| name | level-up |
| description | Use weekly to find and ship one new automation. Walks the 3Ms interview — Mindset (find the candidate) → Method (scope one) → Machine (build it). Trigger on "let's level up", "what should I automate next", "find me leverage this week", or as a Friday ritual. One run = one shipped artifact, or a logged decision to stop. |
The Three Ms of AI is a framework by Nate Herk. See references/3ms-framework.md.
What this skill does
Walks the user through the 3Ms each week to surface and ship one new automation. One interview = one outcome: one shipped artifact, or one logged decision to eliminate, delegate, or stop. It also installs the 3Ms framework into the user's head over time — after 4-6 runs, the user starts spotting opportunities mid-week without prompting because the questions have become internal defaults.
This is the brain-rewire mechanism. The kit doesn't need cron jobs to anchor behavior; it needs /level-up running every Friday.
What /level-up is NOT
- Not
/aios-audit. /aios-audit is structural ("is the AIOS built right?"). /level-up is functional ("what leverage am I missing?"). Run /aios-audit first if structure is messy.
- Not a multi-candidate planner. One run scopes one candidate, ending in one artifact or one logged early-exit decision.
- Not a coach. The user does the thinking. The skill conducts the interview.
When /level-up runs
- First run: after setup is real. Once the user has connected ≥1 MCP/script and run
/aios-audit once. Earlier yields trivial output.
- Cadence: weekly, Friday afternoon. Review the week, surface one automation, ship Monday.
- On-demand any time. Mid-week if a manual task itches.
Inputs the skill reads
context/priorities.md — what the user said matters
context/about-me.md — top_pain, role
connections.md — what's reachable, by what mechanism
references/3ms-framework.md — the framework (used to quote principles back)
decisions/log.md — recent decisions (what's already shipped or considered)
.claude/skills/*/SKILL.md frontmatter — what capabilities exist
- Recent
audits/audit-{date}.md if present
Execution — three phases
Phase 1 — Mindset interview (find the candidate)
Fresh clone, no history? If decisions/log.md is empty and there's no audits/ yet (a brand-new kit), don't stall. Skip the "what did you do in the repo this week" mining and run the interview straight from what the user tells you, plus top_pain in context/about-me.md and context/priorities.md. Suggest running /aios-audit once so future runs have signal, but you can still ship one small automation today.
Surface 1-3 candidates ranked by leverage. Ask these in order, conversationally:
- "Walk me through your week. What did you do 3+ times?" (frequency)
- "Anything that felt manual, boring, or copy-paste?" (drudgery)
- "Anything where you thought 'a smart intern could handle this'?" (delegation)
- "If your workload doubled tomorrow, what would break first?" (constraint)
- "What would let you take on twice the work without breaking?" (growth lever)
Quote relevant Mindset principles when they fit:
- "Sounds like the Default Shift applies — to what extent could AI be leveraged here?"
- "This is the Function Breakdown — you're not automating the whole job, just this one piece."
- "AI is better than you think and improving faster than you think. If it couldn't do this last quarter, it might be ready now."
Output of Phase 1: numbered list of 1-3 candidate opportunities, one-line "why this is leverage" per candidate. Ask: "Pick one to scope."
Phase 2 — Method interview (scope one)
User picks one candidate. Walk the Method pipeline:
Step 1 — Find the constraint. Which bottleneck does this solve, or which growth lever does it open? Tie back to Phase 1 answers.
Step 2 — EAD: Eliminate / Automate / Delegate.
- Eliminate first: "What happens if we just stop doing this?" If the answer is "nothing breaks" → skill exits cheerfully. "Don't automate waste." This is a win, log to
decisions/log.md and stop.
- Automate second: apply 60/30/10 framing. ~60% deterministic, ~30% AI-assisted, ~10% manual.
- Delegate third: if too complex/variable/judgment-heavy → suggest a person. Skill exits with a delegation suggestion, log it.
Step 3 — Map the process. Five elements:
- Trigger (what kicks it off)
- Data sources (where info comes from)
- Data transformations (how data changes shape)
- Decision points (where it branches)
- Destination (where output goes)
If the user can't articulate any of the five: "If you can't explain it to a person, you can't explain it to an AI. Sketch it on paper first, then come back." Skill stops.
Step 4 — Pick the autonomy level.
| Level | Name | What happens |
|---|
| L0 | Manual | No AI |
| L1 | Suggested | AI suggests, human decides every step |
| L2 | Drafted | AI drafts, human reviews and edits |
| L3 | Supervised | AI runs, human validates periodically |
| L4 | Autonomous | AI handles end-to-end |
Default = lowest level that solves the problem. Push back on L4 unless the user has explicitly run lower levels first. "Workflows beat agents. If a decision doesn't HAVE to be made by AI, don't let AI make it."
Side-effect gate (overrides the autonomy level). Before locking the level, ask: does any step send, delete, publish, pay, or deploy? If yes, that step is a mandatory human-gate regardless of the autonomy level chosen — the AI may prepare the action, a human approves the irreversible part. This carries into the scaffolded artifact. (Loop-engineering doctrine: references/agent-loops.md.)
Step 5 — Tie to a KPI. Which of the Three Buckets does this move?
