A1 evolve

Skill N3URAL-A1/a1-specforge/skills/a1-evolve

Self-optimization engine for the a1 skill set. Reads accumulated observations (`.a1/phases/*/observations.jsonl`) and per-skill _learning.md files plus the learning store's `pattern/a1-learnings/*.md` (repo-local `.a1/learnings/` by default; set `A1_VAULT_ROOT` to use an external vault, e.g. Obsidian), clusters recurring patterns, scores by impact (frequency × severity), and proposes concrete diff-level improvements to agent files and SKILL.md / workflow files. Four phases: Collect → Cluster → Propose → Apply (with user confirmation per diff). Confidence thresholds: 1-2 occurrences = monitor, 3-4 = propose, 5+ = high-confidence. Commits applied changes to the skills repo (detected via git). MUST trigger when the user says: "a1-evolve", "improve skills" (alias: "skills verbessern"), "evaluate learnings" (alias: "learnings auswerten"), "what can we improve" (alias: "was können wir verbessern"), "skill optimization", "review retros" (alias: "retros auswerten"), "synthesize learnings", "pattern analysis", "improve the a1 skills", "evolve the agents", or when a1-execute reports 5+ accumulated learnings since last synthesis. Scope: only edits files in the skills repo — never touches ~/.claude/rules/ (user-authored). Do NOT activate for: a single retro entry (just write it via the host skill's Retro block), planning new features (use a1-new-feature), or routine project execution.From its SKILL.md

Install

npx -y skills add N3URAL-A1/a1-specforge --skill a1-evolve

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

2 stars2 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

5.2 KB, as published. Nobody here has run it

a1-evolve — Self-Optimization Engine

Language: English-first; German trigger aliases supported.

Reads what went wrong and right across all a1 skill runs, then proposes precise, targeted improvements to skill and agent files.

Philosophy: Skills improve through use, not through theorizing. Every observation is a data point. Patterns that appear 3+ times across different projects are signal — not noise.

Input sources

This is the ONE canonical list of what a1-evolve consumes. It must match the description above. Any new learning store MUST be added here AND to the collect globs in workflows/01-collect.md in the same commit (framework invariant 4).

#	Source	What it contains
1	`~/.claude/skills/<skill>/_learning.md`	Per-skill _learning.md caches — per-run retros with structured observations
2	`.a1/phases/*/observations.jsonl`	Raw inline observations from agents during execution
3	Learning stores `~/code//.a1/learnings/pattern/a1-learnings/.md`	Canonical learning stores — synthesized per-skill retros, repo-local ONE PER PROJECT; collect reads ALL of them (glob over `~/code/*`), not just the current repo's. `A1_VAULT_ROOT` adds an optional external vault (e.g. Obsidian)
4	a1-fix postmortems	Learning store `wiki/postmortems/` (detail) + promoted lessons in `wiki/lessons/`, plus the normalized retros a1-fix appends to `pattern/a1-learnings/a1-fix.md`

Not an input: ~/.claude/skills/_shared/learnings-index.md is an OUTPUT cache written by this skill after synthesis — it is never read as an input source.

Learning-enabled skill set (honesty note)

"All skills" in the Philosophy line above means "all skills that write a Retro" — not literally every skill in the repo. Verified live against every skills/*/SKILL.md + workflows/*.md file (grep for a Retro heading), the learning-enabled set is 15: a1-execute, a1-analyze, a1-checklist, a1-new-feature, a1-new-project (originally documented group), a1-fix (postmortem-based — a distinct mechanism, see source #4 above, not a _learning.md Retro block), plus a1-plan, a1-roadmap, a1-reconcile, a1-modernize, a1-constitution (added M11-audit-fixes Wave 6, Task 6.1), plus a1-phantom, a1-pr-review, a1-worktree — all three carry a wired, mandatory Retro block despite being deterministic gates or read-only reporters. (a1-check was in this set until M13 retired it into a1-checklist's checks #9/#10; its retro history migrated into a1-checklist/_learning.md.) a1-evolve itself also retros (source of the 15th name above). Only a1-progress (read-only status scan, confirmed no _learning.md and no retro mechanism anywhere in its files) is genuinely excluded today. hero-animation-builder follows no a1 convention (no workflows/ dir) and was never a candidate (see Wave 7 decision doc).

Workflow

See workflows/01-collect.md → 02-cluster.md → 03-propose.md → 04-apply.md

Output

Pattern report — what's recurring, how often, what impact
Diff proposals — concrete edits to specific files with before/after
Applied changes — on user approval, commits to repo

Confidence thresholds

Occurrences	Action
1-2	Log in index, monitor
3-4	Propose improvement (user reviews)
5+	Propose with "high confidence" flag — suggest direct apply

Scope

a1-evolve only proposes changes to files in the skills repo (detected via git):

agents/*.md — agent instructions
*/SKILL.md — skill definitions
*/workflows/*.md — workflow steps

It never auto-modifies ~/.claude/rules/ — those are user-authored.