Your First Workflow
This guide follows a standard implementation cycle. For a standalone request to document existing behavior, follow Updating Documentation on Its Own.
This example adds search by name and email to an existing user directory and
updates its user guide. It uses the standard workflow with the default
AFTER_IMPLEMENTATION verification cadence.
It assumes an existing project set up for STANDARDS, with no active cycle. See Installation and Setup to install the runtime and role skills.
Use each role’s handoff to invoke the next one when it is ready. The examples below show the command for each client; use the one for yours.
1. Start with Auditor
Section titled “1. Start with Auditor”Give Auditor the request and choose standard work:
In Codex:
$auditor Start a STANDARD cycle.Add search by name and email to the existing user directory,and update the user guide to explain how to use it.In Claude Code:
/auditor Start a STANDARD cycle.Add search by name and email to the existing user directory,and update the user guide to explain how to use it.The agent validates the mode, generates a fresh cycle ID, and saves the request
before Auditor begins. Auditor then checks the relevant code, behavior,
conventions, and commands, saving its findings in .standards/CONTEXT.md.
This is the starting role for an existing project. For a new project, follow Starting a New Project.
2. Define the outcome with Scoper
Section titled “2. Define the outcome with Scoper”After Auditor finishes, use its handoff to run Scoper. Explain what users should be able to do and answer questions that would change the result.
The scope might include these acceptance conditions, if you agree to them:
AC-001: Users can find directory entries by name.AC-002: Users can find directory entries by email.AC-003: An empty result displays a clear no-results message.AC-004: The user guide explains how to search the directory.
The IDs connect each outcome to its design and later evidence. Scoper saves the scope and its location in the workflow record.
3. Design the change with Architect
Section titled “3. Design the change with Architect”Run Architect after Scoper’s handoff. It decides how the search should fit the existing application: which components handle the request, what they exchange, and how errors are handled.
Architect accounts for each acceptance condition. The user-guide condition may need no technical design; it still remains work for Documenter. Local coding choices stay with Developer.
4. Approve the plan and work with Developer
Section titled “4. Approve the plan and work with Developer”Developer turns the design into implementation steps and presents a saved plan
for your approval. A step might add the agreed name filter and link it to
AC-001.
Read the plan, request changes if needed, and approve it before coding starts. Choose how to work:
- AUTONOMOUS: the default; work through approved steps.
- STEPWISE: review progress one completed step at a time.
- CODE_WITH_ME: explain the next step and help with code you write.
Developer saves progress and runs implementation checks. Important changes to the plan require approval again. See Working with Developer for the details.
To test outcomes as you build them, choose incremental verification in any collaboration mode. Developer and Tester then exchange checkpoints before their full completion gates and implementation review.
5. Test in a separate chat
Section titled “5. Test in a separate chat”When Developer hands off to Tester, open a fresh chat separate from the implementation conversation and run:
Codex: $tester Continue from .standards/STATE.md.Claude Code: /tester Continue from .standards/STATE.md.Tester reads the saved requirements, design, implementation, and existing tests. It adds or updates coverage where needed, runs appropriate checks, and records the actual results and any gaps.
For this example, Tester checks the search behavior. The guide can remain pending until Documenter supplies its evidence. That does not excuse an untested or failing search condition.
See Tester for the test budget and what happens when a check cannot run.
6. Review the implementation independently
Section titled “6. Review the implementation independently”After Tester’s handoff, open a fresh chat separate from the conversations that produced the requirements, design, project context, code, and tests. Run:
Codex: $reviewer Continue IMPLEMENTATION review from .standards/STATE.md.Claude Code: /reviewer Continue IMPLEMENTATION review from .standards/STATE.md.Reviewer checks the implementation against the requirements, design, and test evidence. It records concrete problems and who must fix them. A different model of equal or higher capability is recommended when known, but optional.
If a correction is needed, follow the saved return instructions before continuing. Reviewer reassesses its findings after the responsible role fixes the work.
7. Document the change
Section titled “7. Document the change”This example uses the default FULL_DELIVERABLE policy. You can instead choose
to
finish after implementation review,
but the required guide in this example would still need Documenter’s evidence
before Reviewer could present the work for sign-off.
After implementation review passes, invoke Documenter from the handoff. It checks how search actually works and updates the user guide and any other required documentation.
Documenter saves what it checked, what changed, and the supporting evidence. The
guide’s completion is linked to AC-004.
8. Check the finished work
Section titled “8. Check the finished work”Documenter hands off to final Reviewer. Use a fresh chat separate from the conversations that authored the work, including documentation:
Codex: $reviewer Continue FINAL_DELIVERABLE review from .standards/STATE.md.Claude Code: /reviewer Continue FINAL_DELIVERABLE review from .standards/STATE.md.Final review checks the assembled result, including the user guide and current evidence for every acceptance condition. Earlier pending work must now be resolved.
After it passes, invoke Synchronizer. Synchronizer checks that the current files, completed assessments, and workflow records agree. If code changed after testing or review, the affected results must still be shown to apply.
9. Decide whether to accept it
Section titled “9. Decide whether to accept it”When the workflow reaches AWAITING_USER_SIGNOFF, read the summary and review
the work. You can:
- Sign off to accept it and finish the cycle.
- Request changes to return work to the responsible role.
- Cancel to end the cycle without accepting it.
Sign-off requires the checks to remain valid for the current files, with no unfinished corrections or blocking questions. Cancellation does not undo project changes. See Human Decisions.
If work stops along the way
Section titled “If work stops along the way”Answer a blocking question when a role needs your decision. If another role must fix a problem, follow the recovery handoff; it records where to return after the correction.
You can continue in another chat from the saved records. Follow the independent-chat requirements when returning to Tester or Reviewer. See Resuming Interrupted Work for a walkthrough.