Skill · Operate and improve

steer-miner

Use when mining local agent transcripts the user points to. Find repeated human corrections and propose skills, lint rules, or constraints as a report. Never edit files.

Written by Kunanon Jarat

Original
You invoke
MIT
Superstacks v6.1.1

View on GitHub

Install this skill

Installs only steer-miner, globally, as a copy

DISABLE_TELEMETRY=1 npx skills add KunanonJ/superstacks -g -s steer-miner --copy -y

Third-party installer; DISABLE_TELEMETRY=1 turns its telemetry off.

Folder layout (steer-miner/SKILL.md). Unzip it into your agent’s skills directory.

Download steer-miner.zip

Flat layout: SKILL.md sits at the zip root, ready to upload as a skill.

Download steer-miner-flat.zip

Copy this folder from the repo at v6.1.1 into your agent’s skills directory

plugins/superstacks/skills/steer-miner/

steer-miner ships in the Superstacks plugin with the other skills. In Claude Code:

/plugin marketplace add KunanonJ/superstacks
/plugin install superstacks@superstacks

Run them one at a time. Codex, Cursor and ZIP steps are on the Skills page.

How it runs

You invoke it.

Its frontmatter sets disable-model-invocation: true, so the agent never loads it on its own. Ask for it by name.

Files

Folder on GitHub →

Origin

Original

Written for Superstacks

Written for Superstacks; sources.lock.json credits no upstream. Released under MIT with the rest of the pack.

License: MIT
Release: v6.1.1 · 2026-10-06

More in Operate and improve

Run tickets end to end and turn repeated mistakes into guardrails.

All 20 skills →

SKILL.md

View on GitHub

---
name: steer-miner
description: Use when mining local agent transcripts the user points to. Find repeated human corrections and propose skills, lint rules, or constraints as a report. Never edit files.
license: MIT
disable-model-invocation: true
---

# Steer miner

Read only the transcript files the user names. Find repeated human corrections. Propose environment changes. Stop.

## Do

1. Confirm the paths. If the user did not name files or a directory, stop and ask.
2. Run the miner:


```text
python <this-skill>/scripts/mine_transcripts.py --path <file-or-dir> [--min-count 2]
```

3. Read the report. Group hits into: new skill, lint/semgrep/ruff rule, `CONSTRAINTS.md` row, or ignore (one-off).
4. Return a report: pattern, examples (short), proposed guard, and which kind from `mistake-to-constraint`.

## Do not

- Do not edit files.
- Do not file tracker issues or send chat messages.
- Do not scan home directories or default transcript folders the user did not name.
- Do not paste secrets, tokens, or full transcripts into the report.