Use this skill when the agent is about to implement a feature or work block. Defines the full cycle: decompose into tasks, present plan, execute with checkpoints, handle deviations (no-drift), verify A/C, and close with context/ and memory/. Applies whenever there is code to implement, even if the user doesn't explicitly mention "iteration" or "plan".
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Use this skill when the agent is about to implement a feature or work block. Defines the full cycle: decompose into tasks, present plan, execute with checkpoints, handle deviations (no-drift), verify A/C, and close with context/ and memory/. Applies whenever there is code to implement, even if the user doesn't explicitly mention "iteration" or "plan".
Iteration Rules — Task Execution
Iteration Flow
Decomposition → split feature into atomic tasks, order by dependency
Plan → present task list to the developer and request confirmation — do NOT implement without confirmation
Execution → per task: implement, write tests, checkpoint each significant block, update context/
Verification → verify each A/C individually, tests pass, no regressions
Closure → update memory/ with what was implemented, mark feature as completed in context/
If a deviation is detected during execution → apply no-drift rule (see section below).
1. Task Decomposition
Each task must be: completable in a continuous block, verifiable (clear output), independent or with explicit dependency, and describable in one sentence.
Order bottom-up: data layer → backend → middleware → frontend → integration → tests. Never start with UI if the backend doesn't exist.
2. Plan
Before touching code, present the plan to the developer:
## Implementation Plan — auth-login**Feature:** Login with email and password
**A/C to cover:** 9 criteria
### Tasks (10)
| # | Task | Depends on | Complexity |
|---|------|------------|------------|
| 1 | Users schema + migration | — | low |
| 2 | POST /api/auth/register | 1 | medium |
| 3 | POST /api/auth/login | 1 | medium |
| 4 | JWT Middleware | 3 | medium |
| 5 | RegisterForm + validation | — | medium |
| 6 | LoginForm + validation | — | medium |
| 7 | Route protection | 4 | low |
| 8 | Lockout after failed attempts | 3 | medium |
| 9 | Unit tests | 1-8 | medium |
| 10 | Integration tests | 1-8 | high |
### Proposed Technical Decisions- bcrypt for password hashing (standard, well supported)
- JWT in httpOnly cookie (more secure than localStorage)
- Zod for form validation (consistent with the project)
**Do you confirm this plan or are there adjustments?**
The agent does NOT start implementing without plan confirmation.
3. Execution
When starting each task, announce what will be done. When completing, confirm what was done. If there were changes from the plan, explain why.
Block checkpoints: after completing a logical group of tasks (e.g. entire backend), checkpoint with summary of what was done, A/C covered, and next block. Ask if there are adjustments.
Context updates: update .docs/context/iteration-YYYY-MM-DD.md at least every 2-3 tasks. Do NOT wait until everything is finished. Format: see skill docs-structure.
4. Verification and Closure
Pre-Closure Checklist
Before declaring a feature as completed:
Each A/C verified individually
Tests pass (unit + integration if applicable)
No regressions in existing functionality
Lint and type-check pass without errors
No TODOs or commented-out code left behind
New routes/endpoints protected if applicable
Errors handled (no empty catches or silent errors)
Closure
Upon completion: update memory/ with what was implemented (format: see skill docs-structure) and mark feature as completed in context/ with status 🟢.
No-Drift Rule
During implementation, if the agent detects a deviation, it MUST pause and report before continuing.
Drift types:
Scope creep — the task requires implementing something outside the feature → propose the minimum for the A/C, or create a derived feature
Missing dependency — something needed isn't configured → report the dependency, propose as additional task or mark A/C as blocked
Inconsistency — an A/C contradicts existing code → ask which is correct, do NOT assume existing code is right
Design decision — multiple valid options with no definition in rules/ → present options with trade-offs, let the developer decide
Report format:
⚠️ DRIFT DETECTED — Type: [scope/dependency/inconsistency/design]
Context: [what I was doing]
Problem: [what I found]
Impact: [what's affected if I continue or not]
Options:
A) [option] — [trade-off]
B) [option] — [trade-off]
Recommendation: [which and why]
Cross-Cutting Concerns
After implementing each significant block of code, walk through this checklist. For each applicable item, consult the corresponding skill and apply its rules to the newly created code. Only mark the task as completed when all applicable items are met.
i18n (user-visible text) — only if the project has i18n configured (i18n dependencies in package.json and existing translation files). If applicable: consult frontend/i18n-react-rules or frontend/i18n-nextjs-rules based on project type. Do NOT hardcode UI strings. If the project has no i18n, skip this item without reporting drift
If a feature has > 15 tasks or > 15 A/C → propose splitting into sub-features, each with its own feature.md in .docs/features/. If the developer prefers to keep it as one, group tasks into phases with more frequent checkpoints.
Gotchas
The agent tends to start coding without presenting the plan — always request confirmation first
The agent may resolve a deviation silently — all drift must be reported to the developer
Empty tests or trivial asserts (expect(true).toBe(true)) don't count as tests
Commented-out code as "placeholder" is not valid implementation — implement fully or don't implement
Don't wait until finishing everything to update context/ — do it every 2-3 tasks
Tasks too large ("implement the backend") are not atomic — decompose further
Tasks too small ("create file X") create unnecessary overhead
Declaring done without verifying each A/C individually is a frequent mistake
Memory is only for accomplished facts — never include TODOs or future plans
