Plan deep dive

/deep-dive

Phase-by-phase deep inspection of any multi-phase plan. Stop at each phase, assess completeness, quality, and future-readiness before moving on.

Structured execution spec: protocol.yon. Read it for the canonical rules and step sequence; this file is explanation. The two must stay in sync — if you edit one, update the other and refresh the @STAMP date.

Caller Options. Before executing, run the Caller Options protocol (caller-options/SKILL.md): triage this invocation for material optionality across the venues declared in front-matter; if one path clearly dominates, proceed silently; otherwise surface the options. For very long plans, delegated keeps the main context clean while a sub-agent walks the phases and returns the verdict log.

Iterative Depth Protocol. /plan-create produces the plan. /plan-deep-dive pressure-tests it phase by phase. Never batch or skip phases — the value is in the forced stops.

When to escalate to plan-create

If phase-by-phase inspection surfaces gaps that would require redesigning the plan rather than tweaking it (missing phase, wrong sequencing, fundamental approach issue), stop the deep-dive and recommend /plan-create to redesign.

Phase 1 — Load & Enumerate

Identify the active plan. Load implementation_plan.md, PLAN.md, or the current multi-phase artifact (e.g. a multi-phase plan presented inline in the conversation, or a plan document at any path the user has referenced). Extract the ordered list of phases.

Verification step: before entering Phase 2, present the loaded artifact's title and first substantive line to the user and ask: "Is this the correct and current plan?" Do not proceed until confirmed.

Gate: no phases found → cannot proceed.

Phase 2 — Iterative Phase Drill-Down

For each phase, execute the full drill-down loop below. Do NOT advance until the current phase passes all checks.

Running verdict log: maintain a verdict table updated after each Step 5. This log is cited verbatim in Phase 3 — do not reconstruct from memory.

| Phase | Verdict | Key Findings | Amendments Applied |
|-------|---------|-------------|-------------------|

Step 1 — Read the Phase

Read the current phase in full. Summarize scope, deliverables, and dependencies.

Step 2 — Gap Assessment: Are we MISSING anything?

Check for: omitted edge cases, unstated assumptions, undeclared dependencies, missing error handling, missing validation, incomplete coverage.

Minimum bar: at least one named gap, or explicitly state "No gaps found — [reason]." A blank or "looks good" is not acceptable.

Step 3 — Quality Assessment: Can we do it BETTER?

Check for: simpler approaches, performance gains, better abstractions, reduced scope creep, alignment with existing patterns and skills, DRY violations, naming drift.

Minimum bar: at least one named quality issue, or explicitly state "No quality issues found — [reason]."

Step 4 — Future Assessment: Is this FUTURE-READY?

Check for: extensibility, breaking change risk, forward compatibility, scalability bottlenecks, tech debt introduced, migration paths needed later.

Consider 3-month and 6-month horizons explicitly. Example prompts: "What breaks if this phase's output format changes in 3 months?" / "What would need to be re-done if the team doubles in 6 months?"

Minimum bar: at least one named future risk, or explicitly state "No future risks found — [reason]."

Step 5 — Phase Verdict — Present

Synthesize findings into a verdict. Present the verdict and all findings to the user. Do not apply amendments yet.

• PASS — no significant gaps, quality issues, or future risks found
• AMEND — gaps or improvements found; plan can proceed after targeted fixes
• RETHINK — stop and redesign before proceeding. Triggered when any of:
  • The phase's primary deliverable is undefined or cannot be stated concretely
  • The phase has an undeclared dependency on work not yet planned or completed
  • The phase's approach directly contradicts a constraint or goal stated in the plan

Update the running verdict log with this phase's result.

Step 6 — User Gate + Apply Amendments

Present the phase verdict and findings summary. The user must acknowledge at least one specific finding by name before the gate passes — a blanket "yes" or "proceed" is not sufficient. If the verdict is PASS with no findings, the user must confirm "no findings noted."

After user approval: apply any amendments to the plan artifact. Then loop back to Phase 2, Step 1 for the next phase.

Phase 3 — Synthesis

After all phases pass, produce a consolidated deep-dive report. Cite per-phase verdicts directly from the running verdict log — do not reconstruct from memory.

• Per-phase verdicts — copied from the running verdict log
• Total amendments made — count and list
• Residual risks — findings not resolved by amendments
• Future considerations — 3-month/6-month items surfaced during Step 4 passes
• Overall confidence score — 1–10, where 10 = all phases PASS with zero findings, 1 = multiple RETHINK with unresolved issues. State the score and the primary reason it is not higher.

Verdict examples:

• AMEND example: "Phase 3 AMEND — missing rollback step identified; added explicit rollback instruction to the phase before advancing."
• RETHINK example: "Phase 2 RETHINK — primary deliverable (the API schema) was undefined; could not assess completeness. Returned to /plan-create to define it before re-running."

Rules

• MUST stop at each phase — never batch or skip
• MUST propose concrete fixes when finding gaps — not vague observations
• MUST consider 3-month and 6-month horizons when assessing future
• MUST apply amendments to the plan artifact only after user approval in Step 6 — never before
• MUST NOT output "LGTM" without substantive analysis
• SHOULD check sibling skills for established patterns when referencing existing work

Downstream

Suggest the natural next command — these mirror the next-skills: front-matter:

• /plan-execute — if all phases PASS
• /plan-create — if RETHINK on any phase

Human output. This skill's handler-facing output obeys the human-output contract (human-output/SKILL.md).

Next skills. On completion, run the Next Skills protocol (next-skills/SKILL.md): surface the next-skills recommendations from front-matter for the caller to pick. Offer only — never auto-invoke.

Self-improvement. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): if this run surfaced a concrete, blocking-or-recurring weakness in this skill, propose a specific fix for the handler to approve. Conservative — silent otherwise. Never auto-apply.