Procedures
A procedure is how your project does one kind of job. Quirework makes procedures findable and followable, and supplies the mechanics for the five that recur when a work is written with more than one model.
What a procedure is
A procedure is a Markdown file in the project: how this work drafts a chapter, checks a claim, researches a source, or makes an edition. A catalog says what each one requires reading first, which tools it uses, and which other procedures may follow it.
quirework sop list
SOPs by task (existing names; no workflow is launched):
Authoring — What do I do when I sit down to write?
write-a-chapter: Write or revise a chapter of Side Quest, from the primer through the gates.
Editions — How do I make another shape of the finished book?
variant-common [protocol]: what every edition of Side Quest keeps
variant-poems: Make the poems edition of Side Quest, one poem of light verse per chapter, drafted blind on Codex and composed against the spec.
variant-six-pager: Make the six-pager edition of Side Quest: the book as one narrative memo for Hedda, drafted blind and composed against the spec.
variant-young-readers: Make the young readers' edition of Side Quest, for readers of about ten who read well and alone, drafted blind chapter by chapter and composed against the spec.
Use sop NAME to read a procedure; sop explain NAME to see its dependencies.
sop explain prints a procedure’s required reading in dependency order,
with its length, so the companion knows what to read before it starts.
Follow-up procedures are listed and never launched. Whether one is
triggered is the operator’s call.
quirework sop explain write-a-chapter
write-a-chapter: Write or revise a chapter of Side Quest, from the primer through the gates.
Required SOP reading, dependencies first (not an execution plan):
procedures/workflows/write-a-chapter.md (420 words)
Total: 420 words
Mechanical tools: quirework spec notes --primer, quirework spec trail, quirework spec trail add, quirework spec units --check, quirework spec notes --check, quirework spec trail --check, quirework config check, quirework dev check
Conditional follow-ups (operator decides whether triggered):
None cataloged.
SOP reading dependencies only; root guidance, directory cards and task-specific inputs still apply. No follow-up is executed.
sop skills generates thin entry points for the companions, so you can
name a procedure in conversation and the companion opens it. The entry
point only routes. It opens the procedure through sop show, which
prints it verbatim and records that it was read, and the procedure stays
the one place its content lives.
Five that recur
Each project owns its procedures and its editorial rules. Quirework supplies the mechanics for five patterns, and what sets them apart is who sees what. Each pattern below closes with how the sample book used it, marked as such, since that book and its author are invented.
| You want | Use | Each seat sees |
|---|---|---|
| Real alternatives to choose between | A prose workshop | One frozen packet, never another writer’s draft |
| An answer argued out through disagreement | A writers room | Its own packet, then the shared board between rounds |
| Several models working together, as themselves | A roundtable | The shared packet, every turn as it happens, and a shared folder |
| What the page tells a reader | A blind read | Only the material a reader would see |
| An edit tested against the whole work | A consistency review | The guide, the spec, the notes and the exact change, read-only |
Prose workshop
Several writers solve the same bounded problem, each blind to the others. Every writer gets the same frozen packet in a private workspace, without the project’s instruction files or the companion’s memory, so the experiment tests the brief rather than the habits of whoever prepared it. The drafts are sealed under letters before anyone reads them, and the writers’ identities open only after the editorial judgment is recorded.
In the sample book, the first chapter came from a workshop of three
writers on three different models. The invented author read A, B and C,
chose A, took one paragraph from B, and only then opened the key. The
sealed drafts and the key are in the book’s provenance/workshops/.
Writers room
Several seats argue one question over numbered rounds and read each other between them. Each seat keeps its own session, so a later round resumes with everything it has read and reasoned, and is handed only what changed. A seat writes to its own outbox, and promotion copies each post onto the board, hashed and attributed. The digest of a round concatenates its posts whole and never summarizes them.
In the sample book, the fourth chapter was argued by a room of four
seats on three harnesses over six rounds, with the author in the chair.
Each seat held different material on purpose. The seat arguing for Otto,
the assistant the book could have made a villain, was given nothing
beyond the brief, so that the best defense came from a seat that had not
been told what the book thinks. The plan, the personas and the board are
in provenance/rooms/.
Roundtable
A roundtable seats models as themselves. Each is named for its model family, told which others are at the table, and told that the supervisor reads everything. Turns are sequential, so a seat reads what the one before it wrote minutes earlier, and every seat has the same tools and a folder the table shares. A person can take a seat as well, described in the roster as a person, never as one more model.
In the sample book, the back cover was written at a roundtable of three
models, and the text that went on the book was the author’s choice. The
board, the drafts and each model’s final text are in
provenance/roundtables/.
Blind read
A blind reader meets the work as a reader would: no spec, no notes, and no author’s explanation, unless the question requires them. The absence of that context is the test. A reader can be handed one chapter per launch, which freezes exactly the bytes it saw, or walk the whole book with its own tools, writing a reaction to each chapter before it opens the next.
Blindness does not make a reader right. It makes the reaction answerable to the material the reader actually saw.
In the sample book, the whole book was read by two readers on two
harnesses, each alone, with a closing question shown only after the last
chapter. Both asked for gaps to be filled, and neither asked for more
pages, which settled that the book was the length it wanted. Each
reaction is in provenance/readings/.
Consistency review
Consistency has a mechanical layer and a judgment layer. The checks cover what can be stated exactly. The review asks whether an edit that is coherent on its own is coherent with the rest of the work. An independent reviewer gets the project guide, the spec, the notes, a brief and the exact Git range, with the repository read-only and its history in reach, and without the operator’s saved conclusions.
In the sample book, when the author corrected the story of the name, a
review of the change found two problems. The young readers’ edition now
defined “quire”, which its own notes say to show in a sentence. Two new
lines of the poem did not scan in the metre its procedure assigns. The
reviewer’s reports are in provenance/reviews/.
Models and harnesses
The companion is the operator, and a model seat is something it convenes. One procedure can mix seats from different makers and harnesses, and a person. Preparation refuses a capability the selected harness cannot provide. The receipt records the boundary that was actually enforced or audited, the effort the model actually ran at, and the model the provider actually served.
Runs you can trust later
A run is a plan of seats and ordered turns, with an explicit model and effort for each. Preparation freezes the prompts and inputs and hashes them. Launch intent is recorded before a process starts, and a pending request is reconciled before anything is resubmitted, so an answer is never bought twice.
A run stopped by a quota limit continues in its own session instead of starting over. A turn cut off by an error waits for your decision, with its session intact. A compaction is reported even when the answer looks fine. Earlier answers and rejected attempts stay exactly as they were.
Reading the record
Captured sessions support procedures of their own. provenance search
reads your own turns, by keyword or by meaning. input compiles your
verbatim words. metrics estimates the division of labor between you
and the machine, with its caveats attached. coldread puts the record
in front of blind analysts and caches their reports by content, so a
changed question misses the cache on purpose.
The workflow guide covers each pattern, run plans, and recovery in full.