moralespanitz/plan

<SUBAGENT-STOP> If you were dispatched as a subagent, skip this skill. Execute your task and return results. </SUBAGENT-STOP> <HARD-GATE> Do NOT produce a vague timeline or milestone list. Every task must be specific enough that someone with no context could execute it. If a task takes more than 1 day, break it down further. </HARD-GATE>

Plan Skill — Concrete Execution Plan

The difference between a plan that works and one that doesn't is specificity. "Set up the codebase" is not a task. "Fork IDSIA/recurrent-fwp, rename to dopamine-read-gate, run python train.py to confirm baseline, record output" is a task.

Step 1 — Read the session

Read the lab notebook to understand exactly where the researcher is:

cat .research-loop/sessions/<slug>/lab_notebook.md

Summarize back to the researcher:

"You've selected [hypothesis/route]. You have [what's done]. The next phase is [what's needed]."

Ask one clarifying question if anything is ambiguous. Then proceed.

Step 2 — Write the plan

Break the work into tasks. Each task must have:

• A specific action (verb + object)
• The exact file, repo, command, or person involved
• How to verify it's done (what does success look like?)
• Time estimate (be honest — hours, not days)

Format:

## Execution Plan — [hypothesis name]
Date: <date>

### Phase: [name]

**Task 1: [action]**
- Do: [exact steps]
- Verify: [how you know it worked]
- Time: [estimate]

**Task 2: [action]**
- Do: [exact steps]
- Verify: [how you know it worked]
- Time: [estimate]

...

Step 3 — Create todos

Use TodoWrite to create one todo per task. Mark the first one as in_progress.

Show the researcher the plan and ask:

"Does this cover everything? Any task that's unclear or needs breaking down further?"

Wait for approval. Then say:

"Starting Task 1. Tell me when you're ready."

Step 4 — Track progress

As each task completes:

1. Mark it complete in TodoWrite
2. Append a one-line note to lab_notebook.md: [date]: Task [N] complete — [what was done]
3. Surface the next task explicitly: > "Task [N] done. Next: [Task N+1 description]. Ready?"

What makes a bad plan

• Milestones instead of tasks ("Week 1: build the model")
• Tasks without verification steps ("set up environment")
• Tasks that depend on unknowns without surfacing them ("get lab data" — do you have the contact? have you emailed?)
• No time estimates

If you catch yourself writing any of these, stop and break it down further.