What are you trying to do?
Understand an existing product · Map a flow · Prove the result · Recover drifting work
Skills
- Skills I publish. My collection on skills.sh.
- External skills. Agent coding methods, UI skills, and tools from other authors.
Prompts
-
Ask for proof of work instead of a summary. Ask a worker for evidence you can inspect.
-
Find implicit state machines with Pstack. Model each subsystem and chart its key flows.
-
Review untested branches. Decide which checks need tests and which states the types can rule out.
-
Could a human working for free do it? Test what it would take to hand off the work.
-
Restate the goal. Check that the agent understood your goal before it starts.
-
Find a better method. Reconsider the method you assumed the work required.
-
Design a system to do the work. Reframe a task as a system to build.
-
What is worth building now? Look for work that became practical as the cost changed.
-
Deeper without bigger. Ask what better practice would add to the same task.
-
The best next question. Get the question that would move the work forward.
-
The same with half the code. Ask for a smaller implementation of the same thing.
-
Stay at the frontier. Resolve the decisions you can see now.
Playbooks
- Describe how an existing product behaves. Trace what people experience when a flow is interrupted.
- Take a breadboard from sketch to proof. Prove an upload survives reopening.
- Prove the thing the user can actually do. Find where to act and what would prove it worked.
- Get an agent unstuck when it keeps missing the point. Find the first miss and choose a repair.
- Work in independently valuable steps. Use each delivery to pick the next step.
- Turn repeatable agent work into a script. Decide when the steps are ready to automate.
- Give the agent a way to verify its work. Fast checks for flows and invariants.
- Give a long task a worklog. Keep decisions and open questions from the chat.
- Coordinate work in Codex. Split work across agents and chats, and keep a shared worklog.
- Intent specialists. Six reusable agent roles and when to use them.
- Automate agreed rules and surface open decisions. Bring open contract questions to review.
Field reports
- Why I turned off automatic agent memory. Stopped. Four tasks with and without access to past lessons.
- AgentsView learning. Prob nah. Search worked. Raw logs caught failures.
- T3 Code with Opus and DeepSeek. In progress. Opus directs DeepSeek workers.
References
- Four Dark Corners of Requirements Engineering. From specification to outcome.
- Towards Self-Driving Codebases. Give agents a way to run and test their changes.
- The Topology of an LLM Agent Loop With Tool Use. Cost and access across agent setups.
- The Case Against Formal Verification, 50 Years Later. What proofs leave open about the spec.