Skill: act

Description

Reads the last session journal entries, open todos, and active plans to infer current intent — then proposes and executes the single most useful next action without being asked.

Trigger phrases:

• "act"
• "what should we do next"
• "pick up where we left off"
• "what were we doing"
• "continue"
• "resume"
• (also fires automatically at Start Session if there are open High-priority todos)

Steps

Step 1 — Read context (parallel)

Read all three sources at once:

1. [MEM]/session_journal.md — last 40 lines (recent session summaries)
2. [MEM]/todo.md — full file (High/Medium priorities only)
3. [MEM]/plans/ — list open (non-archived) plan files, read the most recently modified one

Read([MEM]/session_journal.md, offset=last 40 lines)
Read([MEM]/todo.md)
Glob([MEM]/plans/*.md, excluding plans/archive/)

Step 2 — Infer intent

From the context read, identify:

• What was being built/fixed — look for the last task mentioned in session_journal.md
• What's blocked or open — todos marked [ ] with High priority
• What plan is active — any plan file with Status: In Progress or Status: Draft

Rank candidates:

1. An in-progress plan with an identified next step
2. A High-priority open todo
3. The last thing mentioned in session_journal.md

Step 3 — Propose one specific action

State exactly what you propose to do. Be specific — not "continue the feature" but:

"Last session we finished the Java method for adminSendIntroEmail but hadn't wired the HFmtWebservice endpoint yet. I'll add the routing line and test it now."

Or:

"Your todo.md has one High-priority open item: 'Fix session tab scroll position reset on mobile'. I'll investigate the CSS/JS cause now."

Format:

## Picking up

**Last context:** [1 sentence from session_journal]
**Open item:** [specific todo or plan step]

**Proposed action:** [exactly what I'll do in the next 2 minutes]

Go? (yes / skip / show me the full context)

Step 4 — Execute on confirmation

If user says yes / go / do it:

• Execute the proposed action immediately
• No re-asking, no re-summarizing
• After completion, update [MEM]/todo.md (check off or add follow-up) and [MEM]/session_journal.md

If user says skip:

• Show the next candidate from the ranked list
• Repeat step 3

If user says "show me the full context":

• Print the relevant section from session_journal.md + the matched todo/plan

Notes

• Never propose more than one action at a time — focus wins over completeness
• If context is genuinely ambiguous (no journal entries, empty todo, silent STATUS.md), say so and ask what to work on — act cannot invent intent from nothing; it surfaces what's already been decided
• This skill is intentionally opinionated — it picks one thing and drives it
• Pairs with search-memory when the proposed action needs prior-art research first
• "act" on Clankbrain kit tasks = look at D:\claude-memory-starter changes
• "act" on webapps tasks = look at the TicTacWisdom/InfoGrasp codebase