- More output / more customers
- More value per unit of work
- Less cost
Plus a specific metric (response time, error rate, conversion rate, time-to-completion). If the user can't name a bucket and a metric, skill stops. "If your automation doesn't move a number, why are you building it?"
Step 6 — Define "done = right." How will the user know the output is correct? Pick the cheapest check that fits: a boolean test (pass/fail), a quick look, a rubric, or your eyes. Default to a functional test or a manual check. For content-quality judgments, route the check to a different model lineage — don't let the model that produced the work grade its own work (references/agent-loops.md). This line carries into the scaffolded artifact.
Output of Phase 2: scoped automation spec written to decisions/log.md as a dated entry with the Method answers + autonomy level + KPI + the "done = right" check. Durable record of what was decided and why.
Phase 3 — Machine handoff (build it)
Ask: "How do you want to ship this?" Options ordered by Boring-is-Beautiful default:
- Prompt-only — saved prompt template the user runs by hand. Zero infrastructure. Highest manual involvement.
- Deterministic skill — SKILL.md that runs a script (no AI step). Best for transformations with clear rules.
- AI-assisted skill — SKILL.md with one AI call inside. Drafts, classifies, summarizes.
- Sub-agent — multi-step agent. Last resort. Only if the work genuinely needs reasoning + tool use.
Default selected = highest non-AI option that solves the problem. User has to explicitly choose more autonomy.
Once chosen, route to the appropriate scaffolder:
skill-creator if available globally (Anthropic-shipped)
skill-builder if the user has it locally (ships with this kit)
- For a sub-agent, route to
agent-builder
- Otherwise write a SKILL.md / agent file inline with frontmatter, location, and contents
Every scaffolded artifact keeps standard frontmatter (name + description, only the fields skill-builder allows — don't add fields just because you can) and carries a short note block in the body, right under the frontmatter:
> **Bike Method — Phase 1 (training wheels).** Run this by hand first; advance only after you've validated it manually.
> **Side-effect gate:** any step that sends / deletes / publishes / pays / deploys stays human-gated, whatever the autonomy level.
> **Done = right when:** <boolean test / a look / a rubric / your eyes>. Route content-quality checks to a second model (different lineage).
> *The Three Ms of AI is a framework by Nate Herk.*
This is a convention, not enforcement — nothing in the kit reads a bike-method-phase field, so the reminder lives in the body where the user actually sees it, instead of an invented frontmatter key that breaks the skill-builder rule. It nudges the user to validate manually first and to keep irreversible steps gated; it doesn't mechanically block skipping. The loop-engineering doctrine behind training-wheels phases, side-effect gates, and verification lives in references/agent-loops.md.
Surface the Machine principles when scaffolding:
- Lego Principle — smallest steps, zero-AI first if possible
- Validation Chain — test each step before chaining
- Iteration Mindset — ship the POC, expand from real usage
Output contract
Every /level-up run produces:
- One
decisions/log.md entry — dated, with the Method spec
- Either one scaffolded artifact (prompt, skill, or agent file) or one logged early-exit decision — eliminate, delegate, or stop. Early exits are wins, not failures, and still get logged.
- A one-screen close — what was scoped, and either what was built (with the Bike Method Phase 1 reminder) or why the run stopped early.
Critical implementation rules
- One interview = one outcome (one artifact, or one logged early-exit decision). No multi-candidate parallel scoping.
- Mindset phase always runs first. Even if the user comes in with a pre-formed idea.
- EAD enforces "eliminate first." If the answer is Eliminate, exit cheerfully — that's a win, not a failure.
- Default to the lowest autonomy level that works. Push back on L4.
- Boring-is-Beautiful default in Machine handoff. Default = highest non-AI option.
- Tie-to-KPI is mandatory. If the user can't name bucket + metric, skill stops.
- Bike Method ships into every artifact as a body note (a convention, not a frontmatter field — nothing reads it).
- Read-only on user files except
decisions/log.md and the new artifact. Don't modify other existing files.
- Attribution on output. Every report and every scaffolded artifact references the framework.
Verification (for the implementer)
- Dry run on a populated profile with no prompt. Expected: skill surfaces 2-3 candidates pulled from recent activity, priorities, and top_pain. Generic output ("you should build a brief") = fail.
- Fresh-clone test. Run on a brand-new kit with an empty
decisions/log.md and no audits/. Expected: skill doesn't stall or error on missing history — it interviews from top_pain + priorities + what the user says, and suggests /aios-audit for next time.
- Eliminate-first test. Feed an obviously eliminate-able candidate. Expected: skill suggests Eliminate, exits, logs the win.
- L4 push-back test. User asks for an autonomous email-replier on first build. Expected: skill insists on L1/L2 first, won't ship L4 without explicit override.
- Boring-is-Beautiful test. Candidate solvable with deterministic Python. Expected: skill recommends
(2) deterministic skill as default.
- Bike Method anti-skip. User scaffolds, asks to advance to Phase 4 immediately. Expected: skill conversationally walks them through what each phase means and asks them to confirm they've validated the lower phases (a nudge — the body note is a convention, not a hard lock).
The Three Ms of AI is a framework by Nate Herk. See references/3ms-framework.md.