Reminders curator
Mine every AGENTS.md and CLAUDE.md across local repos for guidance that repeats, then encode each repeated rule as an event-triggered reminder in the ezzs-way-of-coding hook so it stops being copy-pasted into every repo. Use whenever the user wants to curate or consolidate reminders, deduplicate AGENTS.md/CLAUDE.md guidance, "remove repetitiveness" across agent files, turn agent-file rules into hooks, or asks why the same instruction lives in several AGENTS.md files — even if they only say "my agent files keep repeating themselves".From its SKILL.md
npx -y skills add ezzabuzaid/agent-plugin --skill reminders-curatorAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.0 KB, as published. Nobody here has run it
Reminders Curator
Guidance duplicated across AGENTS.md/CLAUDE.md files costs context in every session of every repo, drifts as copies get edited independently, and still misses new repos. A hook reminder fires once, event-scoped, everywhere — the right home for any rule that has proven it generalizes by appearing in two or more places.
Locations
- Hook + rules:
~/Desktop/January/agent-plugin/hooks/ezzs-way-of-coding/(rules live inrules/<theme>.tsfiles — evidence, change-safety, testing, collaboration, frontend, infra, failures, defaults — assembled byrules/index.ts;rules.test.tsis the black-box suite; both Claude and Codex executehook.tson their hook events) - Explorer CLI:
~/Desktop/January/agent-plugin/tools/explore-agents-md/cli.ts - Predicate reference: the package types at
node_modules/@deepagents/experimental/dist/coding-agent-reminders/types.d.tsandpredicates/*.d.tsin that folder — read them before writing awhen
Workflow
1. Inventory
node ~/Desktop/January/agent-plugin/tools/explore-agents-md/cli.ts --json
Returns every folder holding a CLAUDE.md/AGENTS.md with paths, sizes, and
modified dates. Narrow with repeatable --root <dir> when the user scopes the
sweep.
2. Read and extract
Read each discovered file. With more than ~10 files, fan out subagents —
each reads a subset and returns candidate rules as
{ rule: <one-sentence normalized statement>, file, quote }. Normalize
aggressively: "never use jest" and "we strictly use the node test runner" are
the same rule.
3. Cluster
A rule qualifies for migration only when ALL of these hold:
- It appears (semantically, not verbatim) in 2+ files.
- It is a behavior rule — how to work — not a repo fact. Repo-specific paths, commands, ports, schema names, and domain glossaries stay in their files; a reminder that fires in every repo must be true in every repo.
- It is not safety-critical. "Never push to main" style prohibitions must stay always-loaded in the files; reminders are contextual and may not fire when it matters.
Report the clusters to the user before encoding: rule, where it appears, proposed trigger.
4. Encode
Add one ReminderRule per cluster to the matching rules/<theme>.ts file
(create a new theme file plus one rules/index.ts line when no theme fits):
- target/events — prompt rules (
UserPromptSubmit) for "when asked to X" guidance; tool rules (PostToolUse/PostToolUseFailure) for "when Y happens" guidance; session defaults only for the rare rule that applies to every turn. - when — compose the package predicates (
contentPattern,toolCall,and/or/not). Patterns must be precise enough that the reminder feels earned when it fires: require two independent signals (verb AND object) the way the existingpackage-api-firstrule does. - False-positive discipline — never grep free content for failure-ish or
topic-ish words alone; a successful file read that merely mentions the
topic will trigger it. Anchor on structured signals (event names, tool
names,
toolCall({ state })) or CLI-authored prefixes, the waybatchHasFailedCallinrules/failures.tsmatches^Error:per result instead of grepping the joined batch text. That predicate exists because the word-grep version fired on every file that discussed errors. - message — short, imperative, and carrying the why, so the agent can apply the rule to situations the pattern authors never saw.
5. Test
Every new rule gets black-box tests in rules.test.ts through
evaluateReminderHook (no lifecycle hooks; self-contained AAA): one prompt
that must fire it, and one near-miss that must not — the near-miss is what
keeps patterns honest. Then:
npm test --prefix ~/Desktop/January/agent-plugin/hooks/ezzs-way-of-coding
6. Trim the sources
For each migrated rule, propose — never auto-apply — removing the
now-redundant lines from each source AGENTS.md/CLAUDE.md, quoting every
removal verbatim so the user can judge and restore. Checked-in files are the
user's to approve per repo; apply only what they confirm, as ordinary
working-tree edits they review in git diff.
Output
End with a table: rule → files it was mined from → rule id and theme file → trigger summary → sources trimmed or pending. Remind the user that running sessions snapshot hook config at startup, so new reminders appear from the next session.