Track E.3 — skill-from-session generator (BLOCKED on E.1) #29

New Issue

mathias · 2026-05-25T16:45:30Z

mathias commented

2026-05-25 16:45:30 +00:00

Part of mathias/infra Track E umbrella. Skills repo grows from real usage, not hand-curation.

Blocked on

hyperguild#NN E.1 — needs ingested Claude session corpus to mine.

Goal

New skill skill-from-session in mathias/skills repo (or new tool in hyperguild trainer skill). Mines recent Claude session corpus for patterns and proposes new SKILL.md drafts for human review.

Scope

Pattern detection

Over the last N days of ingested sessions, find:

Repeated tool sequences — e.g. grep → Read → Edit patterns that appear ≥5 times across sessions for the same task type. Candidate for a procedural skill.
Repeated user corrections — when Mathias says "no, that's wrong because X" → cluster by topic → candidate for a feedback skill or anti-pattern entry.
Repeated confirmations — "yes exactly" / "perfect, keep doing that" on non-obvious approaches → candidate for a positive-rule skill.
Repeated surprises — turns where Claude logs "unexpected: …" or where the brain has new failures hall entries → candidate for a watchout skill.

Output

Proposed SKILL.md draft in markdown, written to ~/.claude/projects/<host>/proposed-skills/ for human review. Skill metadata includes evidence: which sessions, which lines, how many occurrences.

Human reviews → either:

Accept → mv to ~/dev/skills/<name>/SKILL.md, commit, tag bump
Reject → delete
Edit → curate then accept

Files

hyperguild/internal/skills/skill-from-session/ (or reuse existing trainer/ skill scaffold)
New brain MCP tool propose_skill(window_days, min_occurrences) returns draft list

Acceptance

After E.1 has been running ≥2 weeks, propose_skill returns ≥3 candidate drafts
At least one candidate has evidence backing (linkable to session ids + lines)
Drafts compile to valid SKILL.md with required frontmatter (name + description)

Pre-mortem

Hallucinated patterns: if the model generating the draft sees 3 instances of grep → Read → Edit and concludes "every grep needs a Read+Edit" that's wrong. Mitigation: require ≥5 occurrences and ≥2 distinct sessions; show evidence not just summary.
Skill bloat: 19 current skills will turn into 50 if any pattern becomes a skill. Mitigation: human-review gate is hard. Default-reject anything below threshold.

infra Track E umbrella
hyperguild#NN E.1 (blocks this)
mathias/skills repo (consumer of accepted drafts)

Part of mathias/infra Track E umbrella. Skills repo grows from real usage, not hand-curation. ## Blocked on hyperguild#NN E.1 — needs ingested Claude session corpus to mine. ## Goal New skill `skill-from-session` in `mathias/skills` repo (or new tool in hyperguild trainer skill). Mines recent Claude session corpus for patterns and proposes new `SKILL.md` drafts for human review. ## Scope ### Pattern detection Over the last N days of ingested sessions, find: - **Repeated tool sequences** — e.g. `grep → Read → Edit` patterns that appear ≥5 times across sessions for the same task type. Candidate for a procedural skill. - **Repeated user corrections** — when Mathias says "no, that's wrong because X" → cluster by topic → candidate for a feedback skill or anti-pattern entry. - **Repeated confirmations** — "yes exactly" / "perfect, keep doing that" on non-obvious approaches → candidate for a positive-rule skill. - **Repeated surprises** — turns where Claude logs "unexpected: …" or where the brain has new `failures` hall entries → candidate for a watchout skill. ### Output Proposed `SKILL.md` draft in markdown, written to `~/.claude/projects/<host>/proposed-skills/` for human review. Skill metadata includes evidence: which sessions, which lines, how many occurrences. Human reviews → either: - Accept → `mv` to `~/dev/skills/<name>/SKILL.md`, commit, tag bump - Reject → delete - Edit → curate then accept ### Files - `hyperguild/internal/skills/skill-from-session/` (or reuse existing `trainer/` skill scaffold) - New brain MCP tool `propose_skill(window_days, min_occurrences)` returns draft list ### Acceptance - After E.1 has been running ≥2 weeks, `propose_skill` returns ≥3 candidate drafts - At least one candidate has evidence backing (linkable to session ids + lines) - Drafts compile to valid SKILL.md with required frontmatter (name + description) ## Pre-mortem - **Hallucinated patterns**: if the model generating the draft sees 3 instances of `grep → Read → Edit` and concludes "every grep needs a Read+Edit" that's wrong. Mitigation: require ≥5 occurrences and ≥2 distinct sessions; show evidence not just summary. - **Skill bloat**: 19 current skills will turn into 50 if any pattern becomes a skill. Mitigation: human-review gate is hard. Default-reject anything below threshold. ## Related - infra Track E umbrella - hyperguild#NN E.1 (blocks this) - mathias/skills repo (consumer of accepted drafts)

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: mathias/hyperguild#29