cold-read

Cold Read (/cold-read)

A cold read answers one question about a draft: can the intended reader make their next decision from the first N lines, and what would they cut?

The reader runs in a fresh context. It must not see the conversation that produced the draft. That conversation is exactly what leaks into agent-written artifacts as process narration and as background the reader already has.

Invocation

/cold-read <path> [--role triager|owner|reviewer] [--n 3]

• path: the draft file, in the form a human will see it. Strip draft-only preview lines first; a leading # <Proposed Title> line stays and counts as line 1.
• --role: who reads it. Default triager. Use owner when the likely reader wrote or maintains the code under discussion (small repositories, your own packages). Use reviewer for pull request bodies.
• --n: lines the reader gets before it must decide. Default 3.

Steps

1. Resolve the role. Without --role, infer it: owner when recent commits on the paths the draft references come from one or two active maintainers; otherwise triager. Pull request bodies default to reviewer.
2. Spawn a fresh-context, read-only reader. Give it exactly three things: the contents of references/persona.md, the role and N, and the draft file path. Nothing else from the current conversation.
3. Require the JSON shape in references/schema.json. Re-run once if the reply is prose. Recompute verdict from decision_line and N yourself; the model's decision_line is the signal, its verdict is a convenience.
4. Act on the verdict.
   • decide_in_n: proceed.
   • decide_later or cannot_decide: move the sentence that carries the decision to the first body line, apply cut_list (keeping any sections required by the repository's PR template), add any missing facts, then re-run once.
5. Report one line in chat: verdict, decision line, words above the fold, and the number of cut lines. Do not restate the draft.

Model choice

A fast model (flash) is sufficient for single-draft gating once references/persona.md anchors decision_line to the first line of the block; both flash and the default model score the regression fixtures identically. For batch baselines over multi-section artifacts, prefer the default model when quota allows: in a 10-artifact calibration sample, flash placed decision_line earlier on 2 of 10 drafts.

Batch mode

For baselines over many past artifacts, run one reader per artifact in parallel, each in its own fresh context, and tabulate verdict, decision_line, words_above_fold, the known_to_reader count, and missing. Never give one reader two artifacts; the second read is primed by the first.

What the reader is for

The reader targets three failure modes of agent-written text:

• Discovery narration: leading with what the agent found, in the order it found it.
• Background the reader already has: a tour of the reader's own code, including a "currently, X does Y" opener.
• Template slots filled for completeness rather than for a decision.

It does not judge style, tone, or grammar.