frame-next
Take a paper's open questions from `fli decisions` and make each one decidable — lettered options and a recommendation written into its own bullet — without ruling on any. Use when asked to frame the
SKILL.md.claude/skills/frame-next/SKILL.md
GitHubFrame next
The loop from an open question to a question the owner can answer with one letter. Arguments: a paper id and the question ids to frame in it (scripts/frame-loop.mjs passes both, with a ## Pre-brief), or one question id.
The session's job ends at the options. The ruling is the owner's, made with fli decide. So the only file written is the paper holding the questions, and only under each question's bullet. DECISIONS.md is left as it is, no question is struck, and fli decide is not run. A session that rules has failed, however good the ruling.
1. Pick
fli decisions --open lists what is open, --json the same as data. Say the paper and the question ids before doing anything else, so a person watching can redirect.
If the paper is dirty in git status and the arguments do not say an earlier frame-loop attempt left those edits, another session holds it: end with busy for every question. When they do say so, the edits are yours — read git diff on the paper and build on them.
2. Brief
The pre-brief carries each question verbatim, what it cites, the paper's sections with line ranges, and the rulings that share its terms. Read the paper's sections a question leans on, never the whole paper blind. Ground each option in the tree: rg for the code it would change, fli outline <file> before reading a big one, and a cited ruling read in full in DECISIONS.md. A search across many packages goes to one Explore subagent, asked for file:line answers only.
3. Frame
Run decision-rules on each question before writing its options — framing is choosing which alternatives are real, and that is a judgment call.
Options are real alternatives a reasonable owner could pick, each stated as what the tree would look like after it, and grounded: a file:line, a ruling, a measured behavior. Two options is the floor; the one the paper argues against belongs among them when it is a live choice. The recommendation names its letter and says why in one or two sentences with the evidence.
Write them under the question's bullet, after its existing prose, before the next top-level bullet — exactly this shape, since packages/cli/core/decisions.js parses it:
- **The question's bold lead, unchanged?** Its existing prose, unchanged.
- **A** — what the tree looks like if A, grounded (`path/file.js:120`)
- **B** — what the tree looks like if B, grounded
- **Recommend A** — why, citing the evidence
A wrapped line is indented four spaces, under the option it continues. The bold lead line stays byte for byte as it is: the question's id is a slug of it, and an edited lead is a new question the loop cannot find.
Settled — an existing ruling already answers it
When a live FJS-D ruling (or shipped code a ruling stands behind) answers the question, frame one option that states that answer, and a recommendation naming the ruling id:
- **A** — as `FJS-D123` rules: what that means for this question
- **Recommend A** — `FJS-D123` settles it: the sentence of the ruling that does
The owner confirms it with fli decide <id> --by FJS-D123, which cites that ruling and issues no new one. A superseded or withdrawn ruling settles nothing — cite what replaced it, or frame the question.
Unclear — the question cannot be framed as written
Two questions in one bullet, a premise the tree contradicts, or no choice left to make. Write nothing under it, and say which of those it is in the report line.
4. Verify
fli decisions --json after the edits, and read each id's state and recommend. A framed question is decidable with its recommendation read back; a settled one is open with the recommendation present. An id that is no longer there means its lead line changed — restore it. Then fli done and clear what it lists.
5. Report
One line per question on what was framed and why the recommendation. The last lines of the report are exactly frame-next: <id> <status>, one per question, status one of framed, settled, unclear (followed by — and the reason), busy or failed. The loop reads the register first and these lines only for what the register cannot show.
Then end the session. One paper per session: its questions share an argument and are framed together so their options agree; a second paper would carry every file the first one read. A batch is bun run frame:loop.