gate-plan

name	gate-plan
description	Authors a gate document for a queue item — a production validation plan that defines what must be proven before work is considered done. Applies the quality bar and self-assessment questions that gate-review audits against.
when_to_use	Use when the user asks to create/plan/write/draft a gate, create a gate for a specific Q-number (e.g. "gate Q4"), write a validation plan, or draft validation criteria for a queue item.

Gate Plan — Gate Document Authoring

Produce a gate document for a queue item. The gate defines what must be proven before the work can be considered done. It is a production validation plan, not a checklist.

Input

The operator provides a Q-number (e.g., "Q4", "4"). This maps to a row in QUEUE.md.

Process

Step 0 — Bind the Input

The Q-number this gate plans against is not a parameter the skill body already holds: the harness appends it as an ARGUMENTS block after this skill's body. That trailing block is the ## Input above — read it as the Q-number to plan against, and resolve it against the matching QUEUE.md row and the frozen Q{n}-why.md in Step 1. This is the binding clause only — gate-plan's Step 1 (Resolve Context) already consumes that input, so no further consume-fix is needed, and gate-plan mints no Q-number; it reads one. The clause is here for defect-class consistency with the input-binding discipline applied across the chassis skills, not to repair a drift.

Step 1 — Resolve Context

1. Read QUEUE.md at the workspace root. Find the row matching the Q-number. Extract the intent, scope, and next action. If the Q-number doesn't exist in the queue, stop and tell the operator. Then, alongside the row — not instead of it — read the companion frozen intent record Q{n}-why.md at the workspace root when it is present. This is the record triage freezes for the item; it carries the direction (where the work points), the named fuzz (the conscious uncertainty), the scope-carve (what was deliberately cut), and the revisit-trigger (the condition that should reopen the decision). Draw on these to author completion criteria that pursue the frozen direction and an Excluded section that honors the carve. The read is best-effort, not a precondition: most live queue items predate triage's freeze and carry no Q{n}-why.md, so when the file is absent, proceed on the QUEUE.md row alone — a missing why-record is the normal path, not a halt.
2. Check for inter-gate dependencies. If the queue item's Next field references other Q-items, or if the intent logically depends on prior work, check whether those items have gates and whether those gates have cleared. This informs the Depends on: field in the gate document.
3. Check if a gate file already exists at the workspace root (Q{n}-gate.md). If it does, stop and tell the operator — do not overwrite without explicit confirmation.
4. If the scope references a project directory, read that project's README or top-level structure to understand what's being validated.

Step 2 — Internalize the Pattern

Read existing *-gate.md files at the workspace root and in gates/ (if more than 3 exist total, focus on the 2-3 with the highest Q-numbers). Active gates live at the workspace root; cleared gates are archived in gates/. Both are valid pattern references. The gate template (references/gate-template.md) defines the structural standard. The existing gates show how that standard is applied in practice.

Read the anti-pattern registry (${CLAUDE_PLUGIN_ROOT}/references/anti-pattern-registry.md). The registry catalogs named failure modes with scope, detection, and prevention fields. During drafting (Step 3), you will identify checkpoints that guard against these anti-patterns and tag them using the convention defined in gate-template.md.

Pay attention to:

• How completion criteria are stated (first paragraph — concrete, claimable, specific)
• How phases are scoped (each phase is a coherent stage of validation, not an arbitrary grouping)
• How checkpoints are verified (positive artifacts, not absence of errors)
• How dependencies between checkpoints are documented
• How bypasses are justified inline
• How the Excluded and Cleanup sections scope the work

Step 3 — Draft the Gate

Write the gate document. Apply these quality requirements:

Completion criteria must be claimable. The first paragraph states what the operator can claim when the gate clears. "Validate the system" is vague. "Targeted extraction across 5 implementations to understand how different engineers solve X" is claimable — it names the scope, the method, and what understanding is produced. A practitioner should read the completion criteria and know exactly what was proven.

