Optional intake · init / log
Organize complex requirements or establish device and source provenance for log analysis. Choose the entry that fits the task.
Open workflow for Claude Code · Codex · CodeBuddy
An AI coding workflow, from plan to proof.
Plan, review, build and verify in one resumable ticket.
- Claude Code
- Codex
- CodeBuddy
Site source version · v2.32.1
Six stages. A traceable path forward.
Set the boundaries. Build the change. Keep the reasons behind each step.
Explore stage inputs, outputs and checkpoints
- Plan
- Review
- Finalize
- Code
- Deep check
- Audit
01Planplan
Turn an idea into a reviewable implementation plan.
- Input
- Goals, constraints and existing code.
- Output
- 01_plan.md: scope, approach and verification plan.
- Checkpoint
- Make exclusions and risks explicit. A design-only request stops here.
02Reviewreview
Find design gaps before implementation.
- Input
- The plan, relevant source and interface contracts.
- Output
- 02_review.md: issues, risks and recommendations.
- Checkpoint
- Ground recommendations in the design or code, not impressions.
03Finalizemerge
Consolidate review feedback into an implementation basis.
- Input
- Plan, review findings and agreed trade-offs.
- Output
- 03_plan_final.md: the finalized design.
- Checkpoint
- Resolve or explicitly retain open questions and scope changes.
04Codecode
Implement the finalized plan with traceable changes.
- Input
- Final design and the current code baseline.
- Output
- Source changes, tests and actual execution records.
- Checkpoint
- Stay within authorized scope. A plan is not evidence of completion.
05Deep checkdeepcheck
Revisit implementation and related call chains for gaps.
- Input
- Source changes, design decisions and execution records.
- Output
- 05_deepcheck.md and coverage records.
- Checkpoint
- Check boundaries, regressions and missing evidence; fix and retest findings.
06Auditaudit
Check whether the delivery conclusion has supporting evidence.
- Input
- Design, code changes and available verification evidence.
- Output
- 06_audit.md: delivery conclusion and unverified items.
- Checkpoint
- Separate verified, pending and blocked. A report is not a device test.
Optional paths, not extra required stages
Independent verification · verify
Explicitly choose build, deploy, listen or test, independently or in combination. Listening or testing alone uses the installed version; it does not implicitly build or deploy.
Advanced optional review · crosscheck
Review a completed ticket independently. Switch agents or models yourself; repeated rounds stay separate from original history and code. No automatic model switching.
Capabilities
Choose your goal. Expand for commands and boundaries; examples are not execution evidence.
Develop & design
Ticket-based development, or just a design. Tell the agent both your goal and where to stop.
More examples & boundaries
Use ICODE to design retries for this module. Stop at the plan; do not change code./icode start Add modulo to the calculator and cover it with tests
/icode planstart follows the main workflow; plan can run independently. Design-only requests must stop before implementation.
Logs & diagnosis
Establish the version, time window and evidence source before analyzing an incident.
More examples & boundaries
Use ICODE to analyze this failure log. Find the cause without changing code./icode logCurrent source, deployed versions and historical logs are different evidence layers. Keep uncertainty explicit when field data is missing.
Design & code review
review examines the current design; crosscheck provides repeatable independent review of completed tickets.
More examples & boundaries
/icode review/icode crosscheckFor multi-model code review, switch agents or models yourself first. ICODE does not automatically switch models; crosscheck leaves original history and code unchanged.
Build & device checks
Choose the actions you need and keep records tied to actual execution.
More examples & boundaries
/icode verify --listen/icode verify --build
/icode verify --build --deploy --listen
/icode verify --deploy --test <target>Using --listen or --test alone checks the installed version. It does not implicitly build or deploy; deployment must be selected explicitly. Replace <target> with your project's test target.
Docs & reusable knowledge
Choose project documentation, Word, PPT or reusable knowledge to match your delivery needs.
More examples & boundaries
Use ICODE to write a newcomer-friendly guide for this project./icode doc
/icode docx
/icode ppt
/icode studySuccessful generation is not visual acceptance. A text-only model cannot inspect images directly; keep visual review limits explicit.
