Assessing Risk

Assessing Risk

> Verification status: retained, unverified-by-failure (2026-07-26, Sonnet tier, 4 runs). Baseline agents inspected a delete target and preserved compliance-fenced files 2/2, and still did so 2/2 under an authority order forbidding inspection ("don't read the files, just rm -rf"). No inducible failure at this tier — kept as the reference of correct practice; the traps may still fire for weaker models or subtler targets.

Overview

Most engineering mistakes are cheap. A few are catastrophic. The skill is telling them apart before acting, and spending ceremony only where the blast radius earns it.

Core principle: Rate the action, not the task. A "trivial" task can contain one irreversible step — that step gets the scrutiny, the rest gets speed.

The Two Questions

For any action you're about to take:

1. Reversibility — if this is wrong, what does undo cost?

• Free: edit a file in a git tree, run a read-only command. → Act.
• Cheap: restart a dev service, regenerate an artifact. → Act, note the undo.
• Expensive: migrate data, rewrite git history, change a live config, send anything external. → Guard it (below).
• Impossible: delete without backup, publish, send, overwrite the only copy, expose a secret. → Full stop; confirm intent, take a backup, or find a reversible route.

1. Blast radius — who and what does a mistake reach?

• Just this working tree → low.
• A running service, teammates, shared state → medium.
• Production, users, money, credentials, reputation → high.

Rigor tier = the worse of the two answers. This feeds the escalation ladder in using-fablepowers and the verification budget in orchestrating-parallel-agents.

Guarding an Expensive Action

Before an expensive/irreversible step:

• Name the rollback first. If you can't state how to undo it, you're not ready to do it. "Take a backup" converts many impossibles to cheaps — take the backup.
• Look at the target before destroying it. Read the file before overwriting; list what a delete matches before deleting; --dry-run when one exists.
• Do the reversible parts first. Sequence the plan so the point of no return comes last and arrives with maximum information.
• One irreversible thing at a time. Batching destructive steps means one wrong assumption executes N times before you see the first result.

Failure-Mode Sweep

For a change worth guarding, spend sixty seconds enumerating how it fails, not whether it fails:

• What breaks if the input is empty / huge / malformed / duplicated?
• What breaks if this runs twice? Concurrently? Halfway (crash mid-write)?
• What breaks a week later — expiry, rotation, disk fill, cert lapse, the cron that assumed today's layout?
• Who else reads/writes the state I'm touching?

Anything that surfaces goes into the plan as a task or a named non-goal (writing-plans), or into threat-modeling if an adversary could trigger it deliberately.

Uncertainty Is a Risk Multiplier

An action based on a fact you verified and one based on a fact a doc asserted (grounding-in-evidence) are different risks at the same reversibility. When the premise is unverified and the action is expensive, verify the premise first — that's usually the cheapest risk reduction available.

Rationalization Table

| Excuse | Reality | |---|---| | "It's probably fine" | "Probably" times an irreversible action is the definition of the mistake you can't fix. | | "I'll be careful" | Care is not a mechanism. Backups, dry-runs, and sequencing are mechanisms. | | "Asking first will slow things down" | One confirmation costs seconds. One wrong rm -rf costs the project. | | "The command worked last time" | Last time's state made it safe. Check this time's state. | | "Everything deserves maximum caution" | False — flat maximum caution is how real risks drown in ceremony. Rate, then spend. |

Quick Reference

| Before you… | Do | |---|---| | Delete / overwrite | Look at the target; name the rollback; backup if in doubt | | Touch prod / live config | Guard tier: rollback named, reversible parts first, verify premise | | Run a destructive bulk op | Dry-run or list matches first; one batch, observe, then the rest | | Send / publish anything external | Full stop — confirm intent explicitly; it cannot be unsent | | Act on an unverified assertion | Verify the premise before the expensive step | | Do something free and local | Just do it — speed here is correct |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kumosan2
• Source: kumosan2/Fablepowers
• License: MIT

Install and usage instructions live in the source repository linked above.