Agent learning layer triage
Skill markoblogo/abvx-agent-skills/skills/agent-learning-layer-triage
Reviewable capability layer for coding agents. Portable skills, delivery gates, workflow patterns, and verification-first engineering.
npx -y skills add markoblogo/abvx-agent-skills --skill agent-learning-layer-triageAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 4 stars4 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.
What its author says it does
Copied from the file, not written here
Route reusable agent learning after an error, success pattern, review finding, or repeated workflow. Use when deciding whether an improvement belongs in the model prompt, memory/context note, durable docs, SKILL.md, checklist, script/tool, eval, golden fixture, or should be rejected as overfit.
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
4.7 KB, as published. Nobody here has run it
Agent Learning Layer Triage
Use this skill when a task produces a lesson and the next question is: where should this learning live?
ABVX does not assume the model itself learns. Most useful operational learning belongs in an auditable layer around the model: context, skills, scripts, gates, and evals.
Learning Layers
Classify the candidate into the cheapest durable layer that solves the repeated problem:
- Leave as prompt/session behavior: one-off correction, low recurrence, no durable value.
- Memory or context note: stable preference, repo fact, operator convention, or short reminder that should be easy to read and revise.
- Durable repo doc: stable repo architecture, setup, verification, deployment, or workflow fact that future agents must discover reliably.
- Checklist: repeated human or agent sequence where order matters but no portable behavior layer is needed yet.
- SKILL.md: reusable behavior with a clear trigger, action rules, anti-patterns, and verification discipline.
- Script or tool: deterministic repeated step where code is more reliable than prose.
- Eval or golden fixture: behavior must be regression-tested, especially after a bug, review failure, or model drift.
- Reject / park: plausible but overfit, too broad, not validated, duplicative, or unsafe.
Prefer the lowest layer that makes the next run materially better.
Triage Questions
Ask in order:
- Did this happen more than once, or is recurrence likely?
- Is the lesson stable across repos, or only local to one repo/session?
- Is the value factual recall, procedural behavior, deterministic execution, or regression detection?
- Can the lesson be verified by a command, fixture, review rubric, or before/after trace?
- Would adding this to always-loaded context increase startup cost more than it reduces future mistakes?
- Does it duplicate a stronger existing skill, script, doc, or higher-priority instruction?
- Could it weaken a safety, trust, authorization, privacy, or destructive-action boundary?
If uncertainty remains, park as a note with evidence instead of promoting it.
Routing Rules
- Use memory/context note for preferences, compact conventions, and stable facts that do not need a workflow.
- Use durable docs for repo-local facts that should survive sessions and be discoverable from the repo.
- Use checklists when sequence is the main value and the steps remain partly human-supervised.
- Use SKILL.md when the behavior should load on demand across tasks and needs trigger discipline.
- Use script/tool when prose keeps producing inconsistent execution for a deterministic operation.
- Use eval/golden fixture when the lesson came from a failure that should not regress.
- Use rejected buffer when the idea is attractive but not yet proven.
Output Shape
Return a compact decision record:
Learning candidate: <one sentence>
Evidence: <trace, PR, review, failure, success pattern>
Chosen layer: prompt | memory | durable-doc | checklist | skill | script | eval | reject
Why this layer: <short rationale>
Artifact to update: <path or destination>
Verification: <how we know the learning helps>
Rejected higher layers: <why not skill/script/eval/etc.>
Next action: <one concrete edit or no-op>
Promotion Guardrails
- Do not turn every useful sentence into a skill.
- Do not put repo-local facts into global skills unless the behavior generalizes.
- Do not create scripts for workflows that still require judgment at every step.
- Do not create evals without a stable fixture or observable pass/fail condition.
- Do not store secrets, private user data, credential hints, or sensitive client context in publishable artifacts.
- Do not weaken existing authorization, safety, or verification gates to make learning feel smoother.
Pair With
skillopt-evolve-skillswhen the chosen layer isSKILL.mdor agent instructions.durable-context-maintenancewhen the chosen layer is repo-local durable docs.goal-loop-designerwhen the lesson should become a bounded loop contract, judge rubric, or stop rule.delivery-preflight-gatewhen repeated failure suggests a missing preflight, push gate, or regression fixture.
Final Report
State the chosen layer, the artifact changed or intentionally not changed, and the verification or evidence threshold for promotion.