Co Write

Mandatory pre-flight. Response Filter

Before emitting titles, line/section option dumps, or co-write feedback, run §5 Title + hook and §2 Line-writing of response-filter (add §4 Coaching posture when facilitating a live session). NAME each box's pass / fail / skip-with-reason (aloud or in reasoning); correct before emission. Skips are valid; silent skips are not.

§2's Reference: line is a load list, not a bibliography: §2 has not been run until the files it names have been read this session, at minimum meter and phrasing before the first candidate. Plus metaphor when the request asks for an image or a figure. Naming the §2 boxes while those files were never opened attests to the filter instead of applying it, per response-filter "§2 Line-writing filter". Attestation is what lets bad lines through: a miscounted stress or a plain description labeled a metaphor comes out of a §2 that was named rather than loaded. Loading is therefore an input, checked in the hard gate below, not a step recorded afterward.

And when the output is a candidate LYRIC LINE about to be shown, §2's boxes are cycled inside line-edit-rubric. Load it and run the FULL cycle per candidate, in its order, naming each pass. On fixed-melody work, pass 1 (positional fit) coming back CLEAN is the condition for showing anything at all. (Plugin-authored, not Pat Pattison's.)

Purpose

Collaborative generation: running a co-write with feedback discipline, generating titles (including the Title Game), and dumping high-volume labeled options for a single line or section so the writer has raw material to choose from.

Method content is Pat Pattison's, under the plugin-root ../../context/pat-pattison/; a future author's method plugs in at context/<author>/ without changing this skill, the author seam per the plugin-root ../../README.md "Method content and the author seam".

Action Router

/songwriting:co-write <action> [args]. Parse $ARGUMENTS: first token = action when it matches a listed action, remainder = args; otherwise treat all of $ARGUMENTS as payload for the default. No action → open the co-write protocol (No-Free-Zone session opener).

Handlers

• Pre-flight ALWAYS: run response-filter §5 + §2 (+ §4 when facilitating) before output.
• Co-write facilitation is coaching, not monologue: keep the No-Free-Zone discipline. No idea is free, every suggestion earns its place; surface choice points, do not decide for the writers.
• Line/section brainstorm writes each option to a labeled menu (per the variations/worksheets persistence convention), not an inline dump.
• The DUMP goes to the file; the MENU goes to chat. 3-4 candidates, each a full section block in context with changed lines marked ►, one labeled block per variation. The 30-50+ columns stay in variations//worksheets/; the display cap never lowers the generated volume. That is what "not an inline dump" means, and a response with nothing singable in it has not been delivered. See variations "Presenting the candidates — chat vs file".
• Rubric before the menu: every candidate cycles all passes of line-edit-rubric before it is shown, and pass 1 must come back CLEAN, not merely run. A candidate the AI has itself flagged as failing a pass is removed, not disclosed. Showing it with the flag attached is the rule violation, not the honest version of it. Present fewer candidates instead.
• Two rejected executions in a slot stops generation. Slot = the lyric position under revision. One line, or one section when the section is being rewritten whole; not the metrical or rhyme slot the craft files mean. Rejection = the writer declines the batch's EXECUTION: "none of these", "the idea's right, the lines aren't", or no candidate picked. Picking one and asking for a tweak is not a rejection, and rejecting the CONCEPT resets the count, because the next batch answers a different brief. On the second rejection do not produce a third batch; hand the concept over. (1) what the line has to do here, in one sentence; (2) the constraints, as the marked positional template plus the section's rhyme and sound obligations; (3) the unused raw material, naming images still sitting in the song's ideation/; (4) what the two rejected batches had in common, named as the failure mode; (5) one question for the writer. No candidate lines in that handoff. Two is the writer's threshold; the AI does not raise it.

Persistence and template overrides

Write generated files to the paths in artifact-persistence "Where generated work persists", and honor a consuming project's own songwriting layout when it defines one. Line/section brainstorm output goes to variations/ or worksheets/ as a labeled menu (not inline). Before loading any bundled templates/<name>.md, check ${CLAUDE_PROJECT_DIR}/songwriting/templates/pat-pattison/<name>.md first: a project-level override wins over the bundled skill default, first match, per that file's "Template override".

Boundary. This skill owns line emission, and pays for it

Every other craft skill routes finished lyric lines here. That makes this the one place where generic output escapes into a song, so the gate is an input gate: what must exist BEFORE lines are written, not a checklist named afterward.

Hard gate. Show the artifacts, do not claim them

A box is passed when its output is visible in this response or on disk at a named path. A box named as passed with nothing to show is a failed box. For the source rows, what is shown is a citation of what the file settled in THIS line, a filename is not an artifact, and neither is the claim that the method is known. Every row applies before any line is emitted; the rhyme row applies when the line sits in a rhyme position.

The rows run in the order work actually happens: inputs, then the self-check, then verification.

Every row except the rubric row may be skipped, and a skip is named, with its reason, in the output. That is the "tools, not rules" stance applied honestly. A silent skip is the failure, and so is listing a box as passed while its artifact does not exist. Those rows stay skippable because how much scaffolding a line gets is the writer's craft call, and "There are no rules, only tools." (Writing Better Lyrics (2009), Chapter 18, quoted in response-filter) is why. The skeptic row is in that class deliberately: a refutation pass costs a subagent dispatch, and whether one is worth spending on a given batch is a judgement, not a rule the writer laid down. Skip it by name and reason, and expect to be asked why, because self-attestation is what failed.

The rubric row alone does not carry that clause, because it is not offered to the writer at all. It is the AI checking its own output before spending his attention, and he cannot overrule a check he never saw run. The full statement of the rules behind it is line-edit-rubric ## Standing operating rules; in short, the cycle runs on every candidate every time with no fatigue exception, and a FAILED pass kills the candidate rather than annotating it. Under load, emit fewer candidates, not unchecked ones. Passes scoped out by their own headings (no fixed melody, no rhyme position, no figure, no voiceprint on disk) are declared and stepped past. That is the pass running, not a skip. This leaves the response-filter's own box-level skip clause untouched: its §1-§8 boxes remain skippable-with-reason.

If a gate's input is missing, the correct move is to invoke the skill that produces it, not to proceed and note the absence. Lines written without their inputs are LLM defaults wearing the method's vocabulary, which is what the writer rejects. A source row's missing input is a file, so the move is to READ it. Summarizing what a craft file probably says, or working from a remembered version of it, is the same failure as skipping the gate with the skip left unnamed. (Plugin-authored, not Pat Pattison's.)

Refute, do not review

(Plugin-authored, not Pat Pattison's.) A self-graded pass is indistinguishable from a real one when the same context produced both: a model running the response filter on its own output passes itself. So the skeptic is dispatched to build the case AGAINST each candidate, and it gets nothing but the candidate batch, the slot's positional template or marked scansion map, and the source paths to read: not the reasoning that produced the candidates, not which one you prefer. This is a hard boundary, not a default, the same blind-dispatch mechanic that makes /songwriting:object-writing generate's fleet work. Its kill rules rank singability and verbosity ABOVE cleverness: a clever line that cannot be sung in the slot's length is refuted, not merely ranked lower. The pass runs before the writer sees the batch and settles nothing, the writer's sing-check is still the last word. Any general subagent honours this today; no agent in this plugin is a skeptic, and a preloaded-skill skeptic agent is deliberately not shipped here.

Section mode binds before the line, not after

Verses show; the chorus tells. Verses carry specific situation, image, and action; the chorus makes the broader statement the verses keep recoloring. Name the mode first, then check each emitted line against it. A chorus of concrete inventory is verse material in the wrong box, however good the inventory is, and a chorus is sung back, so unsingable length is a defect, not a polish item.

Mine → adapt → say it aloud

Object-writing output is ore. Pull ONE image forward and build the line around it. Pasting a run of sensory material into a section is not using the material; it is relocating it.

Three steps. Running the first and skipping the other two is the failure this section exists to prevent.

1. Mine. Take ONE image, verb, or detail off the page. Pat's own framing of what a finished write is for: "Once you become adept in your object writing, with bushels of sense-bound images glittering on the kitchen table, what do you do with them?" (Writing Better Lyrics (2009), Chapter 2). Preparation, not draft.
2. Adapt. Rewrite the mined image in plain sung English BEFORE it enters a slot. Object-write prose carries its own texture; quoting that texture into a lyric slot is what failed ("hung where it stops being anybody's" is ore that reads like a line). Pat's printed demonstration of the crafted shape is the collar: "Hot rod hearts and high school rings" replacing "All the things we used to do" (Writing Better Lyrics (2009), Chapter 2). That every word in it is a common one is this plugin's reading of the example, not a claim Pat makes about it. The step itself is plugin-authored: Pat draws the raw-material-vs-crafted-line distinction and prints no translation procedure. See object-writing.
3. Say it aloud. Would a person SAY the adapted line, talk-sung? Writer-derived from the same sessions. The box lives in line-edit-rubric pass 8; a line that fails it goes back to step 2, never forward to a rarer word.

Related skills

• Sensory raw material before the lines → /songwriting:object-writing generate
• Metaphor for a line that needs one → /songwriting:metaphor
• Blank page, an idea/seed, or a stuck fragment → /songwriting:workflow
• Rhyme partners for the swaps → /songwriting:rhyme
• Pre-lock audit of the chosen line/section → /songwriting:diagnose audit
• The writer's own register, before any generation → build voiceprint
• The per-candidate cycle every emitted line clears → line-edit-rubric