<invocation_args>$ARGUMENTS</invocation_args>

Mission Statement

What It Is (and Isn't)

A plugin mission statement is a decision-making anchor, not a feature description. It states what the plugin values, what it refuses to do, and how it resolves trade-offs — enabling AI agents to evaluate alignment without asking the human each time.

Mission Statement Is	Mission Statement Is Not
A decision-making anchor	A feature list or capability overview
An explicit anti-pattern registry	A marketing or sales description
A trade-off resolution guide	A roadmap or version history
A verifiable alignment reference	A technical specification

The mission.json Format

The file lives at the plugin root (not in .claude/plan/).

{
  "status": "draft",
  "mission": "One sentence. What this plugin is trying to be — not what it does.",
  "values": [
    "Value statement 1 — concrete, verifiable",
    "Value statement 2"
  ],
  "anti_patterns": [
    "Specific behavior this plugin refuses to do",
    "Another explicit refusal"
  ],
  "escalation_triggers": [
    "keyword", "another keyword phrase"
  ],
  "trade_offs": {
    "correctness_vs_speed": "correctness",
    "breadth_vs_depth": "depth",
    "explicit_vs_implicit": "explicit"
  },
  "out_of_scope": [
    "Thing that looks related but belongs elsewhere"
  ],
  "interview_backlog_item": "#NNN",
  "validated_scenarios": []
}

Field definitions:

• status — "draft" until interview completes and human approves; then "active"
• mission — single sentence; must pass the "bad twin" test (a bad version of this plugin could not claim the same statement)
• values — observable principles that guide decisions; must be verifiable against behavior
• anti_patterns — explicit refusals; what this plugin will not do even when asked
• escalation_triggers — keyword list for fast string-match in alignment checks; no LLM needed
• trade_offs — when forced to choose, which side does this plugin take
• out_of_scope — things that look adjacent but belong in other plugins
• interview_backlog_item — GitHub issue number of the interview task (added after backlog creation)
• validated_scenarios — list of known past decisions that this statement correctly predicts

Development Process

Three phases:

1. AI Draft (immediate, this session) — source: discussion context + plugin files. Output: mission.json with status: "draft". The [draft] tag signals this is a hypothesis, not a decision.
2. Interview (async, via backlog task) — five structured questions asked to the human. Raw answers captured in backlog item. Output: updated mission.json with human-verified values.
3. Validation (after interview) — run 3 known past decisions through the statement. Does it predict the right choice? Output: validated_scenarios populated; status: "active".

The Five Interview Questions

These questions surface actual values, not stated ones. Ask them in order.

Q1 — The Non-Negotiable

"What is the one thing this plugin must never sacrifice, even to ship faster?"

Anchors values[0] — the primary principle.

Q2 — The Bad Twin

"What would a superficially similar but wrong version of this plugin do? What makes it wrong?"

Populates anti_patterns. A good mission statement the bad twin cannot also claim.

Q3 — The Trade-off

"When forced to choose between [breadth vs depth / correctness vs speed / explicit vs implicit], which does this plugin choose, and why?"

Ask all three. If the human says "both" — ask "if you could only have one." Answers populate trade_offs.

Q4 — The Removal Trigger

"What would make you remove this plugin from the marketplace entirely?"

Populates the most severe anti_patterns and escalation_triggers.

Q5 — The Anti-Pattern Example

"Give me a specific example of a 'fix' or 'improvement' this plugin should refuse to make, even if asked."

Most useful for alignment checks. Concrete refusals become escalation_triggers keywords.

AI Draft Procedure

When invoked (during Phase 0.6 of plugin lifecycle, or standalone):

1. Read the plugin's existing files: plugin.json or .claude-plugin/plugin.json, CLAUDE.md if present, SKILL.md files
2. Read discuss-CONTEXT.md if this is a new plugin creation
3. Draft mission.json with status: "draft". Populate all fields from observed design choices and stated preferences.
4. Write mission.json to the plugin root directory
5. Create a backlog interview task via mcp__plugin_dh_backlog__backlog_add with title "Mission interview: {plugin-name}" and body containing the 5 questions and the current draft mission field
6. Update mission.json with "interview_backlog_item": "#NNN" using the created issue number
7. Report: path of draft written, backlog item number, 2-3 sentence summary of draft mission

Validation Scenario Format

After interview, validate by running known decisions through the statement. Add each to validated_scenarios:

{
  "validated_scenarios": [
    {
      "decision": "Refused to add blanket noqa suppression to a linting plugin",
      "predicted_by": "anti_patterns[0] + escalation_triggers",
      "outcome": "correct"
    }
  ]
}

Status becomes "active" when it correctly predicts at least 3 known decisions.

Standalone Invocation

Arguments: <invocation_args/>

• <plugin-path> — Draft mission for an existing plugin. Read plugin files, draft mission.json, create backlog interview item.
• <plugin-path> --interview — Conduct the interview synchronously in this session. Ask Q1-Q5, update mission.json from answers, move to validation.
• <plugin-path> --validate — Run validation scenarios. Ask human to confirm 3 past decisions, check predictions.

Relationship to Alignment Check

The escalation_triggers list is the fast path for alignment checking — pure string matching, no LLM:

1. Check proposed action text against escalation_triggers — if any keyword matches, escalate immediately
2. If no keyword match, check anti_patterns with LLM reasoning against mission and values
3. If action contradicts anti_patterns or moves away from values, return alignment: LOW with the specific violated principle

The mission statement answers "what would the human say if they were watching?" — codified in advance.