Mariomarquezt/multi-flow-skills / skills/multi-plan
multi-plan
UnverifiedMulti-plan
Model names are short names. The exact version of each lives in ONE place:
~/.claude/skills/_shared/models.md. Read it before the first dispatch and write the exact model id in the final layout. Grok here = the Grok row (4.6, reviews).Gemini (Antigravity, added 2026-11-17): Gemini Hydrogen / Carbon / Argon / Flash are in models.md as extra lanes — no mode uses them until the user moves them in. Swap one in only when the user asks. Runner:
$RUNNERS_DIR/antigravity/agy-run.sh(read-only: critics never get--write). One Gemini lane never checks another Gemini lane's work; Gemini Flash on Antigravity and on Cursor is one model.Model rules (repo owner): every Claude part runs on Opus 5.5 — never Fable, never Opus 5 or older. Never OpenRouter unless the user names it or approves it for THIS task. State both rules in any handoff this skill writes.
One idea. I write the plan. The other models attack it. I decide what changes. Up to three times.
The one idea
A plan written by one model carries that model's habits: its favourite stack, its blind spots, its guesses about prices and limits. Other models, on other bills, trained by other labs, attack it independently — none sees another's answer in the same round — so their agreement means something. The session model stays the author and editor: it does not obey the critics, it rules on each point in writing, and a critic that loses can appeal once with new evidence. The loop stops when the critics run out of serious objections, not when a round count is hit.
The session model never critiques its own plan, and no critic is Claude.
Modes — parse the invocation FIRST, state it back in one line
| invocation | critics | max rounds |
|---|---|---|
/multi-plan 2 | Sol (Codex), Grok (Cursor) | 3 |
/multi-plan or /multi-plan 3 | Sol, Grok, DeepSeek Flash (provider picked) | 3 |
/multi-plan 5 | Sol, Grok, DeepSeek Flash, GLM Flash, MiniMax Flash | 3 |
/multi-plan fast | Composer (Cursor), DeepSeek Flash | 1 |
… rounds N | — | N (1 to 3) |
… dry | none | 0 — draft v1, show it, stop |
… report | — | — also publish the HTML report (Step 7) |
Compose freely: 5 rounds 2 report, fast report.
Independent eyes. Two routes of ONE model count as one critic. Grok and Composer share the
Cursor pool, so fast pairs Composer with DeepSeek Flash, never with Grok. DeepSeek Flash and GLM
Flash may land on the same provider; that is fine for plan review (they are different models) but
write the provider of each in the decision log.
Router (suggest mode). Before Step 3, run:
$RUNNERS_DIR/router/router suggest --task "<one-paragraph idea summary>" --items 1 --mode fix --n <critics>
(The router has only fix and build modes — no review. fix is the closest: it ranks
arms by how well they find and judge problems.)
State its pick next to the mode's default in the mode line. Keep the default unless a default
arm's plan is BLOCKED or near empty — then take the router's substitute and say so. If the router
errors, carry on with the default and note it. After the last round, log one outcome per critic
(router outcome --id … --arm … --role reviewer --result pass|fail, --found <accepted findings>).
Where the files go
- In a project folder:
docs/plans/<YYYY-MM-DD>-<slug>/. - Otherwise:
~/Desktop/Claude/plans/<YYYY-MM-DD>-<slug>/.
plan-v1.md, plan-v2.md, … each version, never overwritten
round-<R>/brief.md the exact brief the critics got
round-<R>/<arm>.md each critic's raw answer (+ .transcript)
decisions.md the ruling on every finding
plan-final.md the agreed plan (copy of the last version, plus §11 below)
Create a todo per step and work them in order.
Step 1 — Intake
Read the idea. If anything the stack depends on is missing — who uses it and how many, budget, hosting the user already pays for, systems it must fit with, deadline — ask one message, at most 4 questions, each with my proposed answer so a "yes" is enough. Read the project's STATUS.md / CLAUDE.md first if the idea lands in an existing project; do not ask what the files already say. Do not ask about taste the critics can argue.
If the idea is tiny (one script, one page, one obvious stack), say so and offer a plain plan in the session instead. This skill costs several model runs.
Step 2 — Draft v1
Write plan-v1.md from templates/plan.md. Fill every section. The stack table names a
runner-up for each layer and why it lost — this gives the critics something concrete to attack.