Every verification method must produce a positive artifact. A passing verification method must leave evidence that it passed. Command output to quote, a file that exists, a document with specific content, a count that matches an expectation. If a verification method can pass when nothing happens — no output, no artifact, silent success — it's proving a negative. Design verification methods that succeed visibly, not ones that pass by default. When a checkpoint contains multiple verification methods, each method is assessed independently. A positive sibling method does not excuse a negative-proof method within the same checkpoint — a checkpoint with two positive-evidence methods and one negative-proof method (e.g., "no errors in output") has a masking defect where the negative-proof method is hidden by its positive siblings at the checkpoint level. This is a hard requirement: a verification method without a defined positive artifact is a blocking deficiency in the self-assessment (Step 4), not a suggestion to improve.

Every checkpoint must be categorized as structural or operational. Use the [structural] or [operational] tag after the checkpoint ID as defined in the gate template. Structural checkpoints verify existence or form — file present, format valid, command succeeds. Operational checkpoints verify that the deliverable produces meaningful results — pipeline generates output, tool processes real input, integration completes a cycle. Preflight gates must include at least one [operational] checkpoint. A preflight gate with only structural checkpoints has direct Vacuous Green exposure and must not be presented — it fails the self-assessment in Step 4.

Verification must exercise what it claims to validate. If a checkpoint says "study document produced," the verification must confirm the document exists and contains the expected sections — not just that the command exited cleanly. If a checkpoint says "all 5 repos on disk," the verification lists them. Don't write a checkpoint that could pass without the thing it validates actually being true.

Ordering dependencies must be explicit. If a verification phase depends on prior phases producing artifacts, say so. If a checkpoint assumes setup from an earlier phase, document the assumption. An agent executing this gate top-to-bottom should never hit a checkpoint that fails because a prerequisite wasn't stated.

Specify outcomes, not tool sequences. Don't prescribe the mechanical steps to achieve a checkpoint. Specify what the result looks like, not how to get there. The executing agent chooses its own approach — the gate defines what "done" means for each checkpoint, not the procedure.

Bypasses are decisions, not shortcuts. [~] requires inline justification explaining why the checkpoint is being skipped and what risk is accepted. A bypass without reasoning is a gap.

Document cleanup expectations. If the gate produces temporary artifacts, test repos, or scratch state, the Cleanup section should specify what to remove after the gate clears. An executing agent should know what to tear down when done.

Scope out explicitly. If something is deliberately not covered, it belongs in the Excluded section with a reason. This prevents an executing agent from scope-creeping into adjacent validation and prevents the operator from wondering why something was missed.

Multi-vector gates require a coverage matrix. When a gate targets remediations across multiple vectors and multiple lifecycle stages, the gate document must include a coverage matrix mapping every identified vector (rows) to every applicable lifecycle stage (columns), with the enforcement mechanism named in each cell. Empty cells are gaps — they represent vector-stage combinations where no enforcement exists. All empty cells must be addressed (filled with a mechanism or explicitly scoped out with justification) before the gate is presented. This requirement applies only to gates that cross both multiple vectors and multiple lifecycle stages; single-vector or single-stage gates are not required to produce a matrix.

Operator-terminal tags require a named execution constraint. When a specific, documented project-level constraint prevents an agent from verifying an operational checkpoint within a Claude Code session, the gate must tag that checkpoint [operator-terminal] and state the constraint inline. The tag is justified by the named constraint, not by the project's technology stack or framework category. For example, the-range's PostToolUse hook control protocol has a bug where continue: false is silently ignored by the CLI when invoked via the SDK control protocol — operational checkpoints requiring hook enforcement cannot be verified from within a nested Claude Code session, so they require operator execution from a separate terminal. Each [operator-terminal] tag in the gate must name the specific constraint (project, mechanism, consequence) that prevents agent verification. A gate that applies [operator-terminal] tags without per-tag constraint justification has a blocking deficiency.

