# Feedback > Use when receiving feedback on your work — review findings, user corrections, audit results, external reviewer notes — to triage by authority, adjudicate each point on evidence, integrate without regressions, and close the loop with the source. Evidence flips positions; assertions don't. - Skill: `davidlee/feedback` (Agent Skill) - Install (CLI): `npx skillmds@latest add davidlee/feedback` - Raw SKILL.md: https://api.skillmd.com/api/skills/davidlee/feedback/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Security - Author: davidlee (https://skillmd.com/u/davidlee) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/davidlee/feedback --- # Feedback An interrupt procedure for the receiving side of feedback. It begins when feedback arrives and ends when the loop is closed: every point integrated, contested, or explicitly parked — never silently dropped. **What counts as feedback:** asserted claims from a source with intent — a human correction, a reviewer's findings (human or agent), an audit result. **What doesn't:** mechanical signals. A failing test, CI red, or lint output is an *observation*, not an assertion — diagnose it (see the `rigour` skill's differential diagnosis), don't "respond" to it. Two failure poles, both defects: - **Capitulation** — flipping position because someone asserted otherwise. "You're right" followed by unexamined rework is not integration; it swaps your errors for theirs and teaches the source nothing. - **Entrenchment** — defending prior conclusions reflexively, demanding the source prove what you could check yourself, litigating every correction. The rule that resolves both: **positions move on evidence, not on assertion — in either direction.** You may not flip on an unexamined claim, and you may not hold on unverified recollection. ## Authority - **Instructions bind.** What to *do* is the source's call (subject to your ordinary duty to flag consequences). Adjudication below applies to claims about what is *true*. - **User-supplied material and empirical knowledge outrank your recall.** Pasted documents, reported observations, domain knowledge: ground truth relative to anything you merely remember. Never counter them from training recall or fresh secondhand research, and never demand provenance. The one admissible counter is a **directly verified observation** — behaviour you observed, output you produced this session — and it is delivered as a discrepancy report ("the doc says X; the endpoint returns Y"), because the user would rather know than be deferred to. Report, then follow their call. - **Reviewer findings are advisory but undroppable.** Whatever their merit, each one receives an explicit disposition. ## Procedure 1. **Triage** (batches; a single correction skips to 2). Order first by dependency, then by severity × authority. Group findings that mutate the same declaration. Rule first on anything whose resolution changes another finding's fix site or economics; integrating a flat list independently is a reliable way to rewrite the same block several times. State the order if it isn't obvious. 2. **Adjudicate each point on the artifact, not on trust.** Re-read what the finding actually cites; reviewers confabulate lines, behaviour, and whole defects. Separate its **observation** (what is wrong) from its **prescription** (what to do): either can be right while the other is wrong. Verify the observation against the incumbent and design the repair on its own merits. Classify: - *correct* → integrate; - *wrong* → contest, with the verified evidence that shows it; - *ambiguous* → clarify before acting; - *preference* → accept, or negotiate once if it costs something real. Before flipping any position: name **which specific claim of yours the feedback defeats**. If it defeats none and you hold verified evidence, contest. Scale persistence by source: against an agent reviewer, contest as long as the evidence supports it; against the user, one crisp counter carrying your evidence, then defer — noting material dissent rather than burying it. 3. **Generalize.** Name the defect class in one line, then sweep the artifact in scope for siblings and fix them with the instance; class members beyond the scope are captured as follow-up work, not silently expanded into. When the repair changes or withdraws a governing claim, the scope is repo-wide: find the normative owner, inspect both authored tiers through the entity's `show` view, and sweep dependent projections after the repair. A negative search needs a positive control on the same surface. 4. **Integrate without regression.** Check the rework against previously settled decisions. Feedback that conflicts with a settled decision reopens that decision *explicitly* — surface it; don't quietly re-litigate it in the course of a fix. A finding supplies a direction, not a destination: do not turn an overclaim into its unsupported inverse, add machinery the evidence did not justify, or mistake disclosure of a defect for its repair. 5. **Re-verify against the complaint.** The check is "does this address what the source actually raised", not "did something change". Where the finding implies a failing case, make it fail before the fix and pass after. Apply a new or corrected rule to its whole population, re-run its evidence (or mark that evidence stale), and re-read the seams and neighbouring claims touched by the correction. 6. **Close the loop.** Report dispositions point-by-point back to the source: integrated (and how), contested (and on what evidence), clarification sought, parked (and why). Silence on a point reads as acceptance and hides the drop. 7. **Harvest.** Artifact corrections die with the task; corrections to your *process or the source's preferences* are the durable leg — harvest them per `harvest.md`, always with the *why*, or the same correction recurs next session. ## With structured review ledgers If the feedback arrives on a structured ledger (a doctrine RV, a PR review thread), the ledger owns the mechanics — disposition vocabulary, contest moves, resolution states, close-gates. This skill is the conduct *behind* those verbs: adjudicate before disposing, evidence before contesting, no disposition chosen to dodge a gate. Don't duplicate the ledger's records in prose, and don't bypass its gates. Compatible, not prescriptive. For design review specifically, the RV is the revision-history surface. Keep finding chronology and point-by-point responses there, keep `design.md` as the current legible design, and promote accepted durable rulings to the knowledge or governance record that owns them. Before handing a revision back, perform a read-only self-attack on the corrected passages for factual error, internal contradiction, class siblings, stale projections, and regression of earlier rulings. Re-check every issue it suggests before acting; this pass discovers possible defects and does not certify its own repairs. ## Guardrails - Never silently drop a point — integrated, contested, clarified, or parked. - Never flip on assertion alone; never hold on recollection alone. - Do not counter user-supplied material with what you remember or with secondhand research; verified observation is the only counter, and it is a report, not a refusal. - Capitulation theatre — agreeing effusively, then changing nothing or changing everything without adjudication — is a worse defect than honest contest. - Do not relitigate settled decisions under cover of integrating feedback. - Contesting a user's finding twice is entrenchment; contesting an agent reviewer's confabulation firmly is a service. ## Outcomes - Every point has a visible disposition; nothing evaporates. - Position changes trace to evidence; the artifact absorbs the feedback's substance, not its noise. - Classes get fixed, not instances; settled decisions survive rework or are reopened deliberately. - The source learns what landed, what didn't, and why. - Durable corrections stop recurring.