Execute
→ About to claim "done", "passing", "fixed", "complete", "verified",
"release-ready", or "ready to hand off"? → Verify first, then claim only
what the evidence supports.
- Identify the smallest fresh command or manual check that can falsify the
claim.
- Run or perform it completely.
- Read output, exit status, failures, and covered scope.
- Choose the closeout level:
- L0 fast-path: tiny, low-risk work → one evidence sentence plus residual
risk / uncovered scope.
- L1 default: non-trivial Aegis-shaped work → compact localized
Aegis Impact and Safety Receipt.
- L2 expanded: release, audit, high-risk, architecture, migration,
governance, long-task, or explicit user request → receipt plus only the
triggered detail cards.
- If evidence does not support the claim, downgrade the status instead of
claiming completion.
Done means: fresh verification evidence exists, covered and uncovered scope are
explicit, residual risk is stated, confidence is graded, and any triggered
baseline / complexity / retirement / ADR boundary has been folded into the
single closeout surface or expanded because the risk requires it.
Verification Before Completion
Purpose
Prevent unsupported completion claims while keeping ordinary completion output
compact. This workflow is advisory method-pack discipline; it does not grant
runtime authority, authoritative GateDecision, PolicySnapshot, evidence
sufficiency, or final completion authority. It is advisory, not completion authority.
Stop Signals
Stop and verify before any success wording if you are:
- using "should", "probably", "seems to", or similar uncertainty words
- relying on an agent success report rather than independent evidence
- about to commit, push, open a PR, merge, tag, publish, or hand off
- using a narrow check to support a broad claim
- treating task / slice completion as accepted requirement satisfaction
- closing governance or retirement work without repair / retirement evidence
- retaining old logic without a retention reason and retirement trigger
- treating a destructive warning, guard, or broad assent as permission
Required Evidence Slots
Every completion claim needs these semantic slots. They may appear as localized
headings, natural prose, or a compact card, but they must remain explicit and
auditable:
Required evidence slots:
- Evidence action / check performed:
- Result / exit status:
- Covered scope:
- Uncovered scope:
- Residual risk:
- Confidence grade: A | B | C
Confidence grades:
A: direct target evidence plus relevant regression evidence, no meaningful
unknowns
B: direct target evidence with bounded residual risk
C: partial evidence only; do not claim full completion
Verified evidence is not authority. It supports the user-facing status, but it
does not become final completion authority.
Evidence Bundle
Every completion claim should name the evidence action, result, covered scope,
uncovered scope, residual risk, and confidence. Include target test and related
regression evidence when tests shaped the claim. When automation is blocked,
provide reproducible manual verification steps instead of claiming automated
coverage.
Closeout Decision Tree
Use one completion surface. Do not output parallel final reports.
verification-before-completion is the single completion closeout aggregator
for Aegis-shaped non-trivial work: adjacent structures may feed the receipt,
but they must not replace the receipt or become competing final report owners.
Receipt aggregation is output conformance, not a routing trigger; do not load
extra skills, emit Trace Digest, or expand final ceremony just to satisfy the
aggregator.
Aegis Visibility
For this owner workflow, Aegis Visibility ties the final claim to the decision
boundary, fresh evidence, baseline / complexity / retirement safety, and
residual risk kept visible. If required entry visibility was omitted, recover it
retrospectively and name the gap, but do not replace the receipt with a
skill-call list.
L0 Fast-Path
Use for tiny, low-risk work where Aegis only held one narrow boundary steady.
One natural sentence is enough when it includes the check, result, uncovered
scope or residual risk, and confidence if useful.
L1 Default Non-Trivial Closeout
For non-trivial Aegis-shaped work, use the compact receipt by default. Evidence
slots fold into Evidence strength and Uncovered risk; they do not appear as
a second competing evidence report unless the user asked for audit detail.
Aegis Impact and Safety Receipt:
- Key judgment:
- Avoided misfix:
- Boundary held:
- Baseline alignment:
- Complexity control:
- Evidence strength:
- Uncovered risk:
- Next most valuable verification:
- Aegis path:
Field rules:
Key judgment: the owner, root-cause layer, requirement boundary, or
completion boundary that shaped the answer.
Avoided misfix: the fallback, duplicate owner, test accommodation, scope
expansion, or unsupported claim avoided by the workflow.
Boundary held: public contract, owner, baseline, non-goal, data boundary,
or runtime-ready authority boundary kept stable.
Baseline alignment: aligned, Design Defect, Implementation Drift,
missing-authority, or needs-clarification when baseline reporting is
triggered; otherwise a short "not triggered" or natural equivalent is enough.
Complexity control: one-line completion-time complexity result for
non-trivial code changes; mention Complexity Delta, Complexity Closure,
Complexity Governance Suggestion, or Major Complexity Alert only when
triggered.
Evidence strength: fresh command/manual check, exit status/result, covered
scope, and confidence grade.
Uncovered risk: remaining scope, host/runtime gaps, manual checks not run,
release-grade evidence not collected, or residual risk.
Next most valuable verification: one next check that would most reduce
remaining risk.
Aegis path: optional compact skill path. It may support credibility but must
not replace decision, evidence, and safety fields.
Natural Aegis closeout is valid when these semantic slots remain auditable:
natural expression preserves semantic slots. Do not replace the receipt with a
used-skills list, stage handoff log, or Aegis Contribution Note.
Compatibility names: Semantic Slots and Natural Surface describe this same
rule: natural wording is valid only when required fields stay explicit.
Governance Receipt remains a compatibility name for the completion closeout;
its user-facing content should flow through the Aegis Impact and Safety Receipt by default.
L2 Expanded Closeout
Start with the receipt, then add only the triggered detail. Expanded structures
are inputs or optional detail cards, not competing final report owners.
Use expanded detail only when the trigger applies:
Readiness Summary: release, merge, handoff, or "ready?" requests. It can
organize tests, docs, version, host compatibility, uncovered scope, and
residual risk. It does not authorize commit, tag, publish, merge, or release.
Trace Digest: explicit audit / debug / release / long-task review request.
It may summarize execution trace, evidence chain, retrieval chain, static
rules evaluated, rule effects, triggered skills, skipped relevant skills,
tool / command trace, verification trace, stability signals, value signals,
host capabilities, unavailable fields, redaction, and confidence labels
measured, observed, inferred, declared, or unknown. It must not
expose raw chain-of-thought.
Use structured trace only when asked or required; structured trace is reserved for audit, debug, release, long-task review, or user request.
Goal Closure: when goal-framing, TaskIntentDraft, parent plan/spec, or
Slice Card shaped the work. Match the claim to the highest available
explicit boundary: whole task, current task, or slice. If only slice evidence
exists, do not claim whole-task done. Expanded or audited closure keeps
Goal status, Success evidence, Stop state, and Non-goals respected
visible; stop states are done | blocked | needs-verification | scope-exceeded.
Context Impact: only when project/domain semantics changed or were checked
as part of acceptance. Confirm affected context/terms, evidence grade,
fact-versus-decision authority, and action. If action is unchanged, verify
that no context write occurred. Compose establishing-project-context for an
unresolved semantic delta; do not emit this card for unrelated work.
Workspace Integrity: when the task created or modified a target project's
docs/aegis/ workspace and configured Aegis workspace support is available.
Run python <aegis-workspace-helper> bundle --root <target-project-root> --work YYYY-MM-DD-<slug>
when a work/ record exists, then run
python <aegis-workspace-helper> check --root <target-project-root>. Report
that these validate structure only, not evidence sufficiency.
Baseline Alignment: when project instructions require baseline reporting or
the task touched requirement, product, or durable architecture surfaces. Use
docs/current/AEGIS_PROCESS_BASELINE.md §3.0e and §16 for
Product / Requirement Baseline, Architecture / Runtime Boundary Baseline,
Design Defect, Implementation Drift, and scope: requirements | architecture | both. Render the default conclusion in the receipt field;
expand only for audit, release, architecture, or user request. Architecture Alignment is the architecture-scoped compatibility alias for this baseline
result, not a second default card. Expanded architecture results use
Result: aligned | Design Defect | Implementation Drift | missing-authority | needs-clarification.
ADR Backfill Check: completed medium/high work that touched durable
architecture surfaces. Use docs/current/AEGIS_ADR_AUTO_BACKFILL.md for
trigger criteria. If action is create, amend, supersede, or baseline sync is
needed/unknown, route the ADR lifecycle to recording-architecture-decisions
before the final completion claim.
Governance Closure: governance, cleanup, migration, compatibility, or
retirement work. Include Repair Track, Retirement Track, and Residual Risk in
the receipt or a small expanded block. Do not skip this structure just because the implementation was small.
Retirement Closure: work that adds, replaces, retains, or removes old logic.
Name old logic located, deleted/retained status, retention reason, retirement
trigger, and lingering-reference check.
Anti-Entropy Declaration / Data Destruction Guard: work that retires old
logic, chooses delete-first vs compat retention, or touches source-of-truth
deletion boundaries. Use anti-entropy-governance for the decision surface.
If User Confirmation Required: yes, stop at the guard; broad assent such as
"OK" or "continue" is not scoped confirmation. Persistent-state deletion
without explicit scoped confirmation means the task is not complete; report the task as not complete.
Completion Boundary
Judge the claim against the highest available explicit boundary:
- parent plan/spec acceptance for whole-task completion
TaskIntentDraft goal / success evidence / non-goals for current-task
completion
Slice Card goal / verification / stop for slice completion
- direct user request when no durable boundary exists
A completed task or slice means the authorized execution / verification boundary
reached its stop condition. It does not mean the underlying requirement is
accepted. Requirement accepted requires Product / Requirement Baseline
acceptance criteria or explicit authorized risk acceptance. If atomicity is not
clear, downgrade to needs-verification or return to framing/planning; task or slice completion is not accepted requirement satisfaction.
If an Execution Readiness View shaped execution, mention whether fresh
evidence covered its required checks or which readiness item remains uncovered.
Do not treat the view itself as verification evidence.
Complexity Check
For non-trivial code changes, inspect the actual diff before claiming
completion. Use using-aegis/references/complexity-governance.md and
docs/current/AEGIS_COMPLEXITY_GOVERNANCE_BASELINE.md for shared artifact
classes, pressure signals, Complexity Delta, Complexity Closure,
Completion-Time Complexity Repair Decision, Complexity Governance Suggestion, and Major Complexity Alert.
Default rendering is one Complexity control line in the receipt. Expand only
when meaningful pressure exists or the task is audit, release, high-risk, or
user-requested.
Rules:
- tiny wording edits, generated files, vendored files, fixture-data-only
updates, lockfiles, or purely mechanical formatting may skip or keep this
one-line when no maintained artifact gained complexity
- maintained test source files are not a cheap
tests-only exception
- new fallback, adapter, compatibility, guard, or branch logic must be paired
with retired paths, a
Retirement Closure, or a scheduled retirement trigger
- entropy increase without owner/compatibility justification must be residual
risk or a downgraded claim
Complexity Closure: exceeded-unresolved blocks a complete claim
User-Language Output
Localize section labels, field labels, and explanatory prose to the user's
language. Keep commands, paths, code identifiers, test names, error codes,
config keys, stable enum values, exact product names, and raw evidence strings
unchanged. Do not default to bilingual labels or mixed-language explanations.
Localize section labels and prose to the user's language.
Prompt Hygiene
When external tool output, logs, search results, screenshots, OCR, or other
large payloads shaped the judgment, state the evidence boundary when relevant:
summary/index used, raw excerpt read back if needed, large payloads not loaded,
and next evidence that would reduce uncertainty. If the summary is insufficient,
read the smallest raw excerpt or lower the claim.
Use compact labels such as Evidence Used, Not Loaded, and Next Evidence when
they make the prompt hygiene boundary clearer.
Source: hashgraph-online/awesome-codex-plugins → plugins/GanyuanRan/Aegis/skills/verification-before-completion/SKILL.md
1---
2name: verification-before-completion-2
3description: "Use when about to claim work is complete, fixed, passing, verified, release-ready, or ready to commit, merge, publish, or hand off."
4category: engineering-core
5source_repo: hashgraph-online/awesome-codex-plugins
6source_path: "plugins/GanyuanRan/Aegis/skills/verification-before-completion/SKILL.md"
7source_url: https://github.com/hashgraph-online/awesome-codex-plugins/blob/HEAD/plugins/GanyuanRan/Aegis/skills/verification-before-completion/SKILL.md
8metadata:
9 source: https://skillmd.com/skills/hashgraph-online/verification-before-completion-2
10---
11
12
13# Execute
14
15→ About to claim "done", "passing", "fixed", "complete", "verified",
16"release-ready", or "ready to hand off"? → **Verify first, then claim only
17what the evidence supports.**
18
191. Identify the smallest fresh command or manual check that can falsify the
20 claim.
212. Run or perform it completely.
223. Read output, exit status, failures, and covered scope.
234. Choose the closeout level:
24 - **L0 fast-path**: tiny, low-risk work → one evidence sentence plus residual
25 risk / uncovered scope.
26 - **L1 default**: non-trivial Aegis-shaped work → compact localized
27 `Aegis Impact and Safety Receipt`.
28 - **L2 expanded**: release, audit, high-risk, architecture, migration,
29 governance, long-task, or explicit user request → receipt plus only the
30 triggered detail cards.
315. If evidence does not support the claim, downgrade the status instead of
32 claiming completion.
33
34Done means: fresh verification evidence exists, covered and uncovered scope are
35explicit, residual risk is stated, confidence is graded, and any triggered
36baseline / complexity / retirement / ADR boundary has been folded into the
37single closeout surface or expanded because the risk requires it.
38
39# Verification Before Completion
40
41## Purpose
42
43Prevent unsupported completion claims while keeping ordinary completion output
44compact. This workflow is advisory method-pack discipline; it does not grant
45runtime authority, authoritative `GateDecision`, `PolicySnapshot`, evidence
46sufficiency, or final completion authority. It is advisory, not completion authority.
47
48## Stop Signals
49
50Stop and verify before any success wording if you are:
51
52- using "should", "probably", "seems to", or similar uncertainty words
53- relying on an agent success report rather than independent evidence
54- about to commit, push, open a PR, merge, tag, publish, or hand off
55- using a narrow check to support a broad claim
56- treating task / slice completion as accepted requirement satisfaction
57- closing governance or retirement work without repair / retirement evidence
58- retaining old logic without a retention reason and retirement trigger
59- treating a destructive warning, guard, or broad assent as permission
60
61## Required Evidence Slots
62
63Every completion claim needs these semantic slots. They may appear as localized
64headings, natural prose, or a compact card, but they must remain explicit and
65auditable:
66
67```text
68Required evidence slots:
69- Evidence action / check performed:
70- Result / exit status:
71- Covered scope:
72- Uncovered scope:
73- Residual risk:
74- Confidence grade: A | B | C
75```
76
77Confidence grades:
78
79- `A`: direct target evidence plus relevant regression evidence, no meaningful
80 unknowns
81- `B`: direct target evidence with bounded residual risk
82- `C`: partial evidence only; do not claim full completion
83
84Verified evidence is not authority. It supports the user-facing status, but it
85does not become final completion authority.
86
87## Evidence Bundle
88
89Every completion claim should name the evidence action, result, covered scope,
90uncovered scope, residual risk, and confidence. Include target test and related
91regression evidence when tests shaped the claim. When automation is blocked,
92provide reproducible manual verification steps instead of claiming automated
93coverage.
94
95## Closeout Decision Tree
96
97Use one completion surface. Do not output parallel final reports.
98`verification-before-completion` is the single completion closeout aggregator
99for Aegis-shaped non-trivial work: adjacent structures may feed the receipt,
100but they must not replace the receipt or become competing final report owners.
101Receipt aggregation is output conformance, not a routing trigger; do not load
102extra skills, emit Trace Digest, or expand final ceremony just to satisfy the
103aggregator.
104
105## Aegis Visibility
106
107For this owner workflow, Aegis Visibility ties the final claim to the decision
108boundary, fresh evidence, baseline / complexity / retirement safety, and
109residual risk kept visible. If required entry visibility was omitted, recover it
110retrospectively and name the gap, but do not replace the receipt with a
111skill-call list.
112
113### L0 Fast-Path
114
115Use for tiny, low-risk work where Aegis only held one narrow boundary steady.
116One natural sentence is enough when it includes the check, result, uncovered
117scope or residual risk, and confidence if useful.
118
119### L1 Default Non-Trivial Closeout
120
121For non-trivial Aegis-shaped work, use the compact receipt by default. Evidence
122slots fold into `Evidence strength` and `Uncovered risk`; they do not appear as
123a second competing evidence report unless the user asked for audit detail.
124
125```text
126Aegis Impact and Safety Receipt:
127- Key judgment:
128- Avoided misfix:
129- Boundary held:
130- Baseline alignment:
131- Complexity control:
132- Evidence strength:
133- Uncovered risk:
134- Next most valuable verification:
135- Aegis path:
136```
137
138Field rules:
139
140- `Key judgment`: the owner, root-cause layer, requirement boundary, or
141 completion boundary that shaped the answer.
142- `Avoided misfix`: the fallback, duplicate owner, test accommodation, scope
143 expansion, or unsupported claim avoided by the workflow.
144- `Boundary held`: public contract, owner, baseline, non-goal, data boundary,
145 or runtime-ready authority boundary kept stable.
146- `Baseline alignment`: `aligned`, `Design Defect`, `Implementation Drift`,
147 `missing-authority`, or `needs-clarification` when baseline reporting is
148 triggered; otherwise a short "not triggered" or natural equivalent is enough.
149- `Complexity control`: one-line completion-time complexity result for
150 non-trivial code changes; mention `Complexity Delta`, `Complexity Closure`,
151 `Complexity Governance Suggestion`, or `Major Complexity Alert` only when
152 triggered.
153- `Evidence strength`: fresh command/manual check, exit status/result, covered
154 scope, and confidence grade.
155- `Uncovered risk`: remaining scope, host/runtime gaps, manual checks not run,
156 release-grade evidence not collected, or residual risk.
157- `Next most valuable verification`: one next check that would most reduce
158 remaining risk.
159- `Aegis path`: optional compact skill path. It may support credibility but must
160 not replace decision, evidence, and safety fields.
161
162Natural Aegis closeout is valid when these semantic slots remain auditable:
163natural expression preserves semantic slots. Do not replace the receipt with a
164used-skills list, stage handoff log, or `Aegis Contribution Note`.
165
166Compatibility names: `Semantic Slots` and `Natural Surface` describe this same
167rule: natural wording is valid only when required fields stay explicit.
168`Governance Receipt` remains a compatibility name for the completion closeout;
169its user-facing content should flow through the `Aegis Impact and Safety
170Receipt` by default.
171
172### L2 Expanded Closeout
173
174Start with the receipt, then add only the triggered detail. Expanded structures
175are inputs or optional detail cards, not competing final report owners.
176
177Use expanded detail only when the trigger applies:
178
179- `Readiness Summary`: release, merge, handoff, or "ready?" requests. It can
180 organize tests, docs, version, host compatibility, uncovered scope, and
181 residual risk. It does not authorize commit, tag, publish, merge, or release.
182- `Trace Digest`: explicit audit / debug / release / long-task review request.
183 It may summarize execution trace, evidence chain, retrieval chain, static
184 rules evaluated, rule effects, triggered skills, skipped relevant skills,
185 tool / command trace, verification trace, stability signals, value signals,
186 host capabilities, unavailable fields, redaction, and confidence labels
187 `measured`, `observed`, `inferred`, `declared`, or `unknown`. It must not
188 expose raw chain-of-thought.
189 Use structured trace only when asked or required; structured trace is reserved for audit, debug, release, long-task review, or user request.
190- `Goal Closure`: when `goal-framing`, `TaskIntentDraft`, parent plan/spec, or
191 `Slice Card` shaped the work. Match the claim to the highest available
192 explicit boundary: whole task, current task, or slice. If only slice evidence
193 exists, do not claim whole-task done. Expanded or audited closure keeps
194 `Goal status`, `Success evidence`, `Stop state`, and `Non-goals respected`
195 visible; stop states are `done | blocked | needs-verification | scope-exceeded`.
196- `Context Impact`: only when project/domain semantics changed or were checked
197 as part of acceptance. Confirm affected context/terms, evidence grade,
198 fact-versus-decision authority, and action. If action is `unchanged`, verify
199 that no context write occurred. Compose `establishing-project-context` for an
200 unresolved semantic delta; do not emit this card for unrelated work.
