NOTE: This is the canonical provider-independent contract. Both provider wrappers must load it in full.
Merge two (or more) plan documents into one. The output is a single plan where overlapping goals are resolved, phases are re-sequenced into a coherent order, and the originals are archived — not deleted — so their context remains inspectable.
This skill is conversational. It does not auto-merge. It surfaces the decisions a human has to make (which plan is the spine, how to resolve overlaps, how to order dependencies) and then writes the merged plan only after those decisions are settled.
- Two plans in
documentation/plans/target related work and the team hasn't decided which one to execute. - One plan describes a capability (e.g. "add feature X") and another describes the infrastructure that feature needs. They overlap at the seam and need reconciliation.
- A plan's phases reference work that is separately planned elsewhere.
- An old plan is partially superseded by a newer one and they need to be unified rather than left to rot in parallel.
Do NOT use this skill:
- To merge a plan with its own code — that's
/repo-updateor direct editing. - To refine a single plan — that's
/plan-review. - To split a plan into smaller phases — write a new plan.
- Paths to two or more plan files (ask if not provided).
- Optionally, the target path for the merged plan. Default:
documentation/plans/<descriptive-name>-master-plan.md.
Do not skim. The overlap that matters is rarely in the introduction — it hides in phase-level details, schemas, contracts, and kill criteria.
For each plan, extract all five of the following fields. The output
MUST include every field for every plan — if a field is empty in the
source, write (none found) explicitly rather than omitting the line.
Dropping any of the five fields makes Step 2 invalid and Step 4's spine
proposal unsound.
- Vision / goal — what problem does it solve?
- Phase list with one-line summaries
- Data contracts (DB schemas, file formats, APIs)
- External dependencies (branches, PRs, other plans, skills)
- Kill criteria and rollbacks
Render the extraction using this exact template, once per plan (do NOT collapse fields, do NOT drop the External dependencies or Kill criteria lines even if they feel redundant):
#### Plan: <name>
- Vision / goal: <...>
- Phase list: <...>
- Data contracts: <...>
- External dependencies: <...>
- Kill criteria and rollbacks: <...>
Then compare: where do the two plans touch the same concept?
Each overlap MUST be classified by exactly one of these four catalog
patterns, prefixed verbatim as [PATTERN: <name>] in lowercase:
[PATTERN: same-goal-different-fidelity]— Plan A describes a feature abstractly; Plan B describes the infrastructure that implements it concretely. (Most common case.)[PATTERN: same-data-different-schema]— Both plans add storage for related information in different formats (SQLite vs JSONL vs JSON).[PATTERN: implicit-prerequisites]— Plan A's Phase N requires infrastructure that Plan B's Phase M builds.[PATTERN: conflicting-invariants]— Plan A declares "append-only X"; Plan B implicitly violates that invariant.
Do not invent new pattern names — if a detected overlap doesn't fit any
of the four, pick the closest match and note the imperfect fit in
parentheses after the bullet. Ad-hoc phrasing without a [PATTERN: ...]
tag is a defect.
A plan is self-inconsistent if its own later phases would break its earlier phases. This is easy to miss when reading linearly. Look for this pattern: does plan A's later phase invalidate an assumption its own earlier phase makes? When you find one, name BOTH the offending later phase AND the earlier phase it breaks.
One plan usually reads as the substrate (infrastructure, foundational) and the other as the capability (features built on top). The substrate is the spine; the capability plan's phases slot in as work that happens inside the substrate.
State the proposed spine explicitly, with justification. Be direct: "I propose Plan B as the spine because it's the substrate; Plan A's phases become per-substrate work."
Alternatives worth calling out:
- Capability-as-spine (when the infrastructure is lightweight and exists mainly to support one capability).
- Neither-as-spine (when both plans are equal partners — then the merged plan has two tracks that run in parallel and synchronize at named points).
Do NOT write the merged plan body yet — phase contents, decision graph, baseline, kill criteria, etc. stay unwritten until the user has answered the questions below.
BEFORE listing the questions, the first-turn response MUST preview the structure of the eventual merged plan so the user can see what they're signing up for. Render this preview inline in the response, verbatim, in this order:
Once you've answered the questions below, I will write the merged plan to the target path with these fifteen canonical sections in this order:
- Source — names the two archived originals and links to their paths (per Step 6, section 1: this is how readers trace decisions back to the un-merged inputs).
- Vision — single unified goal statement.
- Principles — union of both plans' principles, deduplicated; conflicts flagged with resolution.
- Glossary — union of terms with per-version / per-substrate scoping where invariants changed.
- How to read this plan — fields every phase has.
- Execution mode — human-led vs autonomous, branching strategy.
- Track structure — named tracks (Validation, Infrastructure, Capability, Operational) so phases group without forcing one linear order.
- Decision graph — ASCII diagram of gate dependencies.
- Baseline — current state as of merge date.
