| name | double-checking-before-acting |
| description | Checks a critical agent action against its exact target, expected result, and stop condition before and after execution. Use when an agent is about to make a critical edit, run a command or migration, use a credential or tool, change a dependency or model, make a public claim, or affect a release. Do not use for low-stakes reversible edits. |
Double-Checking Before Acting
Overview
A self-check turns a high-stakes action into a careful one. Before you claim success, name the target, the result you expect, when to stop, the action itself, and the check you will run afterward.
Decision contract
- Claim checked: the target, expected result, and stop condition were named before the action, and the real result was compared against the expected one before claiming success.
- Artifact observed: the proposed action, target, and authority -> a
self-check.md note of action, target, expected/real result, stop condition, evidence, and any mismatch.
- Decision affected: block -- whether to proceed with or abort the critical action: a big edit, command, public claim, trust change, or release.
- Failure class: wrong-target-or-unverified-action (acting on an assumed target, or claiming success without comparing the result).
- Next action: pause and escalate when authority, target, or evidence is unclear; a mismatch halts the action rather than triggering a retry.
When to Use
- A command can delete, move, publish, release, migrate, or change something outside the repo.
- An edit touches public claims, where ideas came from, permissions, credentials, dependencies, models, APIs, or the release.
- An agent is about to make a wide or repeated change, where hitting the wrong target is easy.
- A quick candidate is about to become a public claim, an accepted version, or a release action.
When Not to Use
- The task is a tiny local edit with obvious proof.
- The action only reads data and has no effect on any later decision.
- The change record already requires a stronger human check before the action.
Inputs
- The action you mean to take, the exact target, the current phase, and where your authority comes from.
- The result you expect, the most likely wrong-target mistake, and when to stop.
- The proof command, review check, or evidence link you will need after the action.
Process
- Stop at the key moment — the interval after you have chosen the action and before you commit it — and name the exact action and target.
- Think through the result you expect, the likely error, and what would make the action wrong.
- Act only inside the authority you were given.
- Compare the real result against the result you expected, before you make any claim.
- For trust-bearing actions, get a forceful backup — an independent person, or a different model or context — to check the action. A second pass by the same agent in the same context inherits the same blind spots and is not independent.
- Record any mismatch, gap, or escalation. Do not just retry blindly.
Outputs
- A short self-check note, or a
self-check.md file.
- The action, target, expected result, stop condition, real result, and evidence.
- An escalation note when the result does not match.
Verification
- The target is exact enough to rule out work on the wrong file or the wrong environment.
- The expected result is named before the action.
- The after-action check compares the evidence against what you expected.
Escalation
- Pause if the authority, the target, the expected result, or the evidence is unclear.
- Halt the action if it is unsafe or unclear, regardless of seniority; surfacing the concern is protected, not punished.
- Escalate when the action affects credentials, network effects, data, releases, public trust, or anything that cannot be undone.
Common Rationalizations
- "It is only one command." A single command can do wide damage.
- "The target is obvious." Wrong-target mistakes come from assumed targets.
- "I can check after." Checking after only helps if you named the expected result first.
Red Flags
- The action starts before the target is named.
- The expected result is missing or vague.
- A mismatch is treated as a reason to retry instead of a reason to pause.
- Public wording claims safe, secure, approved, or compliant with no scoped evidence behind it.
Prompt
Self-check this Nuclear-grade agent action before it happens.
Inputs:
- packet:
- current phase:
- intended action:
- exact target:
- authority source:
- expected result:
- likely wrong-target or wrong-state error:
- stop condition:
- proof or after-action check:
Return:
- the risky point being checked;
- the action and the target;
- the expected result;
- the stop condition;
- the evidence to collect after the action;
- whether to go ahead, pause, or ask for help.
Source-lineage note
This skill is an original software-workflow translation of self-checking, pause when unsure, flagging, procedure adherence, and verification practices from DOE-HDBK-1028-2009 as public source lineage. It does not create DOE compliance, formal assurance, safety, security, certification, or regulatory adequacy.
