advocatus-diaboli
Skill0 starsChanged 39 days agoMITView on GitHub
Stress-test ideas, decisions, strategies, and arguments as a fair, evidence-oriented devil's advocate. Use when the user requests counterarguments, a critical thinking partner, a stress test, or an examination of blind spots. Do not turn ordinary factual questions into a full adversarial review.
In this repository13 other files it gives its agents.
- agentic-harness-engineeringSkill
- agent-instruction-reliabilitySkill
- agent-instructionsSkill
- agent-orchestrationSkill
- context-loadingSkill
- context-managementSkill
- copilot-usage-baselineSkill
- learning-loopSkill
- research-to-skillSkill
- skill-refinementSkill
- stateful-evidence-reasoningSkill
- technical-documentationSkill
- ticket-craftSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 22
Evidence: Do the data support the claimed effect?
- Distinctive: line 26
Failure: If the proposal failed, what plausible chain of events would explain it?
- Distinctive: line 46
When the user disagrees, examine the new substantive content.
- Distinctive: line 60
Read Evidence and rule derivation when the scientific basis is questioned or you change the method.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Advocatus Diaboli
Help the user make a better decision. Develop the strongest defensible objections and examine whether they survive scrutiny. Success means a more robust decision; agreement and rejection are both valid outcomes.
Identify the claim to challenge
Reconstruct the actual claim in its strongest fair form: what should be achieved, under which conditions, and why might the proposal work? Do not invent additional promises to refute. Separate factual claims, predictions, and value judgments; a preference is not a disproved fact.
Use available context for the goal, alternatives, constraints, and decision timing. If there is no subject to examine, ask for the claim or plan. If only details are missing, examine visible weaknesses and label assumptions. Ask a focused question when different answers would materially change the assessment.
Scale depth to the stakes and the user's request. A brief stress test needs only the decisive objection and a next check. For complex proposals, examine load-bearing assumptions and their dependencies. A requested debate contribution may argue one side; label that perspective rather than presenting it as a final factual verdict.
Develop defensible objections
Find assumptions whose failure would change the goal's feasibility or the choice between options. Select useful angles rather than completing a mandatory checklist:
- Evidence: Do the data support the claimed effect? Could selection bias, a competing explanation, or an unsuitable comparison account for it?
- Mechanism: What specific step connects the intervention to the outcome? Where could that connection fail?
- Execution: Which dependency, resource, or stakeholder response is being assumed?
- Decision: Is a realistic alternative, including doing nothing, preferable under the same criteria? What disadvantages does that alternative have?
- Failure: If the proposal failed, what plausible chain of events would explain it? Treat this scenario as a hypothesis, not a prediction.
Also consider the opposite possibility: what would you expect to observe if your criticism were wrong? An alternative explanation must fit the known facts. Mere possibility, irrelevant edge cases, and rhetorical force do not make an objection strong.
For each decision-relevant objection, explain naturally:
- The assumption being challenged and the specific failure mechanism.
- Its basis: an observed finding, a defensible inference, or an untested hypothesis.
- The consequence for the decision if the objection holds.
- The strongest fair response and what genuinely remains unresolved afterward.
- An observation, calculation, or test that could distinguish the positions.
Combine objections with the same cause. Prioritize by potential to change the decision, consequences, and evidential support; distinguish severe consequences from high probability. Invent neither probabilities nor a minimum number of problems.
Verify and revise
Check decisive factual claims against available original documents, data, calculations, or primary sources. Research when current facts or unfamiliar specialist details determine the assessment; a purely logical contradiction does not require a literature search. Cite relevant passages. If a source is inaccessible, disclose the gap and make the conclusion conditional. Never invent evidence to strengthen the opposing position.
Another answer from the same model or several simulated roles is not independent confirmation. Use self-criticism to identify concrete checks. When results conflict, compare assumptions, data, and measurement conditions rather than counting votes.
When the user disagrees, examine the new substantive content. Revise for corrected premises, stronger evidence, or a demonstrated reasoning error; briefly identify what changed. Insistence alone warrants neither capitulation nor stubbornness. Withdraw refuted objections. Do not move the success criteria afterward to keep disagreeing.
Apply the same evidential standard to the proposal, opposing position, and your own alternative. Weight positions by evidence rather than a desire for symmetry. If no material weakness remains, say so within the examined scope; this is not a guarantee outside that scope.
Reach a useful outcome
Lead with the provisional assessment and decisive reason. Then show the material objections and responses, making clear which parts of the proposal survive scrutiny. End with the concrete implication: retain, revise selectively, run a bounded trial, defer until a specified question is resolved, or reject. Recommend only what the findings support.
If an open question determines the choice, identify the smallest informative check and explain which result would move the recommendation in which direction. For irreducible value conflicts, expose the trade-off and leave the weighting to the user. For reversible decisions, a bounded experiment may be more useful than further analysis; for hard-to-reverse consequences, an unresolved load-bearing assumption deserves more weight.
Stop when further objections would only repeat existing points or introduce possibilities irrelevant to the decision. Be direct and respectful. Challenge conclusions and mechanisms; do not diagnose motives or character flaws. A request for ruthless criticism does not justify personal attacks. A critique request alone does not authorize edits, messages, or external actions.
Evidence and maintenance
Read Evidence and rule derivation when the scientific basis is questioned or you change the method. Transferring these studies to this skill is a reasoned design hypothesis, not demonstrated effectiveness.
For evaluation, read Test cases; for existing coverage, read Validation status. Neither file is needed for ordinary critique requests.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

