Plan Cross Review

Plan Cross-Review

Dual-model review of the current Plan-mode plan, with a cross-critique pass, then a consolidated proposal list for the user.

Models (exact Task model slugs)

When talking to the user, use the labels. When calling Task, use only the slugs above. Do not substitute other models.

Preconditions

1. There must be an active Plan-mode plan in this conversation (the plan the user is currently building/editing in Plan mode).
2. If no plan content is available, stop and ask the user to open/share the plan or switch into Plan mode first. Do not invent a plan.
3. Do not edit the plan unless the user explicitly asks after seeing the summary.

Workflow

Copy and track:

Plan Cross-Review Progress:
- [ ] 1. Capture current plan
- [ ] 2. Parallel independent reviews
- [ ] 3. Cross-critique pass
- [ ] 4. Consolidate and present to user

Step 1: Capture current plan

Obtain the verbatim current Plan-mode plan text (the exact document/body from Plan mode).

• Prefer the latest plan version if multiple exist.
• Store it once as ORIGINAL_PLAN and reuse that identical string in every Task prompt (Steps 2 and 3).
• Optionally note user constraints separately from the plan block — never merge them into the plan text.
• Forbidden: rewriting, condensing, outlining, "key points", paraphrasing, or relabeling a summary as the full plan.

Plan payload integrity (mandatory)

In every subagent prompt (independent review and cross-critique):

1. Include a fenced block labeled exactly ORIGINAL_PLAN (verbatim — do not treat anything else as the plan):.
2. Paste ORIGINAL_PLAN byte-for-byte / character-for-character — same headings, lists, paths, todos, open questions. No truncation for length.
3. Never invent a section titled like "Vollständiger Plan" / "Full plan" that contains a summary. If you must refer to completeness, only the verbatim block counts.
4. Cross-critique prompts must use the same ORIGINAL_PLAN capture as Step 2 — not a rewrite informed by the reviews, not a "cleaned" plan, not deltas applied.

If the plan is too large for one prompt, split only by contiguous verbatim chunks with explicit part markers (ORIGINAL_PLAN part 1/N …) that concatenate to the original; still never summarize.

Step 2: Parallel independent reviews

Launch two Task subagents in the same turn (run_in_background only if Multitask Mode requires it; otherwise wait for both).

Use subagent_type: "generalPurpose" for both.

Agent A — Fable 5 High

• model: claude-fable-5-thinking-high
• description: Fable plan review

Agent B — GPT 5.6 SOL High

• model: gpt-5.6-sol-high
• description: GPT plan review

Give each the same review brief (below) plus the verbatim ORIGINAL_PLAN block. They must not see each other's output in this step.

Independent review brief (same for both)

You are reviewing a Cursor Plan-mode implementation plan. Do not implement anything.

Evaluate:
1. Completeness vs stated goals
2. Correctness / technical risk
3. Sequencing and dependencies
4. Missing steps, edge cases, verification
5. Scope creep or unnecessary work
6. Ambiguities that should be resolved before coding

Return ONLY this structure:

## Verdict
One of: Ready / Needs revision / Major gaps
One-sentence rationale.

## Strengths
- ...

## Issues
For each issue:
- Severity: Critical | Important | Nice-to-have
- Where: which plan section/step
- Problem: ...
- Suggested change: concrete plan edit (wording or structural)

## Open questions
- ...

## Proposed plan deltas
Numbered, actionable edits the user could apply to the plan (not code).

Step 3: Cross-critique pass

After both independent reviews return, launch two more Task subagents in parallel.

Critical: Cross-critique input for the plan is again the unchanged ORIGINAL_PLAN from Step 1. Do not pass a summarized plan, a “revised” plan, or any reconstruction labeled as complete/full. Only the peer review text is new input.

Blind peer review (mandatory): Cross-critique prompts must not reveal which model wrote the review being critiqued. Do not name Fable, GPT, Claude, OpenAI, model slugs, or “Agent A/B”. Strip any such labels from the pasted review if present. Refer only to a neutral PEER_REVIEW block. The parent agent keeps the mapping privately for Step 4.

Agent C — Fable cross-critique (claude-fable-5-thinking-high, description: Fable cross-critique)

• Input: verbatim ORIGINAL_PLAN + independent review from GPT, labeled only as PEER_REVIEW
• Task: agree/disagree with each issue and proposed delta; note blind spots; flag weak or wrong suggestions.

Agent D — GPT cross-critique (gpt-5.6-sol-high, description: GPT cross-critique)

• Input: verbatim ORIGINAL_PLAN + independent review from Fable, labeled only as PEER_REVIEW
• Same critique instructions as Agent C.

Cross-critique brief

You are cross-checking a peer review of a Plan-mode plan. Do not implement.

You receive:
(1) ORIGINAL_PLAN — verbatim Plan-mode text. Treat only this block as the plan. Ignore any implied rewrite.
(2) PEER_REVIEW — an anonymized review of that same original plan. Do not try to identify who wrote it.

Do not assume the plan was edited. Critique the peer review against the verbatim original only.

For each Issues item and Proposed plan delta in PEER_REVIEW:
- Stance: Agree / Partial / Disagree
- Why (1-2 sentences)
- Better alternative if Disagree or Partial

Then add:
## Added findings
Issues or deltas the peer review missed (same severity scale).

## Cross-pass summary
What should definitely change in the plan vs what is optional noise.

Step 4: Consolidate and present to user

Synthesize all four outputs yourself (parent agent). Do not auto-apply plan edits.

Present in this format:

# Plan Cross-Review

## Consensus
- Shared verdict (reconcile Ready / Needs revision / Major gaps)
- 3–7 bullets both models effectively agree on

## Disagreements
| Topic | Fable 5 High | GPT 5.6 SOL High | Recommendation |
|-------|----------------|------------------|----------------|
| ... | ... | ... | ... |

## Recommended plan changes
Ordered by priority (Critical → Important → Nice-to-have).
For each:
1. **Change:** …
2. **Why:** …
3. **Supported by:** Fable / GPT / both (after cross-pass)

## Optional / contested
Changes only one model still pushes after cross-critique.

## Suggested next step
Ask whether to apply specific recommended changes to the Plan-mode plan (list them), and wait for user confirmation before editing the plan.

Keep the user-facing summary concise. Link or briefly quote plan sections instead of dumping raw subagent transcripts unless the user asks for full reviews.

Rules

• Always run the two independent reviews in parallel, then the two cross-critiques in parallel.
• Never skip the cross-critique pass.
• Never alter, summarize, or relabel the plan payload when calling subagents — especially in the cross-critique round. Verbatim ORIGINAL_PLAN only.
• In cross-critique prompts, never disclose the peer reviewer's model identity (names, brands, or slugs); use only PEER_REVIEW.
• Never modify the plan, code, or files as part of this skill unless the user explicitly requests applying changes afterward.
• If a Task/model call fails (unsupported model), report which slug failed and stop that branch; still deliver whatever completed, clearly marked incomplete.
• Do not mention internal skill authoring details unless asked.