AGENT SKILL

calibration-loop

Converge on an ambiguous preference through concrete either/or instances instead of re-asking the same abstract question. Use when brainstorm divergence is unsure, a doc-review disposition is disputed, or feedback scope classification is ambiguous. Does not resolve unrelated ambiguity or bypass human confirmation on the converged rule.

License 许可
MIT
最近更新
2026-11-03
Stars
★ 2
Forks
0

01 前言 Front Matter

name
calibration-loop
description
Converge on an ambiguous preference through concrete either/or instances instead of re-asking the same abstract question. Use when brainstorm divergence is unsure, a doc-review disposition is disputed, or feedback scope classification is ambiguous. Does not resolve unrelated ambiguity or bypass human confirmation on the converged rule.

03 SKILL.md 正文

Calibration loop

A reusable convergence primitive for genuine either/or ambiguity — not a general Q&A loop. Presents one concrete instance per turn, restates the inferred principle, and stops on stability.

Model tier: cheap — resolve via python3 scripts/sw_bootstrap.py resolve-model-tier.py -- --skill calibration-loop.

When to use

Only when a prior attempt to resolve a preference/disposition abstractly has already failed (the operator said "unsure", two reviewers disagree, or scope classification is genuinely ambiguous) — never as a first resort over a direct question.

Fixed verdict set

Every turn presents one concrete either/or instance (a specific example, not the abstract question again) and accepts exactly one of a fixed vocabulary:

Verdict Meaning
A This instance resolves toward option A
B This instance resolves toward option B
either Genuinely indifferent on this instance (informative — narrows the principle to "doesn't matter here")
neither Instance is malformed or off-tension — discard, generate a better one
more-info Operator needs one clarifying fact before verdicting this instance — answer, then re-present the same instance

Never accept free text as the primary channel — free text is allowed only as an annotation alongside one of the five verdicts above, never as a substitute for one.

Procedure

  1. Frame the tension — restate the two poles (A vs B) in one sentence from the blocked artifact (divergence stance pair, disputed disposition, ambiguous-scope routing).
  2. Present one concrete instance — a specific, realistic case that would force a choice between A and B. Never present the abstract question a second time.
  3. Accept a verdict from the fixed set above.
  4. Restate the inferred principle — after each verdict, state the working rule in one sentence ("so: when X, prefer A; when Y, prefer B") and ask for a quick confirm/correct on the restatement, not a new open question.
  5. Adapt toward unexplored seams — the next instance MUST probe a boundary the current principle has not yet been tested against (not a near-duplicate of the last instance). Track covered vs. unexplored seams explicitly in scratch notes for the run.
  6. Stop on stability — converge when either:
    • two consecutive instances confirm the same restated principle with no correction, or
    • a hard cap of 5 instances is reached (whichever first). On cap-out without stability, record converged: false and escalate to explicit human decision — never silently pick a side.
  7. Write the converged rule into the blocked artifact — the calling consumer (brainstorm, doc-review, feedback) persists the final principle at its own persistence point (see Consumers below); this skill never writes on its own — it returns the converged principle + verdict trail to the caller.

Consumers (wired now)

Consumer Trigger Where the converged rule lands
/sw-brainstorm divergence (unsure routing) Operator "unsure" between ~2 viable stances Requirements doc Key Decisions — chosen stance + converged principle
/sw-doc-review disposition disputes Synthesizer/reviewer disagreement on safe_auto vs gated_auto/manual for a finding class Review-round synthesis notes; principle informs the disposition for that finding and future same-class findings in the run
/sw-feedback ambiguous-scope classification Signal cannot be cleanly classed gap-capture vs brainstorm (or substantial vs trivial) Route record (skills/feedback/references/route-record.md) — converged principle recorded alongside the route decision

Guardrails

  • Never a substitute for a direct, answerable question — only for genuine either/or ambiguity after a first abstract attempt.
  • Never accept the same instance twice in a row as "resolved" — each turn must probe a new seam.
  • Cap-out (converged: false) always escalates to explicit human decision; never auto-resolves silently.
  • This skill informs the blocked artifact — it does not itself freeze, merge, or dispatch anything.

04 同仓库更多 Skills