Check facts the picks rest on (current versions, prices, free-tier limits) with a quick web
search when unsure, and tag each row [fact] or [hypothesis]. Never write a credential into
the plan.
dry mode: show v1 and stop.
Step 3 — Critic round R
-
Copy
templates/critic-brief.mdtoround-<R>/brief.md. Paste the WHOLE current plan and, from round 2 on, every row ofdecisions.mdso far. Fill round and max. -
Run every critic in parallel, read-only, one background Bash call:
S=~/.claude/skills/multi-plan/scripts/critic.sh; D=<plan dir>/round-<R> for a in <arms>; do "$S" "$a" "$D/brief.md" "$D/$a.md" & done; waitArms:
sol grok deepseek glm minimax composer qwen mimo. The script picks the provider for multi-provider models (tokens that expire first), never uses OpenRouter, and never uses Devin unlessPROVIDER=devinis set by the user. Expect 2–10 minutes; run it withrun_in_backgroundand wait for the notice — do not poll. -
A critic that fails (non-zero exit, empty file, no
VERDICT:line): rerun it once. If it fails again, go on without it and say so in the round summary. Never fill its place with a Claude subagent: Claude is the author.
Step 4 — Rule on every finding
Read all answers. Then, in decisions.md:
- Merge duplicates. Same point from several critics = one row, all ids listed,
backers= how many independent critics raised it. - Rule on each row:
- ACCEPT — change the plan as asked.
- PARTIAL — change part of it; the reason says which part and why not the rest.
- REJECT — one-line reason that answers the critic's
why, not "I prefer X". - DEFER-TO-USER — a real trade-off that depends on the user's priorities (cost vs speed, lock-in vs convenience). Goes to "open fights".
- Weight rules:
- A point raised by 2+ critics is ACCEPT unless I can state a concrete reason it is wrong.
- A BLOCKER is never rejected without a written reason that a reader could check.
- A
confidence: lowfact claim is checked (web search) before it changes the plan. - An APPEAL with no new argument is REJECTED as "no new evidence"; with new evidence it is ruled again from scratch.
- A finding that breaks a constraint the user set is DEFER-TO-USER, not ACCEPT.
- Tell the user a 3–5 line round summary: verdicts (e.g. "Sol CHANGE, Grok SHIP, DeepSeek CHANGE"), accepted / rejected / deferred counts, the biggest change.
Step 5 — Revise
Write plan-v<R+1>.md with every ACCEPT and PARTIAL applied. Fill "Changes from the last
version", one line per change with its finding id. Do not sneak in changes no finding asked for;
if I spot something myself, add it as my own row in decisions.md (id R<R>-AUTHOR-n).
Step 6 — Loop or stop
Stop when ANY of these is true; otherwise go back to Step 3 with the new version:
- every critic said
SHIP; - this round produced no accepted BLOCKER or MAJOR (only MINORs, or everything rejected);
- the round cap is reached (3 by default).
Say which rule stopped it.
Step 7 — The agreed layout
Write plan-final.md = the last version plus a section 11. How this plan was agreed:
critics and exact model ids, rounds run, the stop rule, counts of accepted / rejected / deferred,
and any critic that failed.
Then show the user, in this order:
- The stack — the final stack table (layer, pick, one-line why).
- The architecture and phases — short.
- What the critics changed — the 3–7 biggest changes from v1, each with who pushed it.
- Where they still disagree — rejected points that 2+ critics still backed at the end, with my reason, so the user can overrule me.
- Open fights — your call — every DEFER-TO-USER row: the trade-off in one line, my pick. At most 3; if more, keep the 3 with the biggest cost or lock-in effect and list the rest by id.
- Next step — the command to build it:
/multi-build(several parts) or/codex-build(one builder from a frozen plan), with the path toplan-final.md.
report mode: load the artifact-design skill and publish an HTML page with the same six parts,
plus a round-by-round table of verdicts and the full decision log. Private by default.
Step 8 — Close
- Log router outcomes (see Router above).
- Write the vault session log per CLAUDE.md (what was planned, the final stack, open fights).
Honesty rules
- Never report "the models agreed" on a point a critic disputed. Agreement = no live objection.
- Quote critic positions from their files; do not paraphrase a REJECT into a soft "noted".
- If fewer than 2 critics answered in a round, say the round is weak and offer to rerun.
- The plan is a plan: prices, limits and versions tagged
[hypothesis]stay tagged in the final.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |