gymcoach: Skill for Claude Code

ideate

triage manufactures code-health work (tests, small bugs, deps); ideate manufactures product work when even triage comes up dry, so the loop keeps pushing the product toward the goal of being the most complete self-hosted AI training/fitness app - not just keeping it tidy. The output is issues, not code: crisp, single-PR feature ideas a later implement-issue run picks up unattended. Read CLAUDE.md and docs/loops/07-autonomy.md first.

This is the lightweight, recurring cousin of the one-off deep-research workflow. It is deliberately cheap: it runs often (whenever the loop would otherwise idle), so it must cost a tiny fraction of that workflow.

When to run

Only when the loop is genuinely starved of product work, in this order:

• no ready PR to ship, AND
• triage found no trusted code-health issue, AND
• there is no open, actionable, trusted issue to implement.

If there is already an open actionable issue, STOP - implement that first. Never pile new ideas on top of unstarted ones.

Hard budget (the whole point is bounded ideation)

• Never launch the heavy deep-research workflow (the ~100-agent one) or anything close.
• Default to brainstorming from context you already have (the repo, the research memory, the roadmap). A web check is optional and capped at ~6 searches, used only if it adds something the captured research does not already cover.
• Spawn at most ONE helper subagent if you need one; prefer doing it inline.
• File at most 3 issues per run.
• Anti-flood: do not run if there are already >= 3 open enhancement/idea issues - the backlog is fed; let implement-issue drain it first. One productive run per starved cycle; once you file issues, you are done until they drain.

Where ideas come from (grounded, not random)

1. The vision. Aim: the most complete self-hosted AI training/fitness app. The operator broadened the scope (2026-06-11) beyond pure hypertrophy: strength, conditioning/cardio, endurance, mobility, and general fitness are all in scope. Concretely, the CSV importers (lib/import/strong-csv.ts, lib/import/hevy-csv.ts) currently SKIP cardio rows and nothing in the app tracks duration/distance work - that axis is now fair game. Ask "what would a serious lifter, a coach, or a leading competitor expect that we do not have yet?"
2. The captured research (Memory: research-product-direction.md) and its wedge - AI that advises within the user's program, explains why, and respects data ownership. Mine it for not-yet-built items.
3. The current product. Skim app/ routes, lib/, components/ to find concrete gaps and the natural next slice of what already shipped (e.g. readiness now exists -> what uses it next).
4. The README roadmap - unchecked items.
5. Optional bounded web check (<= ~6 searches) for what users want / what is missing in AI training/fitness apps lately - only if it beats the memory. No heavy fan-out.