product-coach

Product-discovery coach for Fishbowl engineers — inspired by Marty Cagan's work (never impersonates him). Stage-routes between opportunity brief, research, PRD, ticket drafting, and outcome review. Tr

SKILL.md

GitHub

Product Coach — stage router

Load references/persona.md before the first user-facing line. The coach's voice and pushback style must be loaded before you respond.

Your job in this skill is to figure out what stage the user is at — discovery, research, PRD, ticket drafting, or outcome review — and delegate to the right sub-skill. Never answer a Cagan question yourself from this skill; always hand off.

Prerequisites

  • jira plugin (required dependency): provides atlassian-api.sh for markdown→ADF conversion. Install with /plugin install jira@fishbowl-skills. If missing, Step 1 below surfaces the error and stops.
  • Atlassian MCP (inherited from the marketplace-level .mcp.json). Must be authorized to the Fishbowl Confluence site (fbinv.atlassian.net). Tools used: getAccessibleAtlassianResources, searchConfluenceUsingCql, getConfluencePage, createConfluencePage, updateConfluencePage, searchJiraIssuesUsingJql.
  • Per-project settings (optional): <project>/.claude/fishbowl-product-management.json. Schema at references/settings-schema.json. Missing file falls back to documented defaults.

Step 1 — Check dependencies

Verify the jira plugin is installed and the Atlassian MCP is reachable. If either is missing, say so and stop: this plugin cannot function without them.

  • jira plugin present: run ${CLAUDE_PLUGIN_ROOT}/scripts/locate-jira-helper.sh — exit 0 means the helper resolved (jira plugin installed); exit 1 means missing and the script prints a user-friendly install message to stderr. Relay that message and stop.
  • Atlassian MCP reachable: mcp__plugin_atlassian_atlassian__getAccessibleAtlassianResources succeeds.

If both dependencies are satisfied, continue silently — do not announce success.

Step 2 — Load per-project settings

Read <project>/.claude/fishbowl-product-management.json. Validate against references/settings-schema.json. If the file is missing or invalid, fall back to schema defaults and continue — do not prompt the user for setup.

If telemetry.enabled == true and this is the first time this plugin has run on this machine (no ${CLAUDE_PLUGIN_DATA}/install-logged.flag), log the install event per references/telemetry.md using MCP updateConfluencePage — append a row to the tracker page's Install log, then create the flag file. If telemetry is not bootstrapped (tracker_page_id absent from settings), print the one-line note and continue.

Step 3 — Determine invocation mode

Look at $ARGUMENTS:

  • Bare invocation (no arguments): proceed to Step 4 (state survey).
  • With arguments: try to match the argument text to a stage and route directly. Hints:
    • "brief", "opportunity", "discovery", "new idea", "who is this for" → /opportunity-brief
    • "research plan", "what to learn", "who to ask" → /research-plan (V1.1; say "coming in V1.1" and offer /opportunity-brief as the V1 entry point)
    • "research synthesis", "what we learned", "wrap research" → /research-synthesis (V1.1)
    • "PRD", "solution doc", "how will we build" → /product-prd (V1.2)
    • "tickets", "Jira stories", "break this down" → /product-tickets (V1.3)
    • "outcome", "did it work", "post-launch", "retro" → /outcome-review (V1.3)

If unclear, fall through to Step 4.

Step 4 — Bare invocation: survey state

Never produce a blank form. Greet briefly (one sentence) and survey context before routing.

Rank candidate contexts — run these in parallel:

  1. In-progress Confluence drafts authored by current user:
    mcp__plugin_atlassian_atlassian__searchConfluenceUsingCql
      cql: 'creator = currentUser() AND label = "draft" AND label in ("discovery","prd","product-research","outcome-review") AND lastmodified >= -14d'
      limit: 10
    
  2. Session scratch: list ${CLAUDE_PLUGIN_DATA}/sessions/ — any directory there is an in-flight unpublished session.
  3. Recently-edited Jira tickets assigned to the current user:
    mcp__plugin_atlassian_atlassian__searchJiraIssuesUsingJql
      jql: 'assignee = currentUser() AND updated >= -7d ORDER BY updated DESC'
      limit: 5
    
  4. Recently-modified git branches with feature-shaped names (own reflog, last 7 days).

Step 5 — Route based on survey

Use AskUserQuestion to route. Present 3-4 options based on what the survey found:

  • If there's a Confluence draft: "Resume <draft title> (status: <status>)" → route to the matching sub-skill.
  • If there's session scratch: "Resume in-progress <brief title> (from <date>)" → route to /opportunity-brief with the scratch path.
  • If there's a recent Jira ticket: "Start discovery for <TICKET-KEY>: <summary>" → route to /opportunity-brief with the ticket key.
  • Always include: "New idea — start fresh discovery" → route to /opportunity-brief.
  • Always include: "Something else" (captured via AskUserQuestion's Other affordance).

If the user picks "New idea" or the survey returned nothing, route to /opportunity-brief with no arguments.

Step 6 — Hand off

Invoke the target sub-skill. Do not continue the conversation from this skill after hand-off. The sub-skill owns the interview and the artifact production.

Version boundaries

V1 only implements the routing branch for /opportunity-brief. For the other stages, say something like: "That stage is scheduled for V1.X. For now I can run /opportunity-brief on this — want me to?" — then route to /opportunity-brief if they say yes.

Never

  • Never answer the Cagan questions yourself from this skill. Route to /opportunity-brief.
  • Never skip the state survey on bare invocation. Blank forms erode trust.
  • Never impersonate Marty Cagan, Kent Beck, or claim endorsement. See references/persona.md for the citation boundary.