Level Up

Find one manual drag on the week, scope it ruthlessly, ship exactly one artifact. The interview teaches the framework as a side effect — after a few runs Alex spots candidates mid-week without prompting.

When to use

• "level up", "what should I automate next", "find me leverage this week"
• Fridays, after /os-audit (audit gaps feed Phase 0)
• NOT for: /os-audit (structural score), stale-content hygiene passes, or the command channel's /audit (autonomy log)

Steps

Phase 0 — Pre-seed (data before questions)

Read, in one batch:

1. knowledge/tacit/bottlenecks.md — rows from the last 14 days, plus any topic appearing 3+ times across the full history (count them)
2. Last 7 days of state/daily-notes/
3. state/skill-feedback.jsonl if present (skill corrections logged by session-wrap)
4. your CLI's usage log (e.g. state/agent-usage.jsonl) if present — flag cost-heavy recurring jobs
5. Latest state/audits/*.md — the audit's top gaps are additional candidates

Present the top 3 observed bottlenecks with frequency counts: "Here's what the data says you keep doing manually — pick one or override." Alex can override with their own candidate, but the data goes on the table first.

Phase 1 — Find the candidate (mindset interview)

Ask conversationally, skipping any question the pre-seed already answered:

1. "What did you do 3+ times this week?" (frequency)
2. "Anything manual, boring, or copy-paste?" (drudgery)
3. "Anything a smart intern could handle?" (delegation)
4. "If the workload 10x'd tomorrow, what breaks first?" (constraint)

Output: 1-3 candidates ranked by leverage, one line of "why" each. Ask Alex to pick ONE. Even if they arrive with a pre-formed idea, run the pre-seed + at least one interview check against it.

Phase 2 — Scope it (method interview)

1. Constraint: which bottleneck does this solve? Tie back to Phase 0/1 evidence.
2. Eliminate first (Eliminate → Automate → Delegate): "What happens if we just stop doing this?" If "nothing breaks" → exit cheerfully, mark the bottleneck row Eliminated, log the decision, STOP. Don't automate waste. Then Automate (aim for ~60% deterministic / 30% AI / 10% manual framing). Delegate last (suggest a person, log it, stop).
3. Process map — five elements: trigger / data sources / transformations / decision points / destination. Hard stop: if Alex can't articulate any of the five — "If you can't explain the process to a person, you can't explain it to an AI. Sketch it first, come back."

4. Autonomy level → Laila tiers:

   Level	Maps to
   L0-L2 (manual / suggested / drafted)	Tier 3 — propose-and-wait
   L3 (supervised)	Tier 1 — auto-execute + notify
   L4 (autonomous)	Requires editing config/autonomy-rules.json — push back hard; L4 only after clean L2/L3 history and an explicit override

   Default = lowest level that solves the problem. Workflows beat agents.

5. KPI tie (mandatory): which Goal in the CRM/goal tracker does this move, and which metric? If no goal/metric can be named: "no KPI, no build" — stop, log Deferred-with-reason.

Phase 3 — Build it

Boring-is-Beautiful ladder, in order — default = the lowest rung that solves it; Alex must explicitly choose more:

1. Prompt-only — saved trigger phrase, zero infrastructure
2. Deterministic script in scripts/ (+ a scheduled job if it needs to recur)
3. AI-assisted skill in skills/ — use templates/skill-template.md, including Verification + Self-improvement sections
4. Subagent in agents/ — LAST resort; read an existing agent file for format first

Ship exactly ONE artifact. If it's a skill, frontmatter carries bike-method-phase: 1 — manual validation first, never scheduled as a background job at phase 1.

Closing (mandatory — the triage loop)

1. Update the chosen bottleneck's row in knowledge/tacit/bottlenecks.md: status Planned → Shipped / Eliminated / Deferred-with-reason.
2. Log the decision: dispatch the decision-logger agent with the scoped spec (constraint, eliminate/automate/delegate outcome, process map, autonomy level, KPI) and today's date → knowledge/decisions/YYYY-MM.md.
3. One-screen close in chat: what was scoped, what shipped, bike-method-phase-1 reminder.

Rules

• One interview = one artifact. No multi-candidate parallel scoping.
• Eliminate is a win, not a failure — exit cheerfully and log it.
• L4 requires an autonomy-rules.json edit and explicit override; never default to it.
• No KPI in the goal tracker, no build.
• Writes only: the new artifact, the bottlenecks.md row, the decision log entry. Everything else read-only.

Verification (MANDATORY)

Before declaring this skill done (and again after edits):

• Cold test: fresh session, "level up". It must surface 3 data-backed candidates with frequency counts from bottlenecks.md/daily notes — generic output ("you should build a brief") = fail.
• Idempotency test: run twice on the same candidate — second run must not duplicate the bottlenecks.md row update or the decision log entry.
• Eliminate-first test: feed an obviously eliminate-able candidate → skill exits with Eliminated status, no artifact.
• L4 pushback test: ask for an autonomous email-replier on first build → skill insists on L1/L2 + Tier 3 first.
• Boring test: candidate solvable with a deterministic script → ladder rung 2 recommended, not a skill or agent.

Self-improvement (MANDATORY)

• After each run: if Alex corrected the output, this skill is wrong, not Alex. Propose the edit to this file immediately, or log {"skill": "level-up", "correction": "...", "date": "YYYY-MM-DD"} to state/skill-feedback.jsonl for the next /level-up.
• If the same correction happens twice, the fix is MANDATORY before the next run.