agentleFS
Sign inStart free
sergeeeysergeeey/Claude-cod-top-2026/skills/core/routing-policy/SKILL.md

routing-policy

Skill9 starsChanged 6 months agoMITView on GitHub

Before starting ANY task: (1) know the PROJECT type, then (2) identify the TASK type and follow the route. Several skills overlap by topic but differ by mechanism (skeptic vs codex-skeptic vs grill-me; research vs research-scout vs deep-research). They are modes, not duplicates. When unsure which to invoke, consult docs/skill-disambiguation.md — it maps each family to "when to take this one". Do not fire all of them; pick by mechanism. Task routing alone is not enough — the SAME task…

  • Reads credentials

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

22 rules. About 0 are common, 18 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈18

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    ALWAYS CHECK before starting ANY task.

    line 12
  2. Distinctive:

    Do NOT use as an execution skill — it routes only, never implements.

    line 16
  3. Distinctive:

    Before starting ANY task: (1) know the PROJECT type, then (2) identify the TASK type and follow the route.

    line 31
  4. Very specific:

    When unsure which to invoke, consult docs/skill-disambiguation.md — it maps each family to "when to take this one".

    line 37
  5. Very specific:

    Read .claude/memory/_auto/project_profile.md (auto-written by the project_classifier hook at SessionStart).

    line 45
  6. Very specific:

    If confidence is LOW / ambiguous, invoke the dispatcher skill — it arbitrates by reading CLAUDE.md/README intent, not folder structure.

    line 47
  7. Distinctive:

    NOT call back into routing-policy for task classification (that would cycle) -- pass it the task type you already have, or classify it …

    line 49
  8. Distinctive:

    Every task passes through these stages in order.

    line 82
  9. Distinctive:

    Hard rule: Never skip Stage 1 to jump directly to Stage 3.

    line 91
  10. Distinctive:

    If not found locally → MCP (Context7, PubMed, WebSearch)

    line 101
  11. Distinctive:

    Read target file(s) in full

    line 110
  12. Very specific:

    Run tests (pytest -x -q or equivalent)

    line 113
  13. Distinctive:

    If tests fail → Read the error, understand the cause, THEN fix

    line 114
  14. Distinctive:

    NEVER write implementation before tests — because tests written after code validate the implementation, not the requirements (post-hoc …

    line 142
  15. Distinctive:

    Read full traceback / error message

    line 149
  16. Distinctive:

    Verify hypothesis with a tool → [VERIFIED]

    line 152
  17. Distinctive:

    If 3 attempts failed → Stuck Detection → STOP → ask user

    line 153
  18. Distinctive:

    If working with PII → confirm redaction hook is active

    line 162
  19. Distinctive:

    If the task does not clearly fall into one type:

    line 178
  20. Distinctive:

    When multiple skills could activate, follow this priority:

    line 193
  21. Distinctive:

    If skills conflict: higher layer wins.

    line 203
  22. Distinctive:

    If task type is unclear after 5 seconds, first check the boyko-agent criterion under "Task Type Identification" above — only default to …

    line 207

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md
<!-- BSV — Brief Skill View | поиск: BSV Скил : routing-policy TL;DR : Маршрутизатор задач — выбирает тип и путь решения Вызов : Автоматически при любой задаче НЕ для : Самостоятельная работа без задачи -->

name: routing-policy description: > [STATUS: review] [CONFIDENCE: high] [REVIEWED: 2026-03-13] ALWAYS CHECK before starting ANY task. Decision routing matrix for task→skill→agent→tools selection. Determines optimal execution path. Triggers: any task, start of work, new request, plan, task, implement, fix, debug, review, add, create, build, change. Do NOT use as an execution skill — it routes only, never implements.

Routing Policy — Task Routing Matrix

Current State (auto-injected)

!`git status --short 2>/dev/null | head -15`
!`git diff --stat 2>/dev/null | tail -5`

How to Use

Before starting ANY task: (1) know the PROJECT type, then (2) identify the TASK type and follow the route.

Picking between similar skills

Several skills overlap by topic but differ by mechanism (skeptic vs codex-skeptic vs grill-me; research vs research-scout vs deep-research). They are modes, not duplicates. When unsure which to invoke, consult docs/skill-disambiguation.md — it maps each family to "when to take this one". Do not fire all of them; pick by mechanism.

Stage 0 — Project Type (sets the methodology baseline)

Task routing alone is not enough — the SAME task needs different rigor in different projects (a "quick fix" in a research repo still affects conclusions; in an MVP it just ships).

  1. Read .claude/memory/_auto/project_profile.md (auto-written by the project_classifier hook at SessionStart). It gives project type + confidence.
  2. If confidence is LOW / ambiguous, invoke the dispatcher skill — it arbitrates by reading CLAUDE.md/README intent, not folder structure. When you invoke it this way, dispatcher will NOT call back into routing-policy for task classification (that would cycle) -- pass it the task type you already have, or classify it yourself once dispatcher returns the project type.
  3. The project type sets the methodology baseline; the task type (below) refines the route:
Project typeBaselineFL tier default
researchEstimandOps L0 + skeptic-triggers, claims [VERIFIED]/[HYPOTHESIS]Full
data-sciencevalidate on REAL data [VERIFIED-REAL]Standard
productionreviewer mandatory + coverage ≥80%Standard
MVPspeed > rigor, tests optionalMicro

Then apply the task matrix below within that baseline.

Absolute Safety Floor (applies on EVERY route — confidence or dispatcher-invocation does not matter)

The table above sets rigor, never an on/off switch for these. They apply whether confidence was HIGH (dispatcher never ran) or LOW/ambiguous (dispatcher ran and returned a verdict):

Always mandatoryEven if project type =
Security/PII/payments/secrets reviewMVP, production, any
Tests for migrations, destructive operations, releasesMVP
Evidence-marking for external claims ([VERIFIED]/[HYPOTHESIS]/[UNKNOWN])MVP, research
User confirmation for irreversible actionsany

MVP × "change token auth handling" is not the "tests optional" MVP row above — it falls under the security floor regardless of project type. This is a minimal addition, not a rearchitecture: routing-policy still owns task routing; dispatcher still owns project-type arbitration on ambiguous cases. The floor text is duplicated intentionally (see dispatcher/SKILL.md's own Safety Floor) so it holds even when dispatcher is never invoked.

4-Stage Workflow

Every task passes through these stages in order. Use them as a mental checklist:

StageGoalMaps to
1 — ResearchUnderstand before actingType 1, or Explore agent for Types 2-5
2 — PlanningDesign before codingTypes 2-3 (brief plan or EnterPlanMode)
3 — ActingExecute with evidenceTypes 2-6 execution routes
4 — ReviewVerify and capturereviewer agent + session-retrospective skill

Hard rule: Never skip Stage 1 to jump directly to Stage 3. Symptoms of skipping: editing without reading, committing without testing, closing task without review.

Routing Matrix

Type 1: Research / Question

Signals: "what is this", "how does it work", "where is it", "why", explain, find, search

Route:

  1. Explore subagent → Read/Grep/Glob (local tools)
  2. If not found locally → MCP (Context7, PubMed, WebSearch)
  3. Mark every fact according to Evidence Policy

Hard rule: local tools first, then MCP. Do not call MCP without first trying locally.

Type 2: Code Change (1-2 files)

Signals: "change", "add", "fix" + specific file/function

Route:

  1. Read target file(s) in full
  2. Brief plan in response (not EnterPlanMode)
  3. Edit/Write
  4. Run tests (pytest -x -q or equivalent)
  5. If tests fail → Read the error, understand the cause, THEN fix
  6. Commit

Type 3: Code Change (3+ files)

Signals: "refactor", "new feature", "migration", multi-file change

Route:

  1. EnterPlanMode (plan_mode_guard fires automatically)
  2. boyko-agent (Opus) → decompose into tasks (correction 2026-07-22: a prior edit this session changed this to "navigator", reasoning "there is no agents/boyko-agent.md" -- that check only looked for the FILENAME. PR #207 (2026-07-17) renamed the agent by changing agents/navigator.md's frontmatter name: field from navigator to boyko-agent; per agents/navigator.md's own "identity is defined by frontmatter, not the filename", boyko-agent IS the real, currently-resolving Agent-tool invocation name -- verified via the live Agent-tool roster, which lists boyko-agent and no longer lists navigator. The file is still named navigator.md for git-history continuity.)
  3. builder agent (Sonnet) → implement each file
  4. tester agent (Sonnet) → tests
  5. reviewer agent (Sonnet, high effort) → code review (was "Opus" here -- agents/reviewer.md is model: sonnet, effort: high; fixed 2026-07-21)
  6. Commit

Type 4: Writing Tests / TDD

Signals: "tests", "test", "coverage", "TDD", "with tests", "cover with tests"

Route:

  1. Load tdd-workflow skill (auto-loaded by trigger)
  2. RED → GREEN → REFACTOR
  3. NEVER write implementation before tests — because tests written after code validate the implementation, not the requirements (post-hoc confirmation bias)

Type 5: Debugging

Signals: "crashes", "error", "not working", "bug", error, fail, broken, debug

Route:

  1. Read full traceback / error message
  2. Explore subagent → search context in codebase
  3. Hypothesis [INFERRED] with chain of reasoning
  4. Verify hypothesis with a tool → [VERIFIED]
  5. If 3 attempts failed → Stuck Detection → STOP → ask user

Type 6: Security / Compliance

Signals: "audit", "security", PII, SQL, auth, payments, .env

Route:

  1. Load security-audit skill
  2. reviewer agent → code review
  3. Check deny-list (17 patterns)
  4. If working with PII → confirm redaction hook is active

Hard Guards — Absolute Rules

These rules override ANY rationalization:

GuardRulePrevents
Read Before EditDo NOT edit a file without reading it firstHallucination Loops
Local Before MCPDo NOT call MCP without trying locally firstWasted tokens
Plan Before Multi-EditDo NOT edit 3+ files without a planChaotic changes
Test Before CommitDo NOT commit without running tests (if tests exist)Broken commits
Evidence Before ClaimDo NOT assert a fact without an Evidence Policy markerHallucinations

Task Type Identification

If the task does not clearly fall into one type:

  • Contains "test/coverage" → Type 4 (TDD) takes priority
  • Contains "error/bug" → Type 5 (Debugging) takes priority
  • Contains "security/PII/auth" → Type 6 (Security) takes priority
  • Touches 3+ files → Type 3 (Multi-file) is mandatory
  • Direction is genuinely unclear (not just "little context in the prompt"), or the task is explicitly exploratory/hypothesis-driven/cross-domain → invoke Agent(boyko-agent, ...) first instead of defaulting to Type 2/3. boyko-agent's own whenToUse names exactly this case ("when direction is unclear... exploratory, cross-domain, hypothesis-driven") — routing straight to Type 2/3 here skips the one component built for it.
  • Anything else → Type 2 (Simple change) by default

Skill Execution Order (Layers)

When multiple skills could activate, follow this priority:

LayerPrioritySkillsRule
1 — SAFETYMandatory, always firstsecurity-audit, routing-policyNever skip. Check before any code change touching auth/PII/payments
2 — QUALITYBefore implementationtdd-workflow, brainstorming (if 3+ files)RED before GREEN. Design before code
3 — EXECUTIONTask-specific workarchcode-genomics, geoscan, notebooklm, mentor-mode, reference-registryLoad on trigger, one at a time
4 — ENHANCEMENTPost-taskgit-worktrees, last30days, research-pipelineIsolation, research, supplementary
5 — REVIEWPost-tasksession-retrospectiveCapture decisions, patterns, next focus

If skills conflict: higher layer wins. Layer 1 can block Layer 3 (security blocks unsafe code). Layer 5 always runs last — it reviews, never blocks.

Gotchas

  • This skill is meta — it routes, not executes. Don't let it become a bottleneck
  • If task type is unclear after 5 seconds, first check the boyko-agent criterion under "Task Type Identification" above — only default to Type 3 (code change) if that criterion doesn't apply

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free