Manage & resume tickets
Check state and verification debt, or manage and resume tickets in the local workspace.
More examples & boundaries
/icode status --pending/icode list
/icode ui
Resume the bound ICODE ticket: check its state and execution receipts before continuing.This website does not connect to your local UI or read tickets. Resuming work retains the original scope and stop points.
How ICODE keeps work within its boundaries
Development that outlasts a conversation
Describe the goal
Name ICODE and state the goal and constraints. Your host agent routes to existing steps, preserving stop points and safety gates.
Ticket-based development you can resume
Plans, changes and review records stay with the ticket. Check its state, execution receipts and unfinished work before resuming the authorized scope in another session.
Multi-model code review
Switch agents or models yourself in your host, then run /icode crosscheck on a completed ticket. ICODE does not automatically switch models; repeatable review rounds stay separate from the original ticket history and code.
Verify with evidence
Explicitly combine build, deploy, listen and test, recording actual results and evidence gaps. Listening or testing alone does not implicitly build or deploy.
Manage tickets locally
The optional /icode ui opens a local workspace using existing steps and gates. No tickets need to be uploaded to this website.
Keep your existing host
Use Claude Code, Codex or CodeBuddy. Switch models between stages. Natural-language routing does not require the optional Runtime or a new model service.
Focused guides
AI coding workflow: from request to verification evidence
Use ICODE with Claude Code, Codex, CodeBuddy and WorkBuddy to run a resumable AI coding workflow with traceable design, implementation and verification evidence.
Multi-model code review for completed tickets
Switch agents or models yourself in Claude Code, Codex, CodeBuddy or WorkBuddy, then use ICODE crosscheck for an independent design and code review.
Deliver code with the reasons to trust it.
Design rationale
Code & tests
Verification records
Not verified
Generic calculator illustration, not a real ticket or test result. A report does not mean a device was verified.
What goes into these records?
Design rationale
Why modulo is needed, which inputs are allowed, and how review findings were handled. Plans stay with the ticket.
Code & tests
Which source changed, which boundaries were covered and which commands ran. Execution records match actual changes.
Verification records
Record build, deployment and test evidence separately, with actual versions and results, not just a completed label.
Not verified
Without access to the target device, mark it not verified. Deliver the remaining conditions and limitations alongside the code.
Install once. Start in your project.
System terminal · preview before installing
git clone https://github.com/ayukyo/icode-skill ~/icode-skill
cd ~/icode-skill
./install.sh --dry-run --client all
./install.sh --client allInstallation & usage notes
Bring a configured AI coding host and model access. Installation may download runtimes and optional MCP dependencies. all installs Claude Code and Codex, adding CodeBuddy only when its configuration exists; --client codebuddy selects it explicitly. The default is Claude only.
After installation · your host agent's project session
/icode help
Use ICODE to design retries for this module. Stop at the plan; do not change code.Installation & usage notes
Restart or reload skills in your host; select the skill explicitly or use /icode if needed. Use it only when you name ICODE or resume a bound ICODE ticket. It does not take over unrelated requests. These are host agent instructions, not shell commands.
Start with a sentence. Or choose a step.
Choose your goal. Expand for commands and boundaries; examples are not execution evidence.
Design only
Use ICODE to design retries for this module. Stop at the plan; do not change code.Full development flow
/icode start Add modulo to the calculator and cover it with testsListen to the installed version
/icode verify --listenReview after switching models yourself
/icode crosscheckFind verification debt
/icode status --pendingResume interrupted work
Resume the bound ICODE ticket: check its state and execution receipts, then continue from the interruption within the original scope and stop point.Open the local workspace
/icode uiGates and reports support verification; they do not guarantee defect-free code. Missing tools, unperformed checks and evidence gaps must stay explicit. Hosts, models and physical devices require their own validation.
Version notes
v2.32.1
Fix automatic installation, hash verification and persistent registration of the owned Word renderer. Public distributions require real rendering qualification. The current renderer recipe supports Ubuntu-compatible Linux x86_64; other platforms remain pending.
v2.32.0
Natural-language entry: name ICODE and describe your goal and constraints while keeping existing commands and gates. The date is this note's review date, not the release date.