Workflow for initializing or upgrading a workspace: context, discovery, documentation, etc.
SkillScore57Maintained
Automatic
◆DevOps & SRE pack
Last verified 8/27/2026
What it changes
Passed 10 of 17 test cases with the skill, and 4 of 17 without it.
Without skill26%
With skill57%
Checks it passed that the model alone missed
creates or reads gain.json during context detection
assigns model gpt-5.6-terra-high to the architect subagent
assigns model gpt-5.6-terra-medium to the reviewer subagent
and 8 more in the full report
9 fixed4 passed either way8 still failing2 broke
Pass-rate lift
+30 pp
Tokens
+36%
Agent turns
0%
Response time
-64%
Model gemini-3.6-flashMeasured 8/27/2026v413 of 23 cases passed with skill · 17 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 9 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 9
Prompt
In phase 1 workspace mode detection, the workflow checks whether the project is an install, upgrade, or plugin configuration. What specific JSON configuration file is created or read to store questions for later gap resolution?
What was checked, without and with the skill
Checked
Without skill
With skill
creates or reads gain.json during context detection
Missed
Passed
case-21 · gemini-3.6-flash · 8/27/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md159 linessingle file
<init_workspace_flow>
<description_and_purpose>
Problem: Workspace initialization is multi-phase, order-dependent, and must handle install/upgrade/plugin modes without overwriting human content.
Validation: State file tracks every phase with file inventory; verification confirms all files exist.
</description_and_purpose>
<workflow_phases>
<prerequisites phase="0", applies="ALL">
All Rosetta prep steps MUST be FULLY completed
MUST ALWAYS use todo tasks ledger, ASAP. Phases are sequential. Independent tasks can run in parallel.
MUST FOLLOW THIS WORKFLOW EXACTLY AND FULLY.
MUST extensively use subagents as this is a large workflow.
Each phase updates agents/init-workspace-flow-state.md. Optional phases marked as skipped. Keep state file very brief.
No rush, Take your time, Be thorough, ACCURACY > SPEED
Composite workspace: documentation phases to create top-level index referencing sub-repository docs.
IF state.file_count >= 100 (set by Phase 3): pass "USE SKILL large-workspace-handling" to Phase 5, 7, 8 subagents.
Create agents/init-workspace-flow-state.md.
Conditional phases:
If you have already in context "RUNNING AS A PLUGIN": MUST NOT EXECUTE "shells" phase 2
Else MUST EXECUTE "shells" phase 2
Note: rosetta@rosetta is an MCP connector, not a plugin — it follows the normal path (shells phase 2 executes)
If user says to initialize rules, subagents, agents, workflows, commands it ONLY means to execute "shells" phase 2.
Upgrade from R2 to R3 is exactly the same process as define here, but you already have some files available, which you can reuse.
Additionally tell subagents: "If you want to use shell commands, prefer to combine individual shell commands into single simple shell script and execute it, but already available tools ALWAYS take precedence."
When subagents already available, you are orchestrator and senior team lead and effective manager. Orchestrator makes process poka-yoke and reliable itself, provides clear context and instructions, uses subagents as his team, tells WHAT to do and HOW to think, does not work on tasks for subagents itself nor provides mechanical tasks nor paraphrases instructions, but appends context, ensures subagents provide grounded information, provides already known references to files, instructions, phases, steps, skills, keep agents focused. Prompt subagents to report honestly, concise, terse, and we exact links to files it created/modified.
Remember: subagents always start with fresh context on every run. User can not see orchestrator and subagent communication.
Subagent prompt must be concise, terse, factual, specific, DRY, grounded, etc.
ENFORCE provided required subagent models and effort levels.
Detect mode: install, upgrade, or plugin. Set state.mode, state.plugin_active, state.composite, state.existing_files. Creates/reads gain.json. GAIN.json questions keep for questions phase.