Cross-domain reachability requires a delivery checkpoint. When a gate modifies artifacts consumed external to the current domain, the gate must include at least one checkpoint verifying the changes are reachable to that domain. "External to the current domain" means the artifact's consumer operates in a different domain than where the gate executes — a chassis skill modified in Workshop but consumed by Forge agents, a vault artifact produced in Research but referenced in Lab. The checkpoint must verify reachability through the actual distribution path (plugin cache version, vault file presence, deployed configuration), not just source-level correctness (repo commit, file content). A gate that validates source changes without verifying consumer reachability has proven correctness but not delivery. This requirement applies only to gates modifying cross-domain artifacts; gates whose artifacts are consumed entirely within the authoring domain are not required to include a delivery checkpoint. When the distribution path requires operator action or depends on platform mechanics outside the gate agent's control (e.g., plugin cache refresh, manual deployment), the gate may scope out delivery verification in the Excluded section with reasoning that names what the agent cannot control and why the in-scope checkpoints already cover correctness. This is a valid resolution, not a blocking deficiency. When the delivery checkpoint's distribution path is a version bump — a plugin-cache or marketplace manifest version — the checkpoint and any prose describing the bump must express the bump relative to the version read at execution time ("read the current value, advance one patch level"), never as a hard-pinned target (0.6.0 → 0.6.1) or a present-state literal about the live artifact ("ships in 0.6.0"). Sibling gates may ship in the authoring→execution gap, so a version literal fixed at authoring time names a moving target and goes stale silently — and prose drifts even when the checkpoint itself stays relative, so the rule binds both. A frozen historical baseline a cleared predecessor established stays legitimate to name — a retrospective record of a past value ("prior value was 0.6.1, confirmed from git history") states a fixed fact that cannot drift; only a forward-target or present-state literal on a moving target is prohibited.

Source-anchored decision-token legend (conditional). When a gate cites decision-record tokens — d-NNNN, ADR-NNNN, or an equivalent external decision ID — the gate must carry a Decision-Token Legend (see references/gate-template.md). For each cited token, resolve it against its own source once, at authoring time — not against a secondary citation (another gate, a triage note, a prior review), which can itself be wrong — and record in the legend one decision title and one <corpus-path>#<record-id> source pointer. Before presenting the gate, run the deterministic citation checker (${CLAUDE_PLUGIN_ROOT}/skills/gate-review/scripts/check-citations.py <gate-file>, run from the workspace root so pointers resolve) and confirm a GREEN or N/A exit; a RED exit is a defect to fix before submission, and an UNRESOLVABLE exit means a cited corpus is not on disk in this environment. The requirement is conditional: a gate that cites no decision-record tokens carries no legend and runs no checker. Anti-pattern tags ({AP-nn}) are not decision-record tokens and do not trigger this requirement. The legend is the authoritative resolution of a token's identity (the authority rule, defined once in references/gate-template.md) — do not restate a divergent identity inline.

Predecessor-gate inheritance (conditional). When the queue item reuses a predecessor gate's load-bearing semantics — its runtime, fixtures, contracts, or baseline values — the gate must set the Builds on: field (distinct from the blocking Depends on:; see references/gate-template.md) naming that predecessor, and transcribe the load-bearing fraction (not the whole predecessor) into a matching ## Inherited from {predecessor} section, so a cold gate-work executor reaches that context from the single file it loads. Before presenting the gate, run the deterministic inheritance checker (${CLAUDE_PLUGIN_ROOT}/skills/gate-review/scripts/check-inheritance.py <gate-file>, run from the workspace root) and confirm a GREEN or N/A exit; a RED exit (a declared Builds on: with no matching Inherited from section) is a defect to fix before submission. The requirement is conditional: a gate that reuses no predecessor's load-bearing semantics sets Builds on: None (or omits the field) and the checker returns N/A. The checker enforces section presence and predecessor-name match, not transcription completeness — that the section carries the right load-bearing fraction remains the author's responsibility and gate-review's judgment (${CLAUDE_PLUGIN_ROOT}/adrs/ADR-003-gate-to-gate-inheritance.md). A Builds on: relation is distinct from a Depends on: one: a gate may build on a cleared predecessor it does not depend on.

