dan-myles/skills

Two agents that run your software work inside T3 Code.
**Bob** is the engineer. Hand him a goal and he picks the right playbook, learns the code, designs the change, builds it, proves it works on the real app, gets it reviewed by a second model family, opens the PR, and babysits it until it's ready to merge. He never merges without you.
**Frank** is Bob's chief of staff. He watches your T3 threads, GitHub, Linear, and Slack, triages what comes in, sends small confirmed bugs straight to Bob threads, lines up everything else for you to pick, and briefs you at 8:30 and 4:30. He never writes code, never merges, and never messages anyone on your behalf.
It's a remix of three things:
- **pstack** by Lauren Tan: the playbooks, the principles, and the engineering bar (one job, verified, no slop).
- **Matt Pocock's skills**: grilling, the round-by-round interview that settles every decision before complex work starts.
- **Custom workflows** built for T3 Code: cross-provider subagents, worktree threads, report-backs that wake the coordinator, scheduled briefs, PR watching, and replies that are drawn, not written.
What it feels like
Small changes go straight through. Bob only stops to ask when the work is complex and needs a call only you can make, and then he asks with option cards and his pick marked. Answer **Send it** to take his recommendation.
The cast
| Skill | What it's for |
|---|---|
| /bob | The front door. Sorts the task into a playbook and runs it to a merge-ready PR. |
| /frank | The chief of staff. setup, brief, sweep, status ("what are my bobs doing"), or just talk to him. |
| /bob-grill | Settles every open decision on complex work, round by round, with a recommended answer for each. |
| /bob-how, /bob-why | How code works, and why it ended up that way. Answers come back as interactive pages. |
| /bob-architect | Two designers on two model families, a judge, one merged design. |
| /bob-adversarial-review | Reviewers on different models try to break your diff. |
| /bob-no-comments | Comment Sicko, a reviewer who hates comments, strips the diff before review. The code a comment was excusing gets fixed instead. |
| /bob-blast-radius | What a change breaks outside its diff, with the safety claim proven by running code. |
| /bob-verify | Writes a repo's own bob-verify- skill so Bob can drive the real app and prove things work. |
| /bob-models | Pick the model for every role, checked against T3's live catalog. |
| /bob-improve | Bob learns from real sessions: reflect, eval a skill change, recommend models, make repeated mistakes impossible. |
| bob-principle-* | 24 short engineering principles Bob reads when a step needs one. |
**Bob's playbooks:** Investigation (explain or diagnose), Bug, Feature, Refactor, Perf, Prototype, Babysit, Orchestrate (a whole multi-day program), Cleanup.
Built for T3 Code
- **Every role has the right model.** Opus 5.5 runs Bob threads, GPT-6.1-Sol writes the code, Sonnet reads, and every check comes from the other model family than the work it checks. Set it all in one file.