| name | learn |
| description | Use when a task depends on tech you don't actually know — an unfamiliar library, a fast-moving framework, a tricky API, an unfamiliar codebase — or whenever "I think it works like…" is load-bearing. Just-in-time and task-scoped; distinct from levelup, which updates the durable knowledge base. |
MasterMind — Learn
The invokable form of ~/.mastermind/engineering/core/agent-loop.md → "Learn the stack before you build." Topic/task:
$ARGUMENTS. Goal: reach genuine, current understanding — enough to build to standard — fast, and
cheap on context.
1. Detect what's actually used
Read package.json/lockfile, configs, and representative source to see the exact stack (framework +
versions, styling, state, data, test runner) and the team's conventions. Installed versions matter —
APIs drift, so learn the version that's actually here.
2. Map the skill-tree
Use the relevant roadmap.sh role/topic map as the checklist of what matters and to spot your gaps.
Don't learn the whole tree — learn the branch the task touches.
3. Learn to current standards
Read the primary docs for the specific APIs the task needs (verify against the installed version)
and skim one battle-tested example (the field's curriculum.md lists them). Prefer primary sources;
never rely on stale memory (the field's learning-sources.md).
4. Ground it in this codebase
Grep for how the pattern is already used here and match it. Consistency beats novelty.
5. Grill your assumptions before you build
The costly bugs come from assumptions you never checked. Before committing code:
- List what you believe about the APIs you'll use (behavior, signatures, return values, limits,
versions, edge cases) — specifically ("
getFile streams any size" — true?).
- Answer from the source yourself first — verify each against the docs/types/actual source (for the
installed version). Confirm ✓, correct ✗, or mark unknown ❓. Never ask the user what the code/docs answer.
- For each remaining unknown, propose — don't just flag. State a best-guess resolution + confidence
("probably streams; ~70%") so a human can confirm or correct, not author it. Serve interdependent
unknowns one at a time. Anything still unconfirmed becomes an explicit risk or a tiny
spike — never a
silent guess. No load-bearing assumption ships unverified.
Output & economy
Return a tight working brief: the stack + versions, the few APIs/patterns the task needs, the gotchas
(and the grilled assumptions ledger: claim → verified/corrected/unknown → source · confidence),
and links to the primary sources — not a tutorial. Delegate wide reading to a subagent to protect
context. If you learned something durable and reusable, capture it via levelup.
