Surprise Me - goal-directed creative leverage
The first workable approach tends to get defended instead of questioned. This skill finds high-leverage moves the current plan missed, in any kind of work: a research question, a hypothesis, an experiment design, an engineering design, a plan. Be ambitious: the smallest thing that satisfies the request is the default to beat. The output is a short ranked table of ideas plus one already-executed proof, never a brainstorm essay.
Process
- Fix the goal, the judge and the default. Restate, in one sentence, the external decision rule that defines success (the reviewer, the metric, the deadline, the acceptance criterion). Every idea is scored against that sentence, nothing else. If a plan already exists, state it in two lines and name its main mechanism: that is the default the ideas must differ from.
- Inventory what already exists before inventing anything. Grep the repo, the data files, and the sources for dormant capabilities: disabled self-checks, unused flags, data already collected but never aggregated, tools pointed at the wrong target. The inventory includes the agent's own unused capabilities the owner has not asked for yet (scheduled monitors, browser verification of rendered pages, multi-agent document QA, published dashboards, cross-model review panels) - list the ones that serve the goal, each as a concrete offer with a cost. When the work belongs to another session's plan, target that plan for the step-7 filing rather than leaving the ideas only in chat. The best surprises are recombinations of assets already on disk.
- Generate only ideas the judge can verify. An idea that requires the reader to trust us is worth less than one they can re-run. Prefer: quoting reproducible evidence verbatim, running a control experiment, machine-deriving a statistic, reframing existing facts. Reject: adjectives, promises without dates, anything that tunes an instrument toward the desired outcome. At least one idea must not use the default's main mechanism. For a research question or claim, write several competing hypotheses, including the dull ones (a bug, noise, an effect already known), each with the observation that would rule it out, and search for prior work by name before calling anything a gap. Run the analogy move once: (a) restate the goal's purpose in one sentence with no domain nouns; (b) list who else serves that purpose with a different mechanism - three sources from this repo or its dependencies, two from an adjacent field, at most one far. Near first: near sources yield more usable ideas than far ones; (c) for each source, one line mapping source relation to target relation, and one line naming where the analogy breaks. No named break, no idea. Model the mechanism, never the nouns.
- Check each idea against any policy or freeze line. Mark it "do now" (tooling/framing side) or "commit for later" (frozen-artifact side, e.g. after a freeze lifts). An idea on the wrong side of the line becomes a named commitment, not an action.
- Rank by goal-value per unit cost and present as a table:
# | idea | when | owner | cost | deliverable. Present the survivors; cut anything that does not beat the ideas above it. Say what the top idea improves over the default (quality, time, risk) and what it costs; if the default is still best, say so in one line. - Execute the cheapest read-only idea in the same turn (a grep sweep, a framing sentence, a control run in a scratch dir) and report its first result alongside the table. Proof beats proposal. Choose a check that can fail, and write down the result you expect before running it.
- File the survivors. Propose appending the table to the project's active plan document so the ideas outlive the session, each with when/owner/cost/deliverable filled in; the append itself waits for the owner's go (rule 3).
Scale modes
- Solo (default): the 7-step process above, one context. Right for a daily "what am I missing".
- Panel (wide or high-stakes decisions): fan out fresh-context agents with deliberately different lenses, one per lens the decision needs - a dormant-assets hunter (step 2 only), a reframing hunter (same facts, stronger story), an adversary (attacks the current plan and proposes what its holes imply), an analogist (the step-3 analogy move only), and a fresh start (gets the goal, the judge, the constraints and the sources but not the current plan, and designs its own approach) - then merge, dedupe, and rank their ideas yourself. Different ideas come from different briefs, so each agent commits to its ideas instead of listing options.
Before spending on a winner
An idea costing more than ~1 hour must, before execution: (a) carry a measurable success criterion
written down in advance ("byte-identical output or it changed nothing"), and (b) survive an
adversarial pass - write the idea as a short plan file and run this plugin's plan-reviewer
agent on it blind, rather than re-implementing interrogation here.
Rules
- Never manufacture an idea that moves a number by choosing what to measure - ideas must survive an adversarial reader. This is the canonical anti-tuning rule; step 3 and the pre-spend gate point here.
- For measurement or audit projects, phrase the goal as validity or credibility ("a defensible result"), never as a metric's direction ("a lower error rate") - the second corrupts step 3 silently.
- Experiments that consume real resources and any mutation of a shared plan document need the owner's go; only the one cheap read-only proof in step 6 runs unprompted.
- Every claimed fact in an idea's pitch traces to a command run this session.
- Respect the repo's hard invariants (immutable results, protected sources, signed artifacts) - creative means recombining, not bending.
- If the goal is ambiguous, ask one question to pin the judge before generating anything.
Use $ARGUMENTS as the goal; if empty, default to the project's nearest deadline or active
objective.
