raisedadead/challenge

Preamble

Running: challenge — Adversarial review of the draft from multiple critical lenses. Expect: A findings table with severity ratings, a pre-mortem summary, and an approve/revise/reject gate.

dp-cto:challenge — Adversarial Review

<!-- Two-layer deep-reasoning signal: challenger.md frontmatter `effort: high` sets agent-scope depth; this keyword sets skill-scope depth when the skill body is loaded without agent dispatch. Both are intentional. -->

Use ultrathink for this review — adversarial lenses demand deep reasoning, not surface pattern-matching.

Read references/anti-rationalization.md before proceeding.

Entry Condition

Stage must be drafted (document authored by /dp-cto:draft). The stage machine hook enforces this.

Step 0: Locate the Drafted Document

1. Read the stage state file (.claude/dp-cto/*.spec-stage.json) to find the current session context
2. Locate the drafted document (typically in docs/, specs/, rfcs/, or project root)
3. Read the document to confirm it exists and is complete
4. If no document is found: "No drafted document found. Should I run draft first?" STOP.

Step 1: Determine Lens Set

Read references/adversarial-lenses.md to discover the available lenses, contradiction detection protocol, pre-mortem format, and finding format.

Determine which lenses to apply:

• If the user passed specific lens names (e.g., challenge scale security): run only those lenses. Validate each name against the lenses defined in adversarial-lenses.md. If any name does not match, report the invalid name and list the available lenses. STOP.
• If no lens names passed (default): run all lenses defined in adversarial-lenses.md.

Step 2: Spawn 1 Adversarial Review Agent

Spawn exactly 1 dp-cto:challenger named Agent with a prompt constructed from:

The agent prompt must include:

1. The full document path and instruct the agent to read it
2. The list of lenses to apply (from Step 1)
3. Any user-provided context or constraints about the spec

Step 3: Synthesize Findings

After the agent completes:

1. Parse output for [CRITICAL], [WARNING], and [SUGGESTION] findings
2. Deduplicate: if two lenses flag the same section, keep the higher-severity one
3. Read references/findings-template.md for the consolidated findings table format.
4. If zero findings: "Clean bill of health. No adversarial findings." Skip to Step 5.

Step 4: Pre-Mortem Review

Review pre-mortem scenarios from the agent output. Read references/pre-mortem-template.md for the scenario format. If the agent did not produce scenarios, synthesize them from the findings.

Step 5: Present for Approval

Present the consolidated findings table, pre-mortem scenarios, and a summary:

## Challenge Complete
**Document**: {document_path}
**Lenses applied**: {comma-separated list of lenses applied}
**Findings**: {critical_count} critical, {warning_count} warnings, {suggestion_count} suggestions
**Pre-mortem scenarios**: {scenario_count} identified

Use AskUserQuestion with exactly three options:

• Approve — findings accepted as-is, proceed to handoff
• Revise — incorporate feedback to address findings (max 2 revision rounds, then must approve or reject). On revise: update the document, re-run affected lenses, and present again with the same three options.
• Reject — findings reveal fundamental issues, return to draft phase

If the user selects Reject, stop and print: "Challenge rejected. Should I return to draft to revise the document?"

If the user selects Approve, print: "Challenge complete. Should I continue with handoff to produce the implementation-ready package?"

Do NOT invoke handoff. The user decides when to proceed.

<CHAIN> Challenge complete. The next step is /dp-cto:handoff. The user decides when to run it. </CHAIN>

NEVER

| # | Rule | | --- | ------------------------------------------------------------------------- | | 1 | NEVER review the document yourself — the agent applies the lenses | | 2 | NEVER skip default lenses unless user explicitly passes a lens subset | | 3 | NEVER accept all findings uncritically — present for user approval | | 4 | NEVER dismiss all findings — each was found for a reason | | 5 | NEVER skip the pre-mortem — it catches systemic risks lenses miss | | 6 | NEVER auto-invoke handoff — the user decides when to proceed | | 7 | NEVER allow more than 2 revision rounds — after 2, only approve or reject |

Read references/red-flags.md before proceeding. It contains the stop conditions for this skill.