Readback
Memory-safe engineering-peer plugin for Claude Code + Codex: 11 skills, 6 bounded agents, 4 hooks. Claims only what it can prove; closes every turn with one next action.
npx -y skills add menot-you/n-soul --skill readbackAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 15 days oldThe repository was created 15 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Lock shared understanding before substantive work — translate the ask into executable form and read it back as numbered, falsifiable statements the owner confirms by number.
SKILL.md
2.4 KB, 493 tokens by cl100k_base, as published. Nobody here has run it
Readback
Use before substantive work whenever ambiguity affects the result, authority, safety, scope, or a handoff — and at every delegation seam: owner→session, session→subagent, PRD→issue. This skill executes the shared-understanding gate: the numbered list IS the claim of understanding, and per-number confirmation is the gate. Never claim "yes, I understood."
Fast path: for low-risk reversible work with an obvious next action, keep the gate internal — state the working assumption, act inside granted authority, and bind the close to observed verification. Do not turn the gate into ceremony.
Method
- Read the current artifacts and repository state first; anchor every statement in what is actually there, never in memory of it.
- Translate the ask into the executable form — Task (one goal) · Input · Constraints (including what NOT to do) · Output · Verify (observable success criteria). One prompt = one goal: split a compound ask into chained single-goal units.
- A success criterion that cannot be verified is a numbered gap to surface, never a license to improvise.
- Write the readback as numbered, atomic, falsifiable statements. Label inferences and separate them from what the owner said. Give each load-bearing term one verified meaning; an ambiguous term gets its own numbered statement.
- Deliver with default-confirm: "what you do not mark, I take as said" — the confirmer corrects or confirms by number.
- Block: the gated step stays
blockeduntil the readback returns confirmed; the next step does not fire without it. A readback that does not block is decoration. - Apply corrections by number and re-emit only the changed statements, until confirmed.
- Recurse across the seams: a dispatched subagent reads back acceptance criteria and bounded scope before working; an issue a cold agent cannot execute without asking is not ready.
Output
The confirmed readback record — the executable form plus the confirmed numbered statements and any surfaced gaps — attached to the work it unblocks; it becomes part of that work's contract.
The gate is closed when the confirmer has disposed of every number and no load-bearing term is left unverified.
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.