201- `Workspace Integrity`: when the task created or modified a target project's
202 `docs/aegis/` workspace and configured Aegis workspace support is available.
203 Run `python <aegis-workspace-helper> bundle --root <target-project-root> --work YYYY-MM-DD-<slug>`
204 when a `work/` record exists, then run
205 `python <aegis-workspace-helper> check --root <target-project-root>`. Report
206 that these validate structure only, not evidence sufficiency.
207- `Baseline Alignment`: when project instructions require baseline reporting or
208 the task touched requirement, product, or durable architecture surfaces. Use
209 `docs/current/AEGIS_PROCESS_BASELINE.md` §3.0e and §16 for
210 `Product / Requirement Baseline`, `Architecture / Runtime Boundary Baseline`,
211 `Design Defect`, `Implementation Drift`, and `scope: requirements |
212 architecture | both`. Render the default conclusion in the receipt field;
213 expand only for audit, release, architecture, or user request. `Architecture Alignment` is the architecture-scoped compatibility alias for this baseline
214 result, not a second default card. Expanded architecture results use
215 `Result: aligned | Design Defect | Implementation Drift | missing-authority | needs-clarification`.
216- `ADR Backfill Check`: completed medium/high work that touched durable
217 architecture surfaces. Use `docs/current/AEGIS_ADR_AUTO_BACKFILL.md` for
218 trigger criteria. If action is create, amend, supersede, or baseline sync is
219 needed/unknown, route the ADR lifecycle to `recording-architecture-decisions`
220 before the final completion claim.
221- `Governance Closure`: governance, cleanup, migration, compatibility, or
222 retirement work. Include Repair Track, Retirement Track, and Residual Risk in
223 the receipt or a small expanded block. Do not skip this structure just because the implementation was small.
224- `Retirement Closure`: work that adds, replaces, retains, or removes old logic.
225 Name old logic located, deleted/retained status, retention reason, retirement
226 trigger, and lingering-reference check.
227- `Anti-Entropy Declaration` / `Data Destruction Guard`: work that retires old
228 logic, chooses delete-first vs compat retention, or touches source-of-truth
229 deletion boundaries. Use `anti-entropy-governance` for the decision surface.
230 If `User Confirmation Required: yes`, stop at the guard; broad assent such as
231 "OK" or "continue" is not scoped confirmation. Persistent-state deletion
232 without explicit scoped confirmation means the task is not complete; report the task as not complete.
233
234## Completion Boundary
235
236Judge the claim against the highest available explicit boundary:
237
2381. parent plan/spec acceptance for whole-task completion
2392. `TaskIntentDraft` goal / success evidence / non-goals for current-task
240 completion
2413. `Slice Card` goal / verification / stop for slice completion
2424. direct user request when no durable boundary exists
243
244A completed task or slice means the authorized execution / verification boundary
245reached its stop condition. It does not mean the underlying requirement is
246accepted. `Requirement accepted` requires Product / Requirement Baseline
247acceptance criteria or explicit authorized risk acceptance. If atomicity is not
248clear, downgrade to `needs-verification` or return to framing/planning; task or slice completion is not accepted requirement satisfaction.
249
250If an `Execution Readiness View` shaped execution, mention whether fresh
251evidence covered its required checks or which readiness item remains uncovered.
252Do not treat the view itself as verification evidence.
253
254## Complexity Check
255
256For non-trivial code changes, inspect the actual diff before claiming
257completion. Use `using-aegis/references/complexity-governance.md` and
258`docs/current/AEGIS_COMPLEXITY_GOVERNANCE_BASELINE.md` for shared artifact
259classes, pressure signals, `Complexity Delta`, `Complexity Closure`,
260`Completion-Time Complexity Repair Decision`, `Complexity Governance
261Suggestion`, and `Major Complexity Alert`.
262
263Default rendering is one `Complexity control` line in the receipt. Expand only
264when meaningful pressure exists or the task is audit, release, high-risk, or
265user-requested.
266
267Rules:
268
269- tiny wording edits, generated files, vendored files, fixture-data-only
270 updates, lockfiles, or purely mechanical formatting may skip