change-my-mind
Skill0 starsChanged 4 months agoMITView on GitHub
Attach falsification conditions and update triggers to any verdict, forecast, plan, or belief — the specific evidence, event, or date that would change it — and turn a static conclusion into a living one with a provenance ledger. Use whenever you reach a conclusion, publish a recommendation, make a forecast, take a position, or want a document that stays honest as the world moves. Triggers on "what would change your mind," "how confident are you," writing a decision memo or roadmap, "keep this up to date," post-mortems, and any claim meant to guide action over time. If a conclusion can't say what would overturn it, this skill applies.
In this repository11 other files it gives its agents.
- analogy-checkSkill
- disaggregateSkill
- disparate-impact-auditSkill
- grade-the-evidenceSkill
- hype-checkSkill
- prose-surgeonSkill
- protect-the-skillSkill
- scenario-branchSkill
- steelmanSkill
- validate-the-claimSkill
- value-framesSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
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 63
If the artifact will persist, set up the provenance ledger + version scheme.
≈ “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
Change My Mind
A conclusion that can't name what would overturn it isn't a conclusion — it's a belief wearing a lab coat. This skill makes every verdict falsifiable and every important document alive: it states, in advance, the specific evidence or event that would flip the call, who or what is authorized to trigger a revision, and when to check even if nothing obvious happened. The discipline is simple and it changes everything — a conclusion is retired not because it aged, but because a named piece of evidence moved.
The core deliverable: for every verdict, its kill conditions
For any load-bearing conclusion, specify the four trigger types below. You don't need all four every time, but you need at least the datapoint and the calendar check — a verdict with neither is not falsifiable.
- Datapoint trigger (easiest, most common) — a specific measurable result that would flip the verdict. "This reverses if the next official labor dataset fails to replicate the early-career employment effect." Testable, often on a known schedule.
- Incident trigger (hardest to predict, most important) — a discrete event that forces immediate re-evaluation, not a scheduled one. "Flips to negative if a second, higher-success incident of type X surfaces within 18 months." Don't wait for the calendar; evaluate within a stated window when it fires.
- Threshold trigger (for capability/scale claims) — a level crossing that upgrades or downgrades the evidence itself. "Converts from expert-consensus to measured-fact if a system is confirmed above threshold T."
- Calendar checkpoint — a default re-confidence check on a fixed cadence (e.g. no verdict goes untouched more than 12 months without a recalibration), even absent new evidence. Catches slow drift and stale priors.
Some claims are honestly non-falsifiable — say so
A few conclusions can't be falsified by advance evidence — only judged after the fact (e.g. one-shot tail risks with no base rate). Don't fake a trigger for these. Mark them explicitly as non-falsifiable-in-advance, report them as a distribution of judgment rather than a single value, and note that the honest move is to track the reasoning, not to pretend a datapoint will settle it. Naming this is itself a form of intellectual honesty the reader can rely on.
Make the document live: the provenance ledger
For anything meant to guide action over time, don't silently overwrite when the verdict changes. Keep a visible ledger:
- What the verdict was, and what evidence/event changed it, and when.
- Preserve the prior verdict rather than erasing it — how a conclusion's mind changed is itself valuable, and auditable.
- Version the document (major = the framing/method changed; minor = a new section/verdict; patch = an update to an existing verdict from new evidence).
The ledger is what converts a one-time analysis into a standard that earns trust by showing its work over time.
Procedure
- Restate the verdict/forecast/plan crisply.
- For each load-bearing conclusion, write its triggers: datapoint (required), calendar (required), plus incident/threshold where they apply.
- Mark any genuinely non-falsifiable-in-advance claims as such; report them as distributions, not points.
- List the leading indicators to watch and where they're published/observed — the standing monitoring list.
- If the artifact will persist, set up the provenance ledger + version scheme.
- State the review cadence and who/what may trigger an off-cadence revision.
Output template
VERDICT: <the conclusion, crisply>
KILL CONDITIONS:
Datapoint → flips if: <specific measurable result> (source, cadence)
Incident → re-evaluate within <window> if: <discrete event>
Threshold → upgrades/downgrades if: <level crossing>
Calendar → recheck by: <date/cadence> regardless of news
NON-FALSIFIABLE-IN-ADVANCE: <any claims that only resolve post-hoc — report
as a distribution, not a point>
WATCH LIST: <leading indicators + where to find them>
LEDGER (if living doc):
vX.Y — <date> — verdict: <…> — changed because: <evidence/event> — prior: <…>
REVIEW CADENCE: <interval> | TRIGGER AUTHORITY: <who/what forces an update>Example (compressed)
Verdict: "Our product bet assumes agentic coding tools will reliably handle multi-hour tasks by end of next year."
KILL CONDITIONS:
Datapoint → flips if the standard task-horizon benchmark's doubling trend
breaks across two consecutive quarterly updates (either a plateau or a
jump changes the timeline).
Incident → re-evaluate within 2 weeks if a shipped model is found to have
silently regressed on long-horizon reliability post-deployment.
Threshold → upgrades from "assumed" to "confirmed" when an independent eval
shows >80% reliability at the multi-hour mark (not vendor-reported).
Calendar → recheck the whole assumption every quarter regardless.
WATCH LIST: the task-horizon benchmark (quarterly), independent eval reports,
our own internal reliability logs.
REVIEW CADENCE: quarterly | TRIGGER AUTHORITY: eng lead may force an off-cycle
review if the internal logs diverge from the benchmark.Now the bet isn't a hope — it's a claim with a tripwire. If the trend breaks, you find out on purpose instead of in a retro.
Gotchas
- A trigger must be observable and specific. "If things change a lot" is not a trigger. "If metric M crosses V by date D" is.
- Don't overwrite history. The prior verdict and the reason it changed are the point — preserve them.
- Required minimum: datapoint + calendar. Anything less isn't falsifiable.
- Pairs with scenario-branch (its branch triggers become your update conditions) and grade-the-evidence (a trigger firing often raises the tier of a claim, e.g. consensus → measured).
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.

