Conciseness

'--corrections narrows to user turns that look like a correction. --full keeps long messages intact.' — every flag documented in a single line, commands given bare, no explanations of concepts Claude already knows (no preamble on OpenCode or git rev-parse). The two framing sentences ('Past corrections are a materialized process.') establish the skill's premise in two lines; every token earns its place.

Actionability

Two copy-paste-ready commands ('python3 .agents/skills/mine-agent-transcripts/scan.py --days 30'), a documented env var ('Set OPENCODE_DB to read another OpenCode database'), a specified output format ('Write one line per cluster with links'), and a concrete handoff target ('Pass the cluster list to harden-environment'). Not 4 because the commands cover the common cases fully as written.

Workflow Clarity

Clear three-step sequence (read history, cluster, hand off) with an explicit quality gate ('A cluster needs at least two turns and the same underlying shape') and a fallback branch ('If no constraint is possible, the cluster is a judgment call'). Not 5 because there is no error-recovery or edge-case guidance (e.g. what to do when the scan returns nothing, or how to choose --days 30 vs 90); the read-only scan means the destructive/batch validation cap does not apply.

Progressive Disclosure

No bundle files exist in references/, scripts/, or assets/, so the score reflects the body itself: a ~50-line, cleanly sectioned document (Read the history / Cluster, do not fix / Hand off the clusters / Rules) whose single external dependency (scan.py) is referenced by full path, one level deep, inside a code block — well-signaled and easy to navigate.

Specificity

Names the domain ('past OpenCode sessions in this repository for recurring corrections') and one concrete action ('mine'), matching 'names domain and 1-2 concrete actions, but not comprehensive'. Not 4 because it does not list several specific actions (scan, cluster, hand off are all invisible), and not 2 because the action and target are concrete rather than generic.

Completeness

Explicitly answers 'when' with three concrete trigger scenarios ('Use when the user asks what keeps going wrong, after a session where the agent needed repeated correction, or before hardening the environment') and 'what' with 'to mine past OpenCode sessions in this repository for recurring corrections'. Not 4 because the 'when' clause is already explicit and multi-scenario, not vague or implied.

Trigger Term Quality

'what keeps going wrong', 'needed repeated correction', and 'before hardening the environment' are natural phrases a user would actually say. Missing common variations like 'keeps making the same mistake' or 'learn from past sessions', so good coverage with a few natural terms missing rather than comprehensive.

Distinctiveness Conflict Risk

Clear niche (mining OpenCode session transcripts for recurring corrections) with triggers unlikely to fire for unrelated skills. The only adjacency is the deliberate handoff to 'harden-environment', which is a workflow relationship, not a trigger conflict.