The Smoother AI Codes, the More Business Needs Control Points: Five-Step Claude Edition, Retro Loop, and Project Health Check

The smoother AI codes, the more business needs control points

The change Claude Code brings is real: a requirement that used to take two weeks of scheduling now gets a working draft in minutes. But the smoother it gets, the more you hit a new problem — code output is accelerating while the business owner's sense of control is dropping. The road from requirement to code keeps shrinking, and nobody stops mid-way to confirm "this is actually what I asked for."

Our answer is not to shackle the AI. It is to place a few control points in the flow: where the business owner decides, they decide; where evidence is due, evidence is produced; everything else runs free. That idea is now packaged in three new skills, published this week to the 365Skill public repo: 365-five-step-dev-claude (graded five-step development), 365-five-step-retro-claude (five-step retrospective), and project-health-check (cross-project health sweep).

Five-step, Claude edition: control points at Plan and Verify

365-five-step-dev-claude serves a specific audience: developers who understand the business but do not read code line by line. The skill states it in its opening lines — your control points are Plan (business approval) and Verify (evidence-based acceptance), not the code itself.

Its core mechanism is task grading. Every request is classified against a table before work starts:

• Grade A (trivial): copy, styling, one obvious small bug, no database/API/permission changes — locate, patch minimally, verify in the page; the full five steps are skipped;
• Grade B (standard): new pages, forms, lists, ordinary APIs — all five steps, with Review dispatched to a sub-agent;
• Grade C (high-risk): database schema, permission systems, payments and orders, core module refactors, production deploys — all five steps plus business control-point checks, independent review, and human approval.

When in doubt, grade up. The value of this table: rule intensity follows risk instead of applying one blanket to everything. Low-risk changes are not buried in process; high-risk changes cannot bypass the gate.

One more design worth mentioning is the model-tier gate: if your environment runs tiered models, a Grade B task landing on a premium model stops and asks to downgrade — premium models are reserved for Grade C independent review rather than burned on ordinary pages. The mechanism is optional; environments without tiered-model config simply skip it.

The retro loop: iterate the process on real logs, never on impressions

Once a process is live, who watches whether it drifts? That is 365-five-step-retro-claude's job. It gives the five-step flow its metabolism: read three inputs — the run log (logbook), the current algorithm text, and the conclusion summary of the latest archived retro — then walk a fixed six-dimension analysis: grading quality (graded A but reworked repeatedly = graded too low), checkpoint friction, tier-switch adoption rate, Verify evidence completeness, log completeness, and the basis for rule changes.

The most interesting part is its forensics. The logbook may claim everything was recorded, but the retro cross-checks against git evidence — scanning recent AI-collaboration commits in active projects and matching dates and topics against log entries. A commit without a log entry is a missed recording. An absent entry does not testify for itself; that turns "process discipline" from a slogan into an auditable fact.

The output of a retro is not impressions — it is an approved process upgrade. No rule changes on gut feeling; every change must be backed by real operational data.

Project health check: one page of judgment for unclaimed state

The third skill attacks a different kind of mess: unclaimed state problems scattered across repos — CI red streaks, stale PRs, unpushed commits, orphan worktrees. project-health-check collects them into a single readable report, strictly read-only: any fix is a new development request, graded and routed through the five-step flow.

What separates it from a naive script is its false-positive guards:

• CI red is escalated only after 3+ consecutive failures, and before reporting you compare timestamps against the latest related merge — reds older than the fix are flagged "probably fixed, confirm on next run" rather than reported as failures;
• PRs with an explicit WIP/continuation note in the body count as "deliberately paused," not abandoned;
• worktree ancestry is merged before reporting (a directory hanging off another repo is a worktree, not an independent repo);
• activity is judged by the HEAD commit date of that directory, so worktrees sharing an object database are not misread as active;
• skipped repos must be named in the report with the reason — silently skipping is worse than not checking.

The skill is early experimental — just released, with one full smoke-tested cycle in a single real dev environment. We published it precisely because state-sweep tools are too easily built as hand-wavy scripts, and its false-positive judgment rules need more real-world data to mature.

Together they form a loop

These three skills are not isolated: the five-step flow runs daily execution, the retro iterates the flow itself on real data, and the health check finds who left work unfinished and routes it back into the five-step flow as a new request. Execute → retro → sweep → execute again. Governance starts turning.

They share one design stance: in the age of strong models, put minimal gates only where models are natively weak — obeying rushed instructions, skipping verification, cutting process corners. Rules are easier to delete than to add, and every rule has a cost.

Links

• 365Skill public repo (Apache-2.0, stars welcome): https://github.com/laojin1900/365Skill
• 365-five-step-dev-claude (graded five-step development, Claude Code edition): https://github.com/laojin1900/365Skill/tree/main/skills/365-five-step-dev-claude
• 365-five-step-retro-claude (five-step retrospective): https://github.com/laojin1900/365Skill/tree/main/skills/365-five-step-retro-claude
• project-health-check (cross-project health sweep): https://github.com/laojin1900/365Skill/tree/main/skills/project-health-check

Install in one line: clone the repo and run ./install.sh 365-five-step-dev-claude — the script detects Claude Code, Codex and Kimi on your machine and links the skill in.