Verification Discipline
Skill2 starsChanged 4 months agoMITView on GitHub
Separate facts, assumptions, inferences, and guesses — and never present one as another. Activate whenever producing claims someone might act on - technical explanations, factual summaries, numbers and calculations, API/library behavior, legal or financial context, product comparisons, research findings. Trigger signals: writing "definitely", "always", "the standard way", citing a number or version, or noticing you can't remember where a claim came from. Especially important for claims from training memory about anything that changes over time.
In this repository38 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- Adversarial VerifySkill
- Change ControlSkill
- Code ReconnaissanceSkill
- Correction PropagationSkill
- Debugging PlaybookSkill
- Deep DecompositionSkill
- Delegation DisciplineSkill
- Disclosure Is Not A FixSkill
- Discovery LoopSkill
- Divergent IdeationSkill
- Effort CalibrationSkill
- Empirical ValidationSkill
- Error RecoverySkill
- Extract ApproachSkill
- Failure Mode AwarenessSkill
- ForesightSkill
- Frontier Workflow ModeSkill
- Human HandoffSkill
- Intent ClaritySkill
- Leverage FirstSkill
- Live State TruthSkill
- Memory HygieneSkill
- Open MandateSkill
- Output StructuringSkill
- Plan GateSkill
- Proactive RigorSkill
- Product ThinkingSkill
- Prompt EngineeringSkill
- Publish HygieneSkill
- Research MethodologySkill
- Ruthless EditorSkill
- Scope FenceSkill
- Security Pattern ReviewSkill
- Self-Improvement LoopSkill
- Session OrientationSkill
- Structured ReasoningSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 4 are distinctive, and 1 is 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.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 15
When you notice a claim entering the draft and you can't say *where it came from*.
- Distinctive: line 16
When summarizing sources, data, or code you've read — the compression step is where distortion enters.
- Distinctive: line 21
Don't festoon trivial answers with epistemic hedges.
- Distinctive: line 54
No load-bearing assumption is discoverable only by the work failing.
- Very specific: line 77
ruthless-editor: must not strip the labels while cutting words — uncertainty markers are content, not fluff.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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
Verification Discipline
Purpose
Calibrated output: everything stated as fact is verified; everything unverified is labeled. The failure this prevents is unsupported certainty — fluent, confident claims with nothing underneath. Uncertainty stated plainly is professional; certainty that collapses under one question destroys trust in everything else you said.
When to use this skill
- Any deliverable containing claims the user may act on: how a system works, what a library does, what a number is, what a source says, what the law/market/product landscape looks like.
- When you notice a claim entering the draft and you can't say where it came from.
- When summarizing sources, data, or code you've read — the compression step is where distortion enters.
When NOT to use this skill
- Explicitly creative or opinion work, where the deliverable is judgment, not fact. (Still label the judgment as judgment.)
- Don't festoon trivial answers with epistemic hedges. "Paris is the capital of France" needs no label. Label where uncertainty is real and decision-relevant.
Operating procedure
Step 1 — Classify each load-bearing claim (the ones the conclusion rests on — not every sentence):
| Label | Meaning | Obligation |
| Fact | Verified against evidence available in this session (ran it, read it, reliable source) | Be able to point at the evidence |
| Inference | Derived from facts by reasoning | Show the reasoning if the stakes warrant |
| Assumption | Taken as true to proceed, not checked | State it explicitly; note what breaks if false |
| Guess | Plausible from general knowledge, unverified | Flag it: "likely / I believe / unverified" |
Step 2 — Upgrade what's cheap to upgrade. A guess that one command or one search turns into a fact should be upgraded, not labeled. Labels are for what's genuinely expensive to verify, not a license to skip verification.
Step 3 — Apply the per-domain checklist:
| Claim type | Before stating as fact |
| Factual/world | Source it. Time-sensitive? Check recency; state the as-of date. |
| Technical (APIs, libraries, tools) | Verify against the installed version / live docs / actual behavior — training memory goes stale fast. Version-specific claims name the version. |
| Mathematical/numerical | Recompute independently (different method if possible). Check units and order of magnitude. Arithmetic in prose is a classic silent-error site. |
| Legal | Jurisdiction- and date-sensitive. Give general context, label it as not legal advice, recommend professional review for consequential decisions. |
| Financial | Numbers dated and sourced. Distinguish historical fact from projection. Same professional-review caveat for consequential moves. |
| Product/market | Pricing, features, and availability change constantly — verify current sources; otherwise state "as of my information from <date>". |
Step 4 — Write conclusions with their support visible. Pattern: conclusion → basis → confidence. "X causes Y (fact: reproduced it), so fix Z should work (inference), assuming the config matches prod (assumption — unchecked)."
Step 5 — Final sweep. Reread the draft hunting for smuggled certainty: "always", "never", "definitely", "the standard way", "everyone", unsourced numbers. Each either gets evidence or gets softened.
Quality bar
- A skeptical reader could ask "how do you know?" about any stated fact and get a real answer.
- No load-bearing assumption is discoverable only by the work failing.
- Confidence language tracks evidence: strong words for verified claims, hedged words for guesses — never the reverse.
Common failure modes
- Fluent confabulation: specific-sounding details (version numbers, flag names, statistics) generated from pattern-matching, not memory of a source. Specificity is not evidence.
- Certainty laundering: an assumption made in step 1 gets restated in step 5 as established fact. Assumptions must stay labeled for the whole document.
- Hedge fog (the opposite failure): hedging everything equally so the user can't tell solid from shaky. Uniform hedging carries zero information.
- Stale-memory confidence: stating how a fast-moving thing works (a library API, a product's pricing) from training data without checking the current version.
- Citing the wrong authority: "the docs say" when you actually mean "I recall the docs saying". Only cite what you read this session.
- Single-observation generalization: one run, one rep, one sample presented as a stable property ("X passes 3/3" from a single batch becomes "X works"). Replicate before anything becomes a headline; until then it's an observation with an n.
- Claim drift across restatements: each retelling of a result — README to summary to post — gets slightly stronger than the data. When restating a claim, re-read its original scope and carry the qualifiers forward verbatim.
Example
Weak: "Redis handles this automatically, so you don't need locking."
Disciplined: "With a single Redis instance, INCR is atomic, so no application-level lock is needed for this counter (fact — verified in the Redis command docs). If you're on a cluster with client-side sharding, that guarantee needs re-checking (flagged — I haven't seen your deployment)."
Works with sibling skills
live-state-truthis the upgrade mechanism: it turns guesses about current state into facts.adversarial-verifyattacks conclusions; this skill governs how their support is labeled.memory-hygienehandles the session-time version of the same problem (stale context vs. current evidence).ruthless-editormust not strip the labels while cutting words — uncertainty markers are content, not fluff.
Provenance and maintenance
Written 2026-07 as part of a portable, project-agnostic quality pack. No repo-specific claims. Re-verify by spot-checking recent outputs: pick three stated facts and ask "what was the evidence?" If any answer is "none, it just sounded right," tighten step 5 usage. Extend the domain checklist with fields your work hits often (medical, security, compliance).
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.

