| name | plan-efe |
| description | Given a goal and a set of candidate next actions, rank them by Expected Free Energy (pragmatic + epistemic value) using past similar episodes, and pick the winner. Delegates scoring to the policy-selector subagent. |
Upholds Invariant 3 of predictive-mind: pragmatic and epistemic value
are scored separately. Use this skill when multiple candidate actions
are on the table and the "first plausible one" is not obviously the best.
Typical trigger
- You have a task card but have not committed to a next action yet.
- The dreamer returned multiple candidates with comparable confidence.
- A PostToolUse hook reported surprise, and you need to choose between
/probe, /reflect, or a new action.
Procedure
- Enumerate candidates (2-5 is the sweet spot). Include, at
minimum:
- The action you are currently tempted to take.
- One probe (a
/probe call that would disambiguate an open
hypothesis).
- One "ask the user" option if anything material is unclear.
- Delegate to the policy-selector subagent with:
- The task card (goal + acceptance criteria + open hypotheses).
- The
task_id (so the subagent can call retrieve_beliefs).
- The candidate list.
- Report back the full ranked JSON output and the pick. Do NOT
rewrite or summarize the scores — the raw scores are the
accountability trail.
- If the pick is
ask-user, stop and ask. If the pick is
probe-first, invoke /probe. Otherwise proceed with
/predict then the chosen action.
What separates this from "just think about it"
- Similar-episode grounding: the policy-selector queries
recall_similar_episodes for each candidate, so prior surprises
actually discount similar-looking plans.
- Epistemic budget is explicit: a high-information-gain probe can
legitimately beat a low-confidence direct attempt, even though the
probe does not advance the goal.
- Scores are persisted: the ranked output becomes part of the
session record for later review.
Arguments
$ARGUMENTS — the current goal (free text). The skill will ask you
for the candidate list if you did not include it.
Output shape (from the policy-selector)
ranked: [ { action, pragmatic, epistemic, total, recommendation, ... } ]
pick: { action, rationale }
notes_on_calibration: <honest caveats>
Use the pick's recommendation field to decide the next skill:
execute → /predict then run the action
probe-first → /probe
dream-first → invoke the dreamer on the picked action before acting
skip → drop this candidate, re-plan