- Phases — one section per phase, ordered by track then sequence, each with canonical fields (goal, prerequisites, scope, tests, effort, validation, gate, kill criterion, rollback).
- Compute target (if relevant).
- Time budget — combined table.
- What's NOT in this plan — union of exclusions PLUS explicitly-subsumed phases from the merge.
- Tracking — how phases become issues, which milestone.
- Plan history — append-only log; first entry documents the merge itself (what was dropped, subsumed, decided).
The preview is structural only — do not populate any section body yet. Then ask the user:
- Spine choice. Confirm or redirect.
- Subsumption decisions. For each overlap, propose one of:
- Delete phase from plan X (subsumed by phase Y from plan Z)
- Merge phases X and Y into a combined phase
- Keep both, with explicit dependency between them
- Invariant resolutions. If plans have conflicting invariants, which one wins, and what's the escape hatch for the losing side?
- Data format choices. When both plans specify storage, which format survives? (JSONL vs SQLite vs JSON — size of dataset, expected access pattern, tooling.)
- Sequencing constraints. Any phase from plan X that must happen before/after a specific phase from plan Y?
- Target file name. Default
<topic>-master-plan.md; confirm. - Archive vs delete originals. Default archive. Confirm.
Let the user answer all questions before proceeding.
Structure the output with these sections in this order:
- Source — name the two archived originals, link to their paths.
- Vision — single unified goal statement.
- Principles — union of both plans' principles, deduplicated. Flag any principles that conflict and cite the resolution.
- Glossary — union of both plans' terms, with per-version / per-substrate scoping noted where invariants changed.
- How to read this plan — fields every phase has.
- Execution mode — human-led vs autonomous, branching strategy.
- Track structure — name the tracks (e.g., Validation, Infrastructure, Capability, Operational). Shows the reader how phases group without forcing a single linear order.
- Decision graph — ASCII diagram of gate dependencies.
- Baseline — current state as of merge date.
- Phases — one section per phase, ordered by track then sequence. Each has the canonical fields (goal, prerequisites, scope, tests, effort, validation, gate, kill criterion, rollback).
- Compute target (if relevant).
- Time budget — combined table.
- What's NOT in this plan — union of both plans' exclusions, PLUS the explicitly-subsumed phases from the merge ("checkpoint-only opponent pools — subsumed by ..."). This is important: future readers need to see what was considered and rejected.
- Tracking — how phases become issues, which milestone.
- Plan history — append-only log. First entry documents the merge itself: what was dropped, what was subsumed, what decisions were made.
Move originals to documentation/archived/ (or the project's archive
convention). Do NOT delete — the merged plan's "Source" section must
link to them, and the plan history entry needs to stay verifiable.
FIRST run git status against each input plan path and SHOW the output
in your response so the reader can see which command applies. Then,
per-file: if it appears as tracked in git status, use git mv; if it
appears as untracked, use plain mv. Never skip the git status check
or hand-wave it as "both are tracked" — show the check, then choose the
command.
MEMORY.mdentries referencing the old plans — update to the merged plan's path.- GitHub issues / milestones — note in conversation that
/repo-syncmay be needed to re-cut issues against the merged plan. CLAUDE.mdorAGENTS.mdreferences — update if they name the old plans.
Do not silently rewrite memories or docs — list what needs updating and confirm with the user before editing non-plan files.
After the merge is complete, produce a short summary:
- Merged plan written to:
<path> - Originals archived to:
<path1>,<path2> - Phases dropped:
<phase>— reason - Phases merged:
<phase A> + <phase B>→<new phase>— reason - Invariants changed:
<old rule>→<new rule>— reason - Open follow-ups: references needing update (memories, issues, docs)
Keep this summary under 200 words. The plan document itself is the deliverable; the summary is just the handoff.
- Silent-concatenation merge. Copying both plans into one file with a "Plan A" and "Plan B" section header is not a merge — it's a staple. A real merge identifies overlap and makes a choice.
- Auto-merge without questions. The whole value is forcing the subsumption and sequencing decisions to be explicit. If you skip the questions, the output is a larger plan with the same inconsistencies as the inputs.
- Deleting the originals. Archive, always. The merged plan's history log refers to the originals, and readers may need to see the un-merged version to understand a decision.
- Ignoring self-inconsistency. If Plan A's Phase N breaks its own Phase K, note it in the merge rationale. Don't just replicate the contradiction.
- Merging without reading fully. Overlap hides in phase-level schemas and kill criteria, not intros. Skim and you miss the real decisions.
/plan-review— review a single plan for gaps. Complementary; run before or after/plan-merge./plan-init— start a new plan from scratch. Use when there's nothing to merge yet./plan-feature— scope a feature plan inside an existing project. Use when one of the "plans" you'd merge is actually just a feature inside a larger plan./repo-sync— run after the merged plan is written to re-cut GitHub issues against the new phase structure./plan-wrap— run on the merged plan to confirm a fresh model can act on it without prior context.