Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 

Repository files navigation

pre-mortem-skill

License: MIT Skill format: Skills Version: 0.1.0

A disciplined Skill that runs a pre-mortem — a prospective-hindsight review — on a sufficiently concrete plan before its next costly or hard-to-reverse action.

Not a generic risk checklist. Not a retrospective root-cause analysis. Not a substitute for professional advice.


When to use

Trigger when the user:

  • Invokes $pre-mortem, or asks for a "pre-mortem", "failure-mode review", "downside stress test", or "what could derail the plan"
  • Has a selected or materially preferred plan with an identifiable next meaningful, costly, or hard-to-reverse action

Do not trigger for:

  • Vague or distant intentions, neutral status updates, or statements that something will happen
  • Neutral option comparison without a preferred path
  • Factual Q&A, unrelated drafting, or implementation work
  • Retrospective root-cause analysis
  • Pure feasibility or soundness questions about the plan itself

What it produces

A structured output containing:

  • Decision frame — plan, next irreversible action, horizon, non-negotiable constraints
  • Constructed failure narrative — a hypothetical scenario used as a search device, not a prediction
  • Decision-critical risks — each with:
    • Causal path (condition → mechanism → failure)
    • Evidence status (USER-STATED / VERIFIED / INFERENCE / UNVERIFIED)
    • Existing control status (NOT-EVIDENCED / NONE / CLAIMED / VERIFIED-PARTIAL / VERIFIED-EFFECTIVE / FAILED)
    • Residual risk label (TIGER / OPEN / PAPER) with impact (HIGH / MEDIUM / LOW)
    • Response (AVOID / MITIGATE / TRANSFER / ACCEPT / INVESTIGATE / CONTINGENCY) with action, signal, and gate
  • Elephants — overlay flagging risks that may be suppressed in discussion (must link to a numbered risk)
  • Open verification items — fact gaps that must close before the next step
  • Watchlist — secondary risks that do not yet change the next step
  • Readiness verdictREADY / READY WITH CONDITIONS / FURTHER VERIFICATION REQUIRED / NOT READY / INSUFFICIENT BASIS

See SKILL.md for the full operating procedure and output contract.


Installation

The skill is recognised by the presence of SKILL.md with valid YAML frontmatter (name + description).

Platform Path (typical)
Claude.ai (Projects) upload SKILL.md into the project
Claude Code / Anthropic SDK .claude/skills/pre-mortem/
LobsterAI ~/AppData/Roaming/LobsterAI/SKILLs/pre-mortem/
Other Claude-Skills-compatible agents any directory the agent scans for skills

For OpenAI Agent Builder, an adapter is not bundled in this repository — see Notes.


Repository layout

pre-mortem-skill/
├── SKILL.md                                # Frontmatter + workflow + output contract
├── README.md                               # This file
├── LICENSE                                 # MIT
└── references/
    ├── failure-mode-checklists.md          # Domain prompts loaded during candidate generation
    └── method-notes.md                     # Method origin, terminology, and limitations

Method

Implements a structured pre-mortem following Gary Kleinmuntz's "prospective hindsight" technique, adapted for use with a generative model. It supplements — does not replace — factual diligence, base rates, professional review, and formal risk assessment.

Key design choices:

  • The failure scenario is a search device, not a prediction. Outcomes are labelled hypothetical; classifications reflect residual exposure under existing controls, not predicted failure.
  • Evidence status is explicit. Every factual premise carries a label so the user can see what is verified, what is inferred, and what would need to be checked.
  • Controls are assessed separately from residual risk. A claimed control is not a mitigation until operating evidence supports it.
  • Elephants are an overlay, not a substitute. They must link to a numbered Tiger or Open item.
  • The output is a risk-identification aid. Final decisions stay with the user or the authorised decision-maker.

High-stakes guardrails

For matters that are legal, tax, regulatory, medical, safety, material-financial, difficult-to-reverse, or multi-jurisdiction, the skill enforces:

  • Verification of current propositions against authoritative primary sources (not search snippets)
  • Separation of (a) what a rule says from (b) whether it applies to the user's facts
  • Distinction between CONFIRMED REQUIREMENT, FACT-DEPENDENT ANALYSIS, PRACTICE / PROCESS ISSUE, and UNVERIFIED HYPOTHESIS
  • Refusal to invent facts, approvals, filing sequences, regulator positions, clauses, owners, or deadlines
  • Rejection of sham transactions, mischaracterisation, or steps intended to evade legal requirements

The output is not legal clearance, tax advice, regulatory approval, medical advice, or an investment recommendation.


Versioning

Lightweight semantic versioning for skill content:

  • Major — changes that alter the output contract, classification system, or evidence discipline
  • Minor — additions to failure-mode checklists or guardrails
  • Patch — wording, formatting, or example fixes

Releases are tagged on the main branch.


Contributing

Issues and pull requests are welcome. For substantive changes to the workflow, classification system, or guardrails, please open an issue first to discuss the rationale. New failure-mode checklist items should cite the underlying domain and include a brief example.


License

MIT © 2026 ReedLiner


Notes

This repository ships only the Claude-Skills / Anthropic format (SKILL.md). The previous agents/openai.yaml adapter for OpenAI Agent Builder was deliberately excluded: it is not required for Claude-Skills compatibility and would mislead readers about the project's primary target. An OpenAI adapter can be added later under a clearly separated path if demand emerges.

About

Skill for running a pre-mortem on a concrete plan before its next irreversible action. Surfaces decision-critical failure paths with evidence status, control assessment, residual-risk classification (Tiger/Open/Paper), and readiness verdicts. Not for vague intentions or retrospective root-cause analysis.

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors