envoy

envoy

!checklist init ${CLAUDE_SKILL_DIR} --force

The analysis is settled; the claims have verdicts. envoy is the lens you hold over the paper — the step that carries the finished work out to its community, defends it before the court of peers, and keeps going up the ladder until it lands. It is the sixth and last skill of the inquiry suite: it owns step six of doing computational research — writing the paper, submitting it, defending it through review, and resubmitting after a rejection. It audits (and guides you to run) a gated pipeline, and it will not advance past a GATE until the checklist tool clears it. That gate enforces order — each step done before the next — not the substance of the work inside it; the tool structures the discipline, it does not audit it, so the rigor is yours to supply.

The one mental shift everything hangs on — the paper already exists, and this is the step the human owns most. If the first five steps were done with discipline, the paper is not a blank page: the claims are its argument, the frozen protocol is its experiments section, the mechanism probes are its discussion, the comparison table is its related work. Writing is stringing an existing argument into prose — far less work than most people fear. And precisely because the raw material is done, the remaining work is the most human in the whole pipeline: the story line, the claim wording, the contribution, the choice of venue, the judgment in a rebuttal, and the signature on the claims are all irreducibly yours. The arc is: lay the figure-first skeleton → draft by template → vet for evidence and legibility → submit to the right venue → rebut → persist up the ladder.

The agent is the means, not the author of the argument — and here the line is bright. The agent's legitimate, high-value jobs are real: language polish and page-limit compression, the consistency vet (symbols defined before use, terminology unified, figure numbers aligned with their references — the mechanical errors the human eye misses and the agent almost always catches), the reader test (a clean session given only the paper, asked to restate the claims and each figure's point — its failures are the exact coordinates of where you wrote unclearly), and the agent era's one real change to rebuttal: running the new experiments that forge's one-command regeneration now makes feasible in days. But the cardinal sin is sharp and absolute: the agent must never write a sentence whose evidence it has not seen — every number, every "experiments show", traces to a results-store run id, because an agent helpfully inventing a plausible number that slips into the text is among the most lethal submission accidents there is.

What you cannot delegate — four things, and they are the whole point. Across the entire suite the human keeps exactly four: taste (what to work on), spec (what counts as success), judgment (what to bet on, what to abandon, how to fight a rebuttal), and responsibility (the signature on the claims and their proofs). envoy is where all four come due at once — the story is taste, the contribution is spec, the rebuttal strategy is judgment, the byline is responsibility. Everything else is discipline and the agent. And all of that discipline guards a single thing: the you who saw the results must not get to deceive the you who made the decisions.

What "done" looks like — out the door and through the gauntlet, not "the draft is finished". The step is over when the paper is submitted, the reproduction package is finalized (anonymized, licensed, smoke-tested clean), the rebuttal is fought, and — the common outcome — after a rejection the paper is revised and resubmitted one rung down the ladder, or, on acceptance, the camera-ready checklist is closed. "The draft is finished" is not the terminus; a paper that has run the gauntlet is.