psychology-agent: Skill for Claude Code

Iterate — Autonomous Work Discovery + Execution

Find the single most important thing to do next, then do it.

/sync      = check for inbound activity
/hunt      = discover candidates, present to user
/iterate   = sync → hunt → discriminate → execute → cycle

The user types /iterate and the next most important thing gets worked on.

---

Arguments

Parse $ARGUMENTS to determine scope:

Argument	Behavior
(empty)	Full hunt → discriminate → execute
quick	Quick-wins hunt → discriminate → execute (XS/S items only)
deep	Deep/extrapolate hunt → discriminate → execute
[domain-filter]	Hunt filtered to that domain → discriminate → execute

---

Protocol

Phase 0: Sync (compressed)

Run a quick sync internally — not user-facing output, just input to the hunt.

1. git fetch origin — new remote commits?
2. git log HEAD..origin/main --oneline — anything landed?
3. ls -t transport/sessions/*/from-*.json | head -5 — new inbound messages?
4. gh pr list --repo safety-quotient-lab/psychology-agent --json number,title — open PRs?

If sync finds actionable items (new commits, unread transport messages, open PRs), they become candidates in Phase 1 alongside hunt results.

If clean, proceed silently to Phase 1.

Phase 1: Hunt (compressed)

Run /hunt internally — not as a user-facing presentation, but as input to the discriminator. Use the /hunt protocol (Sources 1-8) but compress output:

• Read TODO.md, MEMORY.md Active Thread, lab-notebook.md Current State
• Scan for candidates across all hunt sources
• Cap at 5 candidates. If hunt surfaces more, pre-filter to the 5 with highest value/effort ratio before proceeding to Phase 2.

Output of Phase 1 is an internal candidate list, not user-facing. Format:

Candidate 1: [action] — Value: H/M/L, Effort: XS/S/M/L, Source: [where found]
Candidate 2: ...
...

Exit condition: If hunt finds zero candidates, report "Project is clean. No actionable work found." and stop.