Stage one: structural checks
Click “⚙ Audit + AI review” and the first reply is instant, because it is not AI. Runic parses the process and walks it, and posts “Structural checks on ‘<document>’ — N findings. Now reading it like an analyst…” or, when nothing is wrong, “— clean: decisions have branches, every step is reachable.” Each finding shows its line as a chip (“L12”); click it and the editor jumps to that line and flashes it. Findings are sorted by line; amber ones are warnings, grey ones are hints.
The warnings, with the message each one posts: single-branch-decision — “This decision only covers ‘<label>’ — what happens otherwise?” (or “has no branches — what happens in either case?”); empty-branch — “The ‘<label>’ branch is empty — what actually happens on this path?”; identical-branches — “This branch is labeled ‘<label>’ like the one on line N — name each branch by its own answer.”; unreachable — “The flow leaves this sequence at ‘.go X’ before it ever gets here — should this line move up or be removed?”; dead-back — “‘.back X’ points at no step — the target must be a fragment of an existing plain step or .input line, without its ‘: caption’.”; ambiguous-back — “‘.back X’ matches N steps (lines a, b) — the arrow lands on line c; copy a longer fragment that names only the intended step.”; input-caption — “The options after ‘: ’ became the arrow’s caption, not the field — write them in parentheses inside the field text.”; orphan-end — “This .end is indented like a .then, so it only ends the ‘<label>’ branch — the decision on line N is never closed. Outdent it to the .if’s indent.”; guard-fallthrough — “The ‘<label>’ branch ends with a retry .back to ‘<step>’, so the flow continues below the .end into ‘<next>’ — the guarded work still runs. End this path instead (.go, or .back to a later terminal step) or nest the steps below inside the other branch.”
The hints: not-a-question — “‘<condition>’ does not read as a question — phrase the decision so each .then is an answer to it.”; top-guard — “‘Who may run this document’ is a .note on the first line, not a decision at the top of the flow — start at the actor’s first real action.” (a first .if that asks who may run the flow, with two branches one of which only refuses); reads-like-two-steps — a step over 45 characters joined by “and”, “&” or “, then”; and no-failure-path — “Nothing in this process branches or fails — does every run really go perfectly?” for a process of eight or more steps with no .if and no .fail. That last one is about the whole document, so it shows a dot instead of a line chip.
Stage two: the analyst
Structure is half an audit. The second stage sends one chat message with a fixed brief: “Review this process as a senior business analyst who must sign it off. For each finding: the exact line text it concerns, what is missing or risky (undocumented failure path, unclear role/permission, data collected but never validated, threshold or SLA absent, a step the brief implied but the document lacks), and the concrete line(s) you would add. Stay specific to THIS document; no generic advice. Answer in the document’s language.” The recent conversation and the archive summaries travel with it, so the analyst knows what you have already discussed.
This call is sent with two explicit flags: the heavy tier on, so it runs on Claude Sonnet 5 with thinking, and web research off. A private process is never sent to a search engine because the word “compliance” appeared in it, so the review never ends with a Sources line. If the call fails you see “⚠ Deep review failed: <reason>”.
It also runs by itself
Every time the AI writes into a process — an applied proposal, a Create with AI draft, the main flow of a generated system, an imported flow, an accepted inline edit — the structural pass runs automatically. If it finds anything, a message appears: “Structural check on ‘<document>’ — 2 issues in what was just written (run ⚙ Audit + AI review for the analyst pass):”. If it is clean, nothing is posted. The automatic check reports; it never edits.
The same checks guard the AI itself
Nine of the codes are hard: single-branch-decision, empty-branch, unreachable, dead-back, identical-branches, orphan-end, guard-fallthrough, input-caption and ambiguous-back, plus unknown-reference when a .do or .go names no document in your workspace or the generated set. Parse errors count ten each, hard findings one each, and that score is what Runic uses to self-repair AI output before you see it. A document that scores above zero goes back to the model once with a validator report quoting each line and a rule per code — add the missing branch, retarget the .back, move the options into parentheses, end the failing branch with .go instead of a .back — and the fix is kept only if it scores lower, adds no parse errors, was not cut short and still contains every original line verbatim. Bare .end lines may go; a .do may become a .go or be renamed to a known name; a line the report named may only turn into the rewrite its code allows.
For chat edits your document is the baseline: findings already present in your text are your structure and are left alone. Sequence diagrams have their own set — a request without a reply, an activation never closed, an .alt or .par with one section, an empty fragment, a lonely actor, a silent receiver, a diagram of eight or more messages with no failure path, five or more actors without pins, an unknown reference — used by the repair pass and by audit_document over MCP. When the fix was kept the reply carries “self-checked” under the message. So by the time you press Audit, the obvious faults have usually been caught upstream, and what remains is the analyst’s job.
Limits
The button exists for processes; it is hidden on sequence diagrams, whose audit runs over MCP and inside the repair pass but has no button yet. Findings are conversation starters, not errors — syntax problems are the parser’s job and show in the editor. The analyst pass counts as one chat message against the 60-an-hour limit and needs a verified email. A .do or .go that points nowhere is the editor’s job too: its Quick fix chip creates the missing process with one click, which is what the self-repair’s unknown-reference code guards against on the AI side.
How to use it
- Open a processThe Audit button only appears for process documents.
- Click “⚙ Audit + AI review”It sits at the bottom of the chat panel, in the row just above the message box, next to 📎 Import flow. The tooltip reads “Instant checks plus an AI analyst pass”.
- Work through the structural listThe first reply lists findings with L-number chips. Click one to jump to that line in the editor; amber is a warning, grey a hint.
- Read the analystThe second reply arrives after “Reading the process… Thinking it through… Writing…”. It names the exact line text, what is missing, and the lines it would add.
- Fix from the chatReply with what to change — “add the missing No branch on line 12” — and apply the resulting diff. The automatic check then confirms the structure held.
Commands & shortcuts
| ⚙ Audit + AI review | run both stages on the open process |
| L12 chip | jump to the finding’s line in the editor |
| warn / hint | amber findings are structural gaps; grey ones are style hints |
| single-branch-decision · empty-branch · identical-branches · unreachable | decisions and reachability |
| dead-back · ambiguous-back · input-caption · orphan-end · guard-fallthrough | the codes added in v10 and v10.3 — all hard codes the AI must fix in its own output |
| not-a-question · top-guard · reads-like-two-steps · no-failure-path | hints |
Tips
- Run the audit before sharing a process; a clean structural pass and a short analyst list read well in a review.
- Answer the analyst in the same thread. “Add a .fail for the payment timeout you mentioned” turns a finding into a diff.
- A guard-fallthrough finding means a failing branch loops back above the decision and then keeps going — end it with .go, or move the steps below into the other branch.
- An ambiguous-back finding tells you which line the arrow currently lands on; copy a longer fragment of the step you meant.
- The automatic check after an applied change is the same stage one. If it is silent, the structure held.
- The structural pass is also exposed to Claude over MCP as audit_document, for processes and sequence diagrams alike.
Questions
Are audit findings errors?
No. They are questions about the process — a decision that only covers one outcome, a step nothing can reach. The parser reports syntax errors separately, in the editor.
Does the analyst search the web?
No. The deep review is sent with research switched off, so a private process never reaches a search engine and the review never carries a Sources line. Ask a separate question in the chat when you want researched context.
Does the audit change my document?
Never. Both stages only post messages. Changes happen when you ask for them and click Apply.
Can I audit a sequence diagram?
Not from the button, which is hidden on sequences. The sequence checks exist — request-without-reply, unclosed-activation, single-section-alt, single-section-par, empty-fragment, lonely-actor, silent-receiver, no-failure-path, unpinned-actors, unknown-reference — and run in the AI’s self-repair and through audit_document over MCP.