Correction Propagation
Skill2 starsChanged 4 months agoMITView on GitHub
When a published claim is corrected, retracted, downgraded, superseded, or reversed, sweep EVERY surface that restates it - sibling docs, other repos, mirrors, READMEs, indexes, memory files - and banner or supersede each in the SAME closeout before it's done. Activate on "I corrected/retracted/updated X", a metric downgraded, a status flipped, a conclusion reversed, a benchmark restated elsewhere, "did the correction propagate", closeout after a correction, or a claim in more than one repo/doc. Method: grep the claim's distinctive tokens (number, phrase) across ALL surfaces, top-of-file first.
In this repository38 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- Adversarial VerifySkill
- Change ControlSkill
- Code ReconnaissanceSkill
- 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
- Verification DisciplineSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 4 are distinctive, and 0 are 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.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 46
Enumerate EVERY surface that could restate it: before grepping one: this doc, sibling docs in the same tree, other repos, the public …
- Distinctive: line 50
Grep all of them for the tokens: Every surface on the list, not just the file in hand.
- Distinctive: line 62
No surface still asserts the old claim un-bannered — verified by a final re-grep, not by memory.
- Distinctive: line 76
If the tokens don't isolate the claim, that's the first thing to fix.
≈ “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
Correction Propagation
Purpose
The pack can MAKE claims (verification-discipline) and ATTACK them (adversarial-verify),
but nothing owns RETRACTING one after it has spread. A claim rarely lives in one place: the
same number, verdict, or headline gets restated in sibling docs, other repos, the public
mirror, READMEs, indexes, issue trackers, and memory files. When that claim is corrected —
downgraded, superseded, reversed, disproved — fixing the doc you happened to be looking at
leaves every other copy asserting the old, now-wrong version. Readers land on the stale one
and act on it.
This skill makes correction a sweep, not a spot-edit: name the claim's distinctive greppable tokens, enumerate every surface that could restate it, grep them all, banner or supersede each hit top-of-file first and dated, then re-grep to prove the sweep is clean. Only then is the correction done. It is the difference between "I fixed the README" and "the string '3/3' now appears in three repos, each bannered or deleted, re-grep returns only bannered hits."
When to use this skill
- You just corrected, retracted, downgraded, or reversed a claim you had published.
- A metric changed value, a status flipped (confirmed → disproved), or a conclusion reversed.
- A benchmark, number, or verdict you're fixing is restated somewhere else too.
- Someone asks "did the correction propagate?" or "is this still cited anywhere?"
- Closeout of any session where a correction happened.
- A claim that appears in more than one repo, doc, mirror, or memory file.
When NOT to use
- The claim never left one file and nothing restates it — a single edit is the whole job.
- You're making or grading a fresh claim, not changing one already out — that's
verification-discipline/empirical-validation. - The change is additive (new claim) and contradicts nothing previously published.
- Pure session-context staleness with no cross-artifact surfaces — that's
memory-hygiene.
The procedure
- Name the claim's distinctive, greppable tokens. The number, the exact phrase, the label — "3/3", "9/12", "replicated", "confirmed", the strategy name. Pick tokens specific enough to grep without drowning in false hits; too vague to grep = can't sweep it.
- Enumerate EVERY surface that could restate it before grepping one: this doc, sibling docs in the same tree, other repos, the public mirror/clone, READMEs, overview/index files, issue trackers, changelogs, and memory files. Write the list — the miss is always a surface you didn't think to check, not a token that didn't match.
- Grep all of them for the tokens. Every surface on the list, not just the file in hand. Include the public mirror path when the claim crossed the public boundary.
- Banner, supersede, or delete each hit — top-of-file first, dated. A reader who lands mid-doc must see the correction, so the fix goes at the top (or wherever leads), not only at the buried line. State the old claim, the correction, and the date.
- Verify the sweep. Re-grep the tokens across all surfaces; the only hits left are the ones you intentionally bannered or kept. An un-bannered hit = the sweep isn't done.
- Only then is the correction done. "Fixed the one I was looking at" is not a corrected claim; a clean re-grep across all surfaces is.
Quality bar
- No surface still asserts the old claim un-bannered — verified by a final re-grep, not by memory.
- The correction is visible top-of-file (or wherever the doc leads), not buried below stale text.
- The sweep covered other repos, the mirror, indexes, and memory files — not just the file in hand.
- Each banner is dated and states what changed, so a later reader can trust it.
Common failure modes
- Corrected the one I was looking at: the classic — fix the doc in focus, leave every sibling copy stale. The trigger for this skill is precisely "there is more than one copy."
- Buried correction: the fix lands at line 31 but line 7 still leads with the wrong claim. A mid-doc reader never scrolls to the retraction. Fix top-of-file first.
- Cross-repo blindness: repo A is fixed, repo B still cites the retracted number. The public mirror is the most common blind spot — it's a separate surface.
- Token too vague to grep: the claim can't be searched, so the sweep silently misses copies. If the tokens don't isolate the claim, that's the first thing to fix.
Works with sibling skills
verification-discipline labels a claim's evidence level WHEN it's made; this skill propagates
a CHANGE to a claim already published — one owns the birth, the other owns the retraction.
memory-hygiene handles session-context-vs-observation staleness; this handles cross-artifact,
cross-repo surfaces that outlive the session. session-orientation's closeout owns updating the
map; this is a closeout duty whenever a correction happened. publish-hygiene guards leaks and
metadata across the public boundary; this guards claim consistency across that same boundary.
Provenance and maintenance
Added 2026-07-09 after a cross-repo audit found a disproven benchmark claim — "only the always-on rules flagged it 3/3" — corrected in one public repo (rules-with-receipts) but still asserted in a sibling public repo (agent-failure-modes, AFM-3), and an internal AB-SNIPPET doc whose title still read "confirmed" after the result was disproved; separately, an analysis's published verdict was never amended after its own successor commit reversed it. In each case the correction was made in one place and never propagated. Re-verify by grepping a corrected claim's distinctive tokens across all repos after any correction — a hit that isn't bannered is this skill's trigger having failed.
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.

