false-competence-check
Check finished-looking agent work against the failure modes that look like competence but are not. Use before declaring a fix done, before opening or approving a PR, when a change was fast and confident, when tests are green on the first try, after a rename, sweep, or cleanup, or when reviewing another agent's "done" claim. Do not use as a substitute for the task's own acceptance criteria.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- false-competence-check
- description
- Check finished-looking agent work against the failure modes that look like competence but are not. Use before declaring a fix done, before opening or approving a PR, when a change was fast and confident, when tests are green on the first try, after a rename, sweep, or cleanup, or when reviewing another agent's "done" claim. Do not use as a substitute for the task's own acceptance criteria.
02 资源文件 1
- 参考文档agents/openai.yaml118 B
03 SKILL.md 正文
False Competence Check
Rule
The most expensive mistakes look like good work.
Decisiveness, speed, cleanup, diligence, and confidence each have a counterfeit. The counterfeit passes casual review because it has the shape of the real thing.
The counterfeits
| Looks like | Actually | Counter-procedure |
|---|---|---|
| Decisiveness | Implementing from stale premises | A prior audit is a hypothesis. Re-locate every symbol by content before building on it. |
| Speed | A shallow search-based fix | A rename searches the entire file, including templates. Claims about large files need uncapped searches and reading every hit. |
| A quick win | A visual fix that dodges the data | Identify the mechanism (paint order, clipping, scale, data truth) before choosing the fix class. A wrong number is withheld or labelled, never dressed. |
| Craftsmanship | A broad refactor disguised as cleanup | Scope out adjacent ugliness. Use anchored edits and verify with counts and a parse. A partial sweep is worse than none. |
| Diligence | Tests that bless the bug | Break the fix and watch the test fail. When a test encodes current behaviour, ask whether that behaviour is the bug. |
| Generosity | A PR with hidden unrelated diffs | Diff from the merge base. Every hunk traces to scope-in; incidental fixes get their own commit and a line in the report. |
| Confidence | False verification | Verify in the environment class where the code runs. Compare suite results against a clean tree. Count deliverables explicitly; search your own diff for placeholders. |
| Completeness | One instance fixed, pattern not hunted | Search every surface that reads the same data or renders the same pattern before declaring done. |
| Progress | Fixing the wrong link in the chain | Trace the whole pipeline first. When a feature yields nothing, check the wiring before the producer. |
| Finished | A code fix presented as data repair | Corrected code does not repair rows it already wrote. Data repair is a separate, dry-run-first, owner-gated path. |
| Defensiveness | Safety relief keyed to a label | Key relief to the degraded condition itself and test the success-with-degraded-inputs path. |
| Validation | Model-generated values passing value-blind guards | Overwrite system-owned fields at the trusted boundary. Validate claimed values against the canonical source. Print "unavailable", never a synthetic zero. |
| Consistency | Grandfathered violations reused as precedent | Ratchets only go down. New code meets the current standard even inside old files. |
Workflow
- Name which counterfeit your current work most resembles.
- Run the one additional check that would falsify the resemblance.
- Report the check and its result, not the reassurance.
Falsify it
Ask: "Which of these does my diff most resemble, and what single check would prove it is not that?" If you cannot name the check, you have not ruled it out.
Stop conditions
- The falsifying check requires access or authority the task does not have.
- The check fails and fixing it would broaden scope beyond the charter.
Report the resemblance and the blocked check rather than claiming done.
Output
CLOSEST COUNTERFEIT · FALSIFYING CHECK · RESULT · RESIDUAL RISK.
04 同仓库更多 Skills
verify-before-cite
Live-verify work state before reporting or relying on it, and reconcile every place that state is recorded. Use before a status report, standup, handoff, or "what's open" answer; when a tracker, queue, memory note, or prior session says something is pending, merged, shipped, or blocked; after any merge, release, or promotion; or when records in two places disagree. Do not use to judge whether a claim about data or numbers is proven.
truth-boundary
Enforce provenance and uncertainty boundaries. Use when a system mixes observed, authoritative, user-attested, derived, estimated, inferred, reconstructed, or unknown values; when missing data could become zero; or when downstream output may overstate certainty. Also use when told to fill a missing, failed, or empty value with zero, a default, or a guess so the work can keep going.
systemic-bug-instinct
Use first, before reading, searching or editing code, whenever a defect, wrong value or balance, stale state, edge-case failure, broken workflow, or misleading UI is reported and a fix is requested, even when the request names the one file to patch. Diagnoses the bug as a possibly systemic failure to find the root mechanism, failure class, blast radius across sibling consumers, fix location, and regression guard. Do not use for a one-line typo or syntax fix, for two components that disagree about one concept (use seam-audit), for a dense screen whose value may be a presentation problem (use operator-ui-audit), or for financial records that contradict each other (use accounting-integrity).
swarm
Design and run the smallest effective multi-agent harness for complex work. Use when the user explicitly says SWARM or asks for a swarm, multi-agent workflow, adaptive orchestration, harness design, parallel investigation, independent falsification, architecture reconciliation, broad audit, competing root-cause hypotheses, or high-risk work that benefits from independent reasoning lanes. Do not use for trivial edits or linear tasks one agent can complete cheaply; the best swarm may use zero subagents. SWARM grants orchestration, not broader implementation, merge, deploy, credential, financial, or destructive authority.
shipping-lanes
Design, run, or adopt a pipeline in which agent-written changes are reviewed and merged without a person in the routine loop, with review and landing matched to each change's risk. Use when setting up or changing how PRs get reviewed and merged by agents; when every change waits on one person, or none gets reviewed; when a reviewer is only reached if the author asks for it; when docs and permission changes take the same path; or when bringing a new repository onto the shared shipping system. Do not use to decide whether one specific merge is authorised (use merge-authority), to judge a finished-looking result (use false-competence-check), or to prove a single guard is reached (use reachability-proof).
seam-audit
Audit the boundary where two components represent the same concept, because each side can pass its own tests while the combined system is wrong. Use when a value looks coherent but is wrong, when a projection or compatibility view stands in for a full record, when sync or merge involves deletes, when status codes or result shapes cross a boundary, or when "unknown" may become zero, false, or success as it crosses between two components. Do not use for a defect inside one component (use systemic-bug-instinct), for grading one claim's certainty (use truth-boundary), or for a financial repair's preview and commit (use ledger-repair).