Skill learner

Skill marcrabadan/praxis/.claude/skills/skill-learner

Turns a knowledge gap a praxis expert hits mid-task — a missing team convention, runbook, or domain rule — into durable skill knowledge via the skill-creator, always proposed through the memory ledger, never mutated silently. Use when the user says learn this, capture or remember how we do X, teach the system a convention, or update a skill with what you just learned. For building a skill from scratch or validating a SKILL.md, use the skill-creator instead.From its SKILL.md

Install

npx -y skills add marcrabadan/praxis --skill skill-learner

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 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

5.7 KB, as published. Nobody here has run it

Skill Learner

The learning loop of the factory. When an SDLC expert is working and hits a gap — it lacks this team's conventions, a runbook, or domain rules it should have applied (e.g. the devops-engineer is asked how the org builds infrastructure in Terraform or Azure and no convention is captured) — skill-learner turns that gap into durable, auditable skill knowledge.

It is a thin orchestration layer, not a second skill factory. It does three things and delegates the rest:

Detect that the gap is real and worth capturing.
Decide where the knowledge belongs (new skill, existing skill, or a reference inside a role expert) and at what tier/granularity.
Delegate the actual authoring to the skill-creator, then propose the result through the memory ledger as pending for the user to accept.

It never edits a skill silently. Propose, don't mutate.

Required reading

Before acting, read references/rules.md — the four rules that keep this loop safe. The new-skill-vs-reference granularity call is governed by ../../factory/ai/promotion-policy.md (promote on evidence, never on anticipation).

When to use

Trigger this skill when:

An expert was asked for org-specific knowledge it does not have — "how do we build infra in Terraform/Azure", "what's our deployment runbook" — and the answer should be remembered for next time.
The user says "learn this", "capture how we do X", "remember our convention", "teach the system", "you didn't know our setup".
An existing skill gave an answer that was wrong or incomplete for this org and should be refined with what was just learned.

When not to use

The user wants to build a skill from scratch or validate a SKILL.md — that is the skill-creator directly. skill-learner is for gaps discovered during work, and it hands off to the skill-creator.
The gap is public knowledge the model already has (plain Terraform syntax, what Azure is). A skill captures org-specific conventions and procedures, not facts the model can already produce. See references/rules.md, Rule 2.
The user just wants the answer once, with no intent to make it durable.

Workflows

Pick the mode that matches the situation and follow its file.

Mode	File	When
Detect & route	workflows/detect-gap.md	Entry point — confirm the gap is real and org-specific, search existing coverage, decide the route.
Refine existing	workflows/refine-existing.md	Coverage exists but is wrong/incomplete — hand to skill-creator's review flow.
Create new	workflows/create-new.md	No coverage — decide reference-vs-new-skill and tier, hand to skill-creator's create flow.
Promote governance	workflows/promote-governance.md	The insight is a standing constraint, not skill knowledge — turn a validated assumption or evidenced gap into a `pending` rule / gate / eval / guardrail via `tools/promote.py`.

Default path: detect-gap → (refine-existing | create-new). When the learning is a harness constraint rather than skill knowledge, use promote-governance.

References

references/rules.md — the four governing rules.
references/examples.md — worked examples (the Terraform/Azure gap, the language-granularity call).

Evals

evals/trigger-evals.json — positive and negative routing cases, with the near-misses centred on the boundary against skill-creator (build-from-scratch / review a SKILL.md) and against public knowledge that should be answered, not captured.

This skill carries evals/ purely as a routing safety net — it has a real collision risk with skill-creator and benefits from explicit negative cases. It stays Tier 3: it orchestrates and delegates, it does not generate, modify, or validate code, so the Tier 4 trigger in ../../factory/ai/skill-tiering.md does not apply.

Output expectations

A one-line statement of the gap and evidence it is org-specific, not public.
A routing decision: create new / refine existing / add reference, with the exact destination (which skill, which reference file).
A tier and granularity call, justified against the promotion policy.
A handoff to the skill-creator (which workflow, with what brief).
A memory-ledger entry logged as pending describing what will be learned.
Validator result for any touched skill.

Stop conditions

The skill is done when:

The gap has been classified as capturable (org-specific) or declined (public knowledge / one-off), with the reason stated.
A routing decision and destination exist.
The skill-creator handoff has been made (or explicitly deferred to the user).
A pending ledger entry records the proposed learning.
The validator passes on any skill that was created or refined.
Nothing was mutated without a corresponding pending proposal.