| name | rdr-propose |
| metadata | {"argument-hint":"<NNNN> [--commit | --no-commit]"} |
| description | Use to move a seeded RDR from problem statement to chosen approach with alternatives weighed. Runs Stage 2 and re-validates stale seeds. Trigger for propose approach, $rdr-propose, or /rdr-propose. |
rdr-propose — Stage 2 (Propose)
Move from a problem statement to a chosen approach with the seriously-considered
alternatives recorded and rejected, so the draft has one coherent solution for
Refine to tighten. Doubles as the front-half resume point: its freshness check
re-validates a seed that sat idle and went stale.
Usage
Codex: $rdr-propose <NNNN>
Claude: /rdr-propose <NNNN>
- Read
rdr-common.md; run §seam-bind + §rdr-resolve
to bind $RDR_RESOURCES, RDR_PATH.
- Run the prompt
02-propose.prompt.md.
Propose owns selection (which approach wins, read from prior art), not
verification (deep spikes / full corpora) — that stays at Resolve, by
design. The split is depth, not avoidance. It runs the freshness check (step 0),
then reads prior art, enumerates 2–4 grounded approaches, recommends one,
premortems it, and writes Proposed Solution / Alternatives / Decision Rationale +
a Pending Critical Assumptions list into the draft. Retrieval-first: prior
art is read before approaches are named (an LLM that enumerates first anchors
on its training prior), with a cite-check — every prior-art claim the choice
rests on is quoted/anchored from source, not paraphrased (the guard against the
inverted-citation defect that overturned an approach three stages late). No prior
art found → an explicit ⚠ no prior-art coverage line. large/foundational
also build a scored Questions-Options-Criteria matrix so the choice falls out
of an explicit comparison, not prose.
Review gate (Stage 02-propose.md)
- The seed was still current (step 0 ran; stale refs/scope fixed; any prior "retained
as history" note acted on and deleted).
- Real alternatives weighed (not one option or strawmen).
- Prior art read before the approaches (grounds the set;
⚠ no prior-art coverage
if none), and the choice's prior-art claims are quoted/anchored, not paraphrased.
large/foundational: the choice falls out of a scored Q-O-C matrix, not prose.
- The chosen approach solves the user's problem; the premortem ran and the choice
survived it.
- Load-bearing assumptions named (Pending is fine) — Stage 4's worklist.
- Technical Design proportionate — enough to commit, not a full implementation.
Next step (rdr-common §next-step)
- If autocommit is on, run §commit for
propose first (see the rdr-common table).
Next: /rdr-refine NNNN — remove contradiction, redundancy, bloat.
- A surfaced contradiction is a real design hole → iterate this stage (re-run
/rdr-propose NNNN) before refining.
/rdr-status NNNN to re-orient.