https://github.com/drubinstein/terry
A collection of Claude Code skills for general software-engineering work — bug reproduction (jamie), root-cause debugging (sherlock), performance (brendan), empirical iteration (andrej), and agent-fleet orchestration (terry).
https://github.com/drubinstein/terry
agentic-coding claude-code claude-code-plugin claude-code-skills debugging developer-tools
Last synced: 5 days ago
JSON representation
A collection of Claude Code skills for general software-engineering work — bug reproduction (jamie), root-cause debugging (sherlock), performance (brendan), empirical iteration (andrej), and agent-fleet orchestration (terry).
- Host: GitHub
- URL: https://github.com/drubinstein/terry
- Owner: drubinstein
- License: mit
- Created: 2026-06-02T12:46:06.000Z (24 days ago)
- Default Branch: main
- Last Pushed: 2026-06-14T21:39:15.000Z (12 days ago)
- Last Synced: 2026-06-14T23:15:17.667Z (12 days ago)
- Topics: agentic-coding, claude-code, claude-code-plugin, claude-code-skills, debugging, developer-tools
- Language: Python
- Homepage:
- Size: 139 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Claude Code skills for software engineering
[](LICENSE)
A small collection of Claude Code **skills** — reusable, named methodologies for
everyday engineering work. Each is a self-contained playbook (a `SKILL.md` plus
optional references and templates) that Claude loads on demand when your request
matches it.
## Why this exists
Claude writes code well but, under pressure, tends to *guess* — skip reproducing a
bug, theorize before measuring, dismiss a failure as "flaky," or call something
done without evidence. Each skill here encodes a disciplined methodology that
resists exactly that, distilled from real long-running agent work. You invoke one
by intent — "does this bug actually reproduce?", "find the bottleneck" — and
Claude follows the matching playbook.
## The skills
| Skill | What it does |
|-------|--------------|
| **[jamie](skills/jamie/SKILL.md)** | Confirms a bug report actually reproduces before you fix it. |
| **[sherlock](skills/sherlock/SKILL.md)** | Finds the cause of a stubborn bug by elimination. |
| **[brendan](skills/brendan/SKILL.md)** | Finds the bottleneck when something is slow (Brendan Gregg's USE method). |
| **[andrej](skills/andrej/SKILL.md)** | Drives an iterative fix or tuning loop, gated on a metric. |
| **[terry](skills/terry/SKILL.md)** | Runs many long-running agent tasks as a paced queue — parallelized and merged, with concurrency capped to avoid OOM and pace token spend. |
| **[ronald](skills/ronald/SKILL.md)** | Runs a controlled A/B experiment to choose a model or prompt. |
Most stand alone, but five compose into a debugging-to-fix pipeline: **jamie**
(confirm it's real) → **sherlock** (find the cause) → **brendan** (find the
bottleneck) → **andrej** (drive the fix, gated on a metric) → **terry** (run many
at once, paced to your machine and budget). **ronald** stands apart — a controlled
experiment for picking a model or prompt with evidence. See each skill's
`SKILL.md` for the full method.
## Example
You paste a bug report: *"CSV export corrupts my file every time."* Claude recognizes
a triage task and loads **jamie** — which stops it from closing the issue after one
green run on the latest build. Instead it matches the reporter's version and file,
reproduces the crash at their data scale, and minimizes it to a one-command repro:
**CONFIRMED**. It hands that repro to **sherlock** to eliminate suspects down to the
root cause, then to **andrej** to drive the fix, gated on the repro until it passes.
## Using these skills
**As a plugin** — point your Claude Code plugin config at this repo (or a marketplace
entry that references it). The manifest at `.claude-plugin/plugin.json` auto-discovers
every `skills/*/SKILL.md`.
**As personal skills** — copy any skill directory into your user skills folder:
```bash
cp -R skills/sherlock ~/.claude/skills/sherlock # or any other skill
```
Either way, Claude auto-discovers them and loads the right one when your request
matches its description — or invoke one explicitly with the Skill tool.
## Layout
Each skill is a directory under `skills/`:
```
skills//
SKILL.md # the methodology — what it's for and how to run it
references/ # deeper detail, loaded on demand (optional)
templates/ # copy-paste worksheets / scripts (optional)
```
The repo is named `terry` for historical reasons — it began as that single
agent-fleet skill — but it's now a general collection, and `terry` is just one
skill in it.
## Contributing
Each skill follows the same shape — a tight `SKILL.md` with depth in `references/`
and copy-paste assets in `templates/` — and is built test-first: write a failing
scenario for a subagent, write the skill, then close the loopholes it finds. New
skills and fixes are welcome via pull request or
[issues](https://github.com/drubinstein/terry/issues).
## License
[MIT](LICENSE) © David Rubinstein