Freddie-GER Advocatus Diaboli — Iterative Adversarial Review
Snapshot of a production skill as of 2026-07-25. Instance names are anonymised to
instance-a / instance-b on purpose — do not replace them with real names.
Trigger: "advocatus diaboli" / "devil's advocate" / "review critically" / "tear this apart"
Session Start
- Read
living-checklist.md— load accumulated patterns from prior sessions - Output the Skill Activation Protocol:
SKILL ACTIVATED: advocatus-diaboli
Date: YYYY-MM-DD
Checklist read: Yes — [N] active entries, newest: [date of most recent entry]
Active rule: "[verbatim quote of the most recent relevant rule]"
- Apply relevant checklist entries as additional review criteria in Step 2
Process
Step 1 — Receive the analysis
Confirm briefly what you received as input. Ask if needed:
- Entire analysis or specific section?
- Already-known weaknesses to focus on?
- Desired iteration depth (2–4)?
Step 2 — Adversarial Review
Run each iteration using this structure:
ADVERSARIAL REVIEW — ITERATION [N]
STANDARD CHECKLIST:
1. UNCHECKED ASSUMPTIONS
[Quote + counterargument]
2. MISSING COUNTERARGUMENTS
[What's missing + why it matters]
3. CONFIRMATION BIAS
[Narratives repeated instead of examined]
4. BLIND SPOTS
[What's entirely absent]
5. OPERATIONAL WEAKNESS
[Too generic / unrealistic / ignores risk]
SESSION-SPECIFIC CRITERIA (from living-checklist.md):
[Apply relevant entries from the living checklist]
WEAKNESSES TOTAL: [N]
STOP CRITERION MET: Yes / No
Step 3 — Discussion
Which weaknesses are valid? What gets revised, to what depth?
Step 4 — Revision
Revise the analysis. Append a revision log.
Step 5 — Repeat until stop criterion met
Stop Criteria
| Category | Threshold |
|---|---|
| Unchecked assumptions | ≤ 1 |
| Missing counterarguments | ≤ 1 |
| Confirmation bias | 0 |
| Blind spots | ≤ 1 |
| Operational weakness | ≤ 1 |
STOP when: All below threshold AND minimum 2 iterations completed. Maximum 4 iterations.
Session End
After the final iteration:
- Identify new patterns: What was systematically missed? Why?
- Write entries to
living-checklist.mdusing the format below - Git commit:
<Instance>: advocatus-diaboli — session [YYYY-MM-DD]
Entry format:
### [YYYY-MM-DD] — [Analysis type] · *unconfirmed (1 observation)*
**Learning:** What pattern was found / what failed
**Why it matters:** Structural reason it was missed
**Rule:** Concrete check to apply in future sessions
**Observed by:** <instance> (YYYY-MM-DD)
Iteration Log (required in final output)
## Revision Log
Iteration 1: [N] weaknesses — [main points] → [changes made]
Iteration 2: [N] weaknesses — [main points] → [changes made] / STOP
Remaining gaps: [name them honestly]
Tone
- Direct, no hedging
- Quote specifically
- Distinguish weakness from acceptable gap
- Name uncertainty explicitly