find-opps

Find Opportunities

Identify new directions of work the project should explore — things that aren't already tracked, aren't obvious from the existing backlog, and that the project owner may not have thought to pursue yet. The output is backlog-ready items — this skill does NOT execute them.

This is the outward-looking complement to /find-tasks, which finds work within existing paths. This skill looks beyond current plans to find new territory.

When to Use

• When /find-tasks lenses 1-6 yield fewer than 2 new items (existing paths are well-mined)
• When the user asks "what else should we be thinking about?"
• Between major milestones when strategy should be reassessed
• When the project feels stuck or stale

Context Required

Before running this skill, you need the project overview already loaded (charter.md, MEMORY.md, state files, backlog, recent plans, references). If you don't have it, gather it now:

1. charter.md — the goals, who the project is for, the audiences its goals target, and its non-goals (a repo set up before it existed keeps its goals in MEMORY.md); and MEMORY.md — project overview, constraints, and knowledge map
2. state/INDEX.md → then every file listed in it
3. backlog/INDEX.md → then every file and subfolder listed in it
4. Active plans — glob plans/*.md, skip INDEX.md, and read the rest (plans/INDEX.md deliberately lists none of them), except an autopilot cycle record (*autopilot-cycle*), which is autopilot's record, not a plan
5. plans/completed/INDEX.md — scan the 5 most recent completed plans
6. references/INDEX.md — scan for existing research to avoid duplicating
7. references/discards.md, if it exists — ideas already turned down, each with its reason and what would reopen it

Isolate the heavy reading when you can. In an interactive session on a frontier model, push the context-load above and the lens-3/lens-4 web research to a read-only reader subagent that returns a distilled overview + web digest with citations — then do the ideation (lenses, Filter, Constraints) yourself on that brief. Reading state files and full web pages doesn't need frontier judgment; the ideation does. Enumerate what the reader should extract, or it drops the signal. Skip this on a small project where there's nothing to isolate. The ops:opps-finder agent cannot spawn subagents; it does the reading itself, already apart from its caller's context.