full-cycle

Full cycle

Carry work from an idea to a PR a stranger can approve. Spend tokens freely: fan out subagents in dynamic workflows wherever breadth, independent opinions or adversarial checks make the answer better. The human spends attention: they steer and correct, the agent does the deep work.

Five words carry this skill. Tier list: every survey ends in a ranked list with one plain reason per row. Recommended: every question to the human carries the agent's own pick. Receipts: every number was really measured; what was not run is said. Skeptic: every finding is reproduced by a second agent before anyone acts on it. Dry: review ends only when rounds stop producing real findings.

Asking the human

Research until you could make the call yourself, then hand the human the call. Each ask lays out what they need to decide and nothing more: the options, what each costs and gains, what it touches. Mark your pick (recommended) with one plain reason. The human confirms, picks another, or redirects.

The human switches context all day, and a thread may span weeks. Each message stands alone: answer first, plain words, one picture where a picture is faster than a paragraph. A Mermaid map for structure and flow, a motion schematic for what changes over time (pr-issue skill, motion.md), a screenshot or slowed real recording for what a user sees. Style is owned by the agent-to-human skill.

Peers

From what the product is, name its peers: open-source apps and libraries that face the same problem. For a coding-agent client that means other agent UIs, editors and CLIs. Read their code, not their marketing. Every research stage answers "how did others solve this problem?" with file paths as evidence.

Stages

Gates marked ask wait for the human. Everything else runs unattended.

1. Survey. A workflow fans out across the problem space — for "optimize the app": rendering, startup, bundle, network, memory, server latency. Each agent measures or reads, not guesses, and checks the peers. Return a tier list of areas. Ask which to take.
2. Deep dive. Inside the chosen area, build or reuse a repeatable measurement, take a baseline, then return a tier list of concrete targets with their measured cost. A target whose cost is an estimate gets a probe: a throwaway fix, measured, reverted. Probes routinely refute the obvious suspect; rank by probe results.
3. Options. For the top target, 2–4 fixes, always including the answer to "is there a stupid simple solution we're not thinking about?" and the peers' approach. Prototype the serious ones in isolated worktrees and measure them interleaved. For each: upside, downside, what else it touches, and a picture of how the system changes. Ask which to build.
4. Pre-review. Before coding, one agent attacks the chosen design with "what have we missed?" and returns the final brief.
5. Implement in its own worktree and branch, following the repo's AGENTS.md/CLAUDE.md and owning specs.
6. Review until dry. See below.
7. Evidence. Rebuild before measuring; stale builds lie. Before and after run interleaved, five or more rounds, median with spread. Render or call counts beat milliseconds when timings are noisy. Run the full test suite once. Record media: slow a real recording until the difference is visible, and add a capture-only on-screen marker (a frame counter, a timestamp) when the change is about time.
8. PR. Write it with the pr-issue skill: before/after table, the options rejected and why, future work, visuals. Slice into one-problem PRs, stacked when they build on each other. Show the human the local media paths before any PR body changes. Ask before opening anything on a repo the human does not own.