Tag checkpoints that guard against known anti-patterns. After drafting checkpoints, review them against the anti-pattern registry read in Step 2. Identify checkpoints where a genuine guarding relationship exists — the checkpoint prevents or detects the failure mode described by a registry entry. Apply the {AP-nn} tag convention from gate-template.md to those checkpoints. Tags are applied selectively: not every checkpoint needs a tag, and forcing tags where no anti-pattern applies degrades the signal. A checkpoint that prevents Vacuous Green (AP-07) gets tagged; a checkpoint that validates file formatting with no anti-pattern relevance does not.

Step 4 — Self-Assess Before Presenting

Before showing the draft to the operator, evaluate it:

1. If every checkpoint were cleared, does that fully justify the completion criteria in the first paragraph? If not, there are missing checkpoints.
2. Does every verification method within every checkpoint produce a positive, observable verification? Flag any that prove negatives or could pass silently — including negative-proof methods masked by positive siblings within the same checkpoint.
3. Are all ordering dependencies between checkpoints and phases documented?
4. Could an agent execute this gate top-to-bottom without the operator having to clarify sequencing, prerequisites, or intent?
5. Is the checkpoint granularity appropriate — detailed enough to be unambiguous, not so granular that it's micromanagement?
6. Does the gate include operational checkpoints that verify first-iteration readiness, not just structural completeness?
7. If the gate targets multiple vectors across lifecycle stages, does it include a coverage matrix with zero empty cells?
8. If the gate modifies artifacts consumed external to the current domain, does it include a checkpoint verifying consumer reachability — or, if the distribution path is outside the agent's control, does it scope out delivery verification with reasoning in the Excluded section? And when the delivery checkpoint's distribution path is a version bump, do the checkpoint and its describing prose express the bump relative to the execution-time value ("read the current value, advance one level") rather than as a hard-pinned target or a present-state literal?
9. Does every [operator-terminal] tag carry a stated project-specific execution constraint — naming the project, the mechanism, and why the agent cannot verify that checkpoint within a Claude Code session? Would a tag justified only by the project's technology category (without naming the specific constraint) fail this check?
10. Do the gate's checkpoints cover relevant anti-patterns from the registry? Where a guarding relationship exists between a checkpoint and a registry entry, has the {AP-nn} tag been applied? Are there relevant anti-patterns with no guarding checkpoint that should be addressed?
11. Does any checkpoint prescribe a specific tool invocation, command sequence, or procedure instead of defining what the result looks like?
12. If the gate cites decision-record tokens, does the Decision-Token Legend carry one decision title and one <corpus-path>#<record-id> source pointer per token, and does the citation checker exit GREEN/N-A? A RED exit, or a missing legend on a token-citing gate, is a blocking deficiency.
13. If the gate builds on a predecessor gate's load-bearing semantics, does it set Builds on: and carry a matching ## Inherited from {predecessor} section transcribing the load-bearing fraction, and does check-inheritance.py exit GREEN/N-A? A RED exit (declared Builds on: with no matching section) is a blocking deficiency.

If the answer to any of 1-10, 12, or 13 is no, or the answer to 11 is yes, fix the gaps before presenting. Do not present a draft you assess below 4/5 confidence and ask the operator to identify what's wrong. That's the operator's time wasted on work the agent should have caught.

Step 5 — Output

Write the gate file to the workspace root as Q{n}-gate.md. Present a brief summary of what the gate covers and how many phases/checkpoints it contains. The operator will review and may request changes.

Related Skills

• gate-review — Audits a gate document against the quality bar defined here. Review sits between plan and work in the lifecycle.
• gate-work — Frames an agent to execute an existing gate. Plan creates the gate; work executes it.
• prime — Session context loading. Prime and gate-plan are independent — prime loads context, gate-plan authors a gate document.

Reference Files

• references/gate-template.md — Structural template for gate documents. Conventions for titles, phases, checkpoints, verification methods, anti-pattern tags, and operational sections.
• ${CLAUDE_PLUGIN_ROOT}/references/anti-pattern-registry.md — Named failure modes with scope, detection, and prevention fields. Read during Step 2 for anti-pattern tagging in Step 3.