dissect

Dissect

Leading words: dissect, ground truth, entity-level, minimal-build, red-team, evidence-backed, stub detection, naming audit, fold test, phase separation, patch in place.

Systematically interrogate and optimize an existing service, written plan, or system. The goal is not to rubber-stamp the current design — it is to arrive at the minimal-build plan that correctly meets the requirements, by challenging every assumption with live evidence and an adversarial review panel.

Trigger: /dissect <target> where <target> is a service name, plan file path, codebase area, or free-form description. If omitted, ask.

dissect is the backward-looking counterpart to deep-deliberation. They share DNA — parallel read-only sub-agents, devil's-advocate red-teaming, human checkpoints, evidence over intent — but point in opposite directions: deliberation designs what to build; dissect audits what already exists.

---

User Input

$ARGUMENTS

---

Core principles (never waive)

1. Ground truth beats intent. What the code actually does overrides what the plan says it should do. Always verify live.
2. Row counts are a liveness signal only. A 0-row table can be load-bearing; a populated table can be redundant. Never use row count as a keep/drop reason.
3. Name what it IS, not what it was named. Entities are frequently misnamed. Read a few actual records before drawing conclusions. A table called contact_preference might store food allergies and golf equipment.
4. Own vs read. "Entities this service reads" ≠ "entities this service owns." Reading a table does not mean you should count it in the domain.
5. Stubs are not live. A function that throws new Error('not implemented'), a cron that no-ops, or a consumer that re-queues to the DLQ is not a running feature. Check before counting it as active behavior.
6. Minimal build. Every proposed change must trace to a requirement. Do not design for hypothetical future use. Three similar things beat a premature abstraction.
7. Naming audit is phase zero. Naming confusion causes the most incorrect verdicts. Run it before any verdict.
8. Sub-agents return evidence, the orchestrator returns verdicts. A sub-agent that declares "VERDICT: DROP" unprompted is overstepping. Synthesis happens in the main context where all evidence is held together.
9. Escalate discrepancies, don't resolve them silently. If code says X and the plan says Y, surface both at the checkpoint and let the human decide.
10. Patch in place — never rewrite the live plan. If the session is in Cursor Plan Mode or Claude Plan Mode, or the target is an existing plan document, the source plan is the document of record. Dissect it, then surgically update it (fix, tune, fill gaps). Never full-file Write it. Never create a second plan that replaces it. Unaddressed sections stay verbatim. Read references/plan_preservation.md before any write. Override only if the human explicitly asks to rewrite from scratch.