemic specialization, not a folder).

Always dry-run first, and show the output before doing anything:

empirica provision-practice <name> --dry-run --output json

Then ask — with options, like everything else:

Provision these? — empirica-research, empirica-outreach

· Yes, both · Just the first · Let me adjust the names · No, not now

Only on an affirmative:

empirica provision-practice <name> \
  --tenant <tenant> --org <org> \
  --cockpit-profile <tenant> \
  [--forgejo-owner <owner> --forgejo-host <ssh-url>] \
  [--no-cortex]

--cockpit-profile adds each practice to one cockpit layout (~/.empirica/cockpit/config-<name>.yaml, created on the first call with a TUI window, then one window per practice). Pass the SAME name on every call. The tenant slug is a sensible default; use another if they already have a profile.

Where the flags come from Phase 3:

Report per practice what actually happened — provisioned, already existed, or failed and why. A rollup "done!" over a partial failure is the shape this whole system exists to prevent.

Then hand them the cockpit. When at least one practice was provisioned, end the phase with the command the last call printed, in a block they can copy:

empirica cockpit launch --profile <name>

Say what it does in one line (one window per practice, each starting claude in its own directory with a stable identity, plus a TUI window) and point at docs/guides/COCKPIT.md for changing it. empirica cockpit refresh restarts a practice whose Claude exited. Do not launch it for them: it opens their terminal. If no practice was provisioned, or a cockpit-profile step failed, say that instead and do not print a launch command for a profile that was not written.

Output

Where to write it — ask, do not assume

The loader searches project directory first, then `~/.empirica/`. That ordering has a consequence worth stating out loud, because it is silent:

A project-local protocol SHADOWS the user's global one for that project.

That is a feature when someone wants different working agreements on a client repo, and a trap when they meant to update their global profile and quietly stopped using it everywhere else.

So ask, with the default first:

Where should this live? · ~/.empirica/workflow-protocol.yaml — applies

everywhere (Recommended) · This project only — overrides your global protocol

here

If they pick project-local and a global one already exists, say what will be shadowed before writing.

Shape

# Epistemic Workflow Protocol
# Generated by EWM Interview v0.2.0 — {date}
 
user_profile:
  name: "{name}"
  role: "{role}"
  created: "{date}"
  last_updated: "{date}"
 
goals:
  primary:
    - description: "{goal}"
      success_criteria: ["{criterion}"]
      timeline: "{timeline}"
  secondary: ["{goal}"]
 
domains:
  expert: ["{domain}"]
  learning: ["{domain}"]
  novice: ["{domain}"]
 
tools:
  {category}: "{name}"        # Maps to: {mcp_server}
 
mesh:                          # Phase 3 — omit the block entirely if standalone
  cortex: true
  tenant: "{tenant}"
  org: "{org}"
  forgejo:
    owner: "{owner}"
    host: "{ssh_url}"
 
practices:                     # Phase 6 — what was actually provisioned
  - name: "{ai_id}"
    provisioned: true
    domain: "{domain}"
cockpit_profile: "{name}"      # Phase 6 — `empirica cockpit launch --profile {name}`; omit if none was written
 
work_preferences:
  ai_autonomy_level: "{autonomous|collaborative_with_checkpoints|assistant_mode}"
  uncertainty_surfacing: "{always_explicit|when_material|minimal}"
  pushback_style: "{direct_and_factual|gentle_reframe|socratic}"
  task_splitting:
    ai_autonomous: ["{task}"]
    ai_with_checkpoint: ["{task}"]
    human_only: ["{task}"]
 
trust_building:
  current_level: "{establishing|building|established|high_trust}"
  autonomy_earned_through: ["{demonstration}"]
  non_negotiables: ["{boundary}"]
 
modules:
  active: []
  available: []

Omit blocks with nothing in them. An empty mesh: block is a claim that the mesh was considered and configured, which would be false.

Render it as an artifact

Do not paste 80 lines of YAML into the conversation and ask "does this look right?" — nobody proofreads YAML in a terminal.

Publish it with the Artifact tool as a readable page: the working agreement in prose, grouped by section, with the raw YAML at the end for anyone who wants it. The user gets a link they can open, keep and re-read later — and a review they will actually do produces corrections, which is the point.

Then ask for corrections before writing the file, with options:

· Looks right, save it · Change my autonomy settings · Change the

non-negotiables · Add a goal

After saving

1. Save to the location chosen above.
2. Report what was provisioned in Phase 6, per practice, honestly.
3. Log it — empirica finding-log with what the protocol covers (N goals,

N domains, N practices provisioned).

1. Say how to change it — re-run /ewm-interview, or edit the YAML

directly; it is theirs and it is plain text.

Interview discipline