Tandem — core PM rules

You are working in a project that uses the Greenfield PM Operating Kit. These rules are mandatory whenever you touch anything under _00-Project-Management/.

Lifecycle command chain (canonical order — single source of truth)

This is the one place the lifecycle command order is recorded (ADR-0047). Every per-command Next: pointer in the lifecycle skills must agree with this chain; if a pointer ever disagrees, this record wins and the pointer is the bug.

The chain shares one planning prefix, then forks at execution-strategist into two honest paths. Follow exactly one path per unit of work — each step then runs exactly once:

Planning prefix: /tandem:draft-okrs → /tandem:draft-prd → /tandem:draft-epic → /tandem:split-into-features → /tandem:split-into-stories → /tandem:refine-backlog → /tandem:execution-strategist

• Single-story path: /tandem:execution-strategist → /tandem:execute-story → /tandem:run-testplan → /tandem:close-out-story
• Batched path: /tandem:execution-strategist → /tandem:start-phase → /tandem:execute-batch → /tandem:close-phase — execute-batch composes the per-story skills (execute-story → run-testplan → close-out-story) internally; never re-run them separately after a batch, and never run a batch without start-phase cutting the phase branch first.

close-phase is terminal (no Next: pointer). Every skill not named in the chain is a non-chain utility (cadence / review / setup — e.g. weekly-monitor, monthly-retro, reflect, session-start, critique, peer-review, document, curate-toolkit, fill-claude-md, mode); the two fork paths above are the only lifecycle orders.

Reference order — where to look

1. Project root CLAUDE.md — pointers + critical gotchas
2. _00-Project-Management/CLAUDE.md — folder semantics
3. _00-Project-Management/90-Standards/SOP.md — full lifecycle, DoR, DoD, frontmatter contract, subagent policy (§18)
4. _00-Project-Management/90-Standards/PROJECT-CONTEXT.md — this project's stack quirks
5. _00-Project-Management/90-Standards/DAILY-WORKFLOW.md — rhythm + worked example
6. _00-Project-Management/90-Standards/CLAUDE-CODE-CONFIG.md — how this plugin maps to Anthropic's Claude Code best practices
7. Template in _00-Project-Management/91-Templates/

Non-negotiable rules

Frontmatter timestamps

Every artefact has three timestamp fields:

created_at: ''      # set on file create; ISO 8601 with offset, quoted string
started_at: ''      # set when status → in-progress
completed_at: ''    # set when status → done | wontfix | duplicate | archived

• Format: YYYY-MM-DDTHH:MM:SS±HH:MM. Always quoted.
• Source of "now": system clock (Get-Date -Format "yyyy-MM-ddTHH:mm:sszzz" or date -u +"%Y-%m-%dT%H:%M:%S%z"), not the chat-stated date.
• Status flip + timestamp set = same edit, not separate.
• Status revert (back to not-started) → clear started_at and completed_at to ''.

Status enum — closed set, exactly 9 values

not-started | ready | in-progress | in-review | done | blocked | wontfix | duplicate | archived

Never invent values. Never use open / shipped / completed / fixed / deferred / Planned.

Folder layout (full | flattened | custom)

The kit's skills and templates name PM sub-folders in the canonical "full" numbering: 30-Epics, 31-Features, 32-Stories, 33-Testplans, 34-Bugs, 40-Decisions, 42-Monitor, 11-Backlog, …

A project may instead use a "flattened" layout (e.g. 01-EPIC, 02-Features, 03-Stories, 05-Test, 04-Bug, 06-ADR, 00-Monitor) or a custom map.

The scripts resolve the real folder names automatically via _00-Project-Management/93-Scripts/lib/pm-paths.js, driven by .claude-pm-config.json ("layout": "full"|"flattened" and/or a per-key "paths" override). pm:install pins this; pm:doctor reports it.

INSTRUCTION TO CLAUDE: when a skill names a folder like 32-Stories, do not assume that literal path — resolve it to the project's actual folder first (read .claude-pm-config.json, or list _00-Project-Management/ and match by role: epics/features/stories/testplans/bugs/decisions/monitor/backlog). On a flattened project: 32-Stories→03-Stories, 33-Testplans→05-Test, 34-Bugs→04-Bug, 40-Decisions→06-ADR, 42-Monitor→00-Monitor.

Story → Testplan pairing — MANDATORY

When creating a STORY under 32-Stories/EPIC-NN/FEAT-NN.M/, create the paired TESTPLAN at 33-Testplans/EPIC-NN/FEAT-NN.M/TESTPLAN-NN.M.PP-<slug>.md in the same response.

• Every AC checkbox in the story maps to ≥1 TC in the testplan.
• Every TC has a runnable Command: (no manual steps, no "have a human verify").
• See 91-Templates/STORY.template.md + 91-Templates/TESTPLAN.template.md.

Bug auto-raise on failure — MANDATORY

Whenever a TC fails or you observe any defect during exploration/code review, file a BUG at 34-Bugs/EPIC-NN/FEAT-NN.M/BUG-<YYYYMMDD-NN>-<slug>.md in the same response, before reporting in chat.

• ID format: BUG-YYYYMMDD-NN where NN is the day's sequential counter within that FEAT folder.
• Slug: kebab-case, ≤6 words, describing the symptom.
• Body includes: reproduction steps, environment snapshot, first analysis hypothesis, suggested fix direction a junior dev can act on.
• Use 91-Templates/BUG.template.md.

DoR gate — before in-progress

Before flipping a story not-started/ready → in-progress, verify the DoR checklist in SOP.md §6. If a DoR item is missing, stop, list the gap, ask.

DoD gate — before done

Before flipping a story in-review → done, verify the DoD checklist in SOP.md §7. MONITOR.md update is part of the same response.

ADR on the spot — MANDATORY for non-obvious decisions

On any non-obvious decision (library choice, schema field name, threshold setting, scope deferral, divergence from defaults), create 40-Decisions/ADR-<NNNN>-<slug>.md in the same response. Number sequentially across the project. Link from the story's decisions: array.

Templates over memory

Every new artefact starts from 91-Templates/<TYPE>.template.md. Do not redraft section headings from memory.

Artefact economy — quality over count

Before creating any EPIC / FEAT / STORY / BUG / BACKLOG artefact, apply 90-Standards/ARTEFACT-ECONOMY.md: choose the smallest level that carries the claim; fewest artefacts that carry the claims; one intake item per initiative (tranches, not siblings); every artefact carries a falsifiable claim; small debt rides the next story touching the same files. Kill-before-promote cadence: triage-backlog before refine-backlog. When unsure between two levels, pick the smaller.

Concision — artefacts and chat (convention, not gate)

One writing convention for every producer skill and every hat; not a new MANDATORY gate.

• Artefacts: Match section length to substance — cover what the ACs need and stop; don't pad sections to look complete.
• Chat: lead with the outcome or verdict; supporting detail after. Plain English, complete sentences, no ceremony or filler openers. Explain a thing once — don't restate what a table, file, or earlier line already says.
• Long enumerations (finding lists, per-item tables, logs) belong in an artefact or report; chat carries the verdict plus the pointer.
• Respect stated output budgets where a skill declares one (e.g. session-start ≤ 25 lines).

Strategy linkage

Every EPIC must have okr: or prd_section: in frontmatter. Reject epics without strategic linkage — ask "What business outcome does this move?" before writing.

Hat protocol

State which hat at session start: Founder · PM · Dev · QA. Don't mix hats in one session.

Conversation Mode (the moat) — enforces the hats across chats

A project-global Mode (plan · dev · dual · neutral) groups and enforces the hats across separate chats. It persists in .tandem-mode.json (repo root, git-ignored) and is injected each message by the UserPromptSubmit hook for joined chats. Plan groups Founder/PM/QA-planning; Dev groups Dev/QA-execution. On an out-of-mode request, nudge (switch / go Dual / one-off) — never hard-block. Set it with /mode <plan|dev|dual|neutral>; it auto-resets to Neutral on close-out-story (last story in phase) and close-phase. Full rules: the mode skill.

MONITOR + dashboard

When a story flips to done, update 42-Monitor/MONITOR.md in the same edit (tick the bar, update shipped count, prepend revision-history one-liner). The dash hook (Stop event) regenerates DASHBOARD.html at session end — you don't need to run npm run pm:dash manually if the plugin is active.

Subagent delegation (SOP §18)

• Editing / decisions / status flips → main thread.
• "Where is X / which files reference Y" → Explore agent (read-only).
• Multi-step research, running tests, anything producing noisy logs → fresh agent.
• Never delegate understanding. Agents return evidence; main thread synthesises.

When in doubt

Bring it to the user. Do not invent rules. Do not silently bend.