Skip to content

khuynh22/agent-dev-team / skills/using-agent-dev-team

using-agent-dev-team

Verified

Routes an incoming engineering request to the right team member and the right workflow, and explains the tier system. Use at the start of any development work when it is unclear who should do it, which skill applies, or how much process a task deserves. Also use when a task turns out to be bigger than it looked and needs re-routing.

Using the Agent Dev Team

This is the entry point. It answers three questions: how much process does this deserve, who should do it, and which workflow applies.

Step 1: Classify the work

ClassLooks likeProcess
TrivialA typo, a constant, a rename with known call sitesDo it. No ceremony. Still run the tests.
BoundedA change to a flow that already exists in this repositoryClarify, short plan in chat, implement, verify
FeatureNew behavior, several files, requirements not fully writtenRequirements, breakdown, then build task by task
ArchitecturalNew subsystem, changes a contract others depend on, hard to reverseRequirements, ADR, breakdown, then build
IncidentProduction is broken right nowStop. Go to incident-response.

When two classes fit, take the heavier one. The ratchet is one way: discovering hidden complexity upgrades the class mid-task. Nothing downgrades it.

Step 2: Route to a team member

SituationAgent
Fuzzy ask, no written requirementsproduct-manager
Requirements exist, no plantech-lead
Fully specified, mechanical, at most two filesintern-engineer
Clear requirement, existing pattern to followsoftware-engineer
Ambiguous, cross-cutting, or a migrationsenior-engineer
Two designs disagree, or a subtle correctness bugprincipal-engineer
A change is written and needs a gatecode-reviewer, then test-engineer
Touches auth, secrets, untrusted input, cryptosecurity-auditor
Measured slowness or a missed budgetperformance-engineer
Deploy, rollback, alerting, runbooksre
Production broken right nowincident-commander
MCU, RTOS, driver, memory budgetfirmware-engineer
New board, boot failure, signal-level bugboard-bringup-engineer
Component, state, layout, Core Web Vitalsfrontend-engineer
Usability, accessibility, interface copyux-reviewer
README, ADR, runbook, changelogdocs-engineer

Route to the lowest tier that clears the ceiling. Over-assignment is not caution; it hides the fact that the brief is too vague, and it costs more.

Step 3: Pick the workflow

PhaseSkillWhen
Definerequirements-interviewThe ask is fuzzy
Definespec-writingRequirements exist and need a written spec
Planwork-breakdownA spec needs splitting into verifiable tasks
Planarchitecture-decisionA non-obvious technical choice needs recording
Planapi-designAn interface others will depend on is being designed
Buildtdd-loopAny implementation
Buildincremental-deliveryExecuting a multi-task plan
Buildfrontend-buildThe change is user interface
Buildfirmware-buildThe change runs on a microcontroller
Verifysystematic-debuggingSomething is broken and the cause is unknown
Verifybrowser-verificationA UI change needs real-browser confirmation
Reviewcode-review-passA diff needs a gate
Reviewsimplification-passCode is more complex than the problem
Reviewsecurity-hardeningA trust boundary is involved
Reviewperformance-passSomething is measurably slow
Shiprelease-and-rollbackGetting it into production
Shipincident-responseProduction is broken
ShipdocumentationDocs are missing, stale, or wrong
Metateam-escalationA task exceeded the current tier's ceiling

Step 4: Work

  1. Announce the class, the agent, and the skill in one line, so the human can override it before work starts.
  2. Follow the skill. If it has a checklist, track it.
  3. Verify with evidence. Quote what you ran and what it printed.
  4. If you hit a ceiling, stop and use team-escalation. Escalating with a handoff packet is the system working, not a failure.

Finding the reference files

Skills and roles cite checklists as references/<name>.md. That path is relative to the agent-dev-team installation, not to the project you are working in, so a plain read from a project directory will miss.

Resolve it once, at the start of a session that needs one, and reuse the answer:

  1. If the harness exposes a plugin root variable, references/ sits directly under it.
  2. Otherwise search for **/references/escalation-ladder.md across the installation directory — ~/.claude/plugins/, ~/.claude/skills/, or wherever the repository was cloned.
  3. Under a copy-mode install the directory is ~/.claude/skills/adt-references/.

A checklist that cannot be found is not a reason to skip it. Say you could not load it and work from the process in the skill body, which is self-contained.

Red flags

ThoughtReality
"I know what to do, I'll skip routing"Routing takes ten seconds and prevents the wrong tier doing the work.
"I could not find the checklist, so I skipped that step"Say so. The skill body carries the process; the checklist adds depth.
"I'll assign it senior to be safe"Over-assignment hides a vague brief. Fix the brief.
"It grew but I'm almost done"Re-classify. Hidden complexity upgrades the class mid-task.
"No skill fits exactly"Take the nearest. Partial process beats none.
"The human is in a hurry, skip the gates"Urgency is when gates pay. Say what you skipped, if you must skip.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 468kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 319kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 319kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 319kUpdated Claude CodeCodexVerifiedInstall