spike

Answer: is this idea worth a proper project? Produce a reproducible exploratory artifact and directional verdict, not production code.

$ARGUMENTS

Before anything else, run the tracker preflight (${CLAUDE_PLUGIN_ROOT}/skills/shared/references/preflight.md); a failure stops here with its single ## Action needed block, before any baseline work starts.

Create/load the spike Task via the tracker skill: discover candidate hackday Epics in the configured project (owned by the tracker skill), let the user choose the parent via AskUserQuestion (${CLAUDE_PLUGIN_ROOT}/skills/shared/references/user-interaction.md), create the Task, set it in-progress, and self-assign. Keep tracker command mechanics in the tracker skill (/sy:tracker), not here.

Routing

• small cohesive evidence ⇒ read directly;
• broad archaeology ⇒ sy:sweep;
• one load-bearing production/data path ⇒ sy:trace, at most the resolved limits.max_depth_agents cap in flight — resolve per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/config-values.md.

Returned briefs are leads. Verify decisive spans and own the experimental conclusions. Seed every agent prompt with known anchors — paths, symbols, data locations — and name ground already covered.

1. Establish the baseline

Locate the real production behaviour and enough representative data to reproduce it offline. Treat production as a baseline/reference implementation, not an oracle: the experiment may exist because production is wrong.

Validate baseline fidelity against known examples before comparing alternatives. Record any irreducible mismatch.

2. Validate reproduction

A failed reproduction is not yet a negative result. First verify:

• premise;
• data selection;
• baseline fidelity;
• ordering/configuration assumptions;
• sampling bias.

If the problem still does not reproduce, report that plainly. Do not manufacture a positive result.

3. Design the experiment

Define:

• falsifiable success bar;
• comparison baseline;
• primary metric(s);
• regression metric(s) in the opposite direction;
• representative and adversarial slices;
• controls for leakage/confounding where relevant.

Surface to the user via AskUserQuestion only when a load-bearing choice changes the experiment or verdict interpretation.

4. Measure gain and regression

Measure both directions: improvement and what gets worse. Use representative shapes and report uncertainty/sampling limits. Prefer direct metrics over proxy storytelling.

Every material empirical or spatial claim must be directly inspectable through a figure, table, or compact numeric evidence. Do not force a chart where a table or scalar is clearer.

5. Stress-test sampling

Probe boundary cases, failure clusters, and slices where the proposed approach should plausibly lose. Look specifically for evidence against the preferred hypothesis.

6. Produce the artifact

Create one standalone, reproducible artifact folder with:

• README/run instructions;
• data anchors or retrieval instructions;
• baseline reproduction;
• alternative implementation;
• metrics in both directions;
• figures/tables/numeric evidence;
• limitations and unresolved questions;
• exact environment/commands needed to rerun.

Repository-specific plotting/style defaults belong to repository standards, not this generic skill.

7. Record the verdict

Before posting it, pressure-test the directional conclusion with sy:debate — unconditionally, not only when the result looks contested: even a clear-looking gain/regression measurement can hide a confound, a sampling bias, or a stronger alternative explanation that only an adversary surfaces. See ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/debate.md.

Post a tracker verdict comment. Its human part is the call being made:

• TL;DR verdict;
• strongest failure cases;
• recommendation: stop, iterate another spike, or start /sy:plan.

Its agent_detail part is what a later session reruns or cites from:

• headline gain and regression numbers;
• baseline-fidelity confidence;
• artifact path/link.

If the idea clears the bar, write a /sy:plan brief that states the opportunity, evidence, success bar, constraints, failure modes, and the next unanswered project-level questions. Link it from the verdict.

Set the spike Task's status to done via the tracker skill only after the verdict is durable. End with the Task key/link, artifact path, verdict, and /sy:plan <brief> kickoff when applicable.

Every subagent dispatch resolves its model and effort from config and passes them as the Agent invocation's actual model and effort parameters, per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/model-dispatch.md; a nested dispatch inherits nothing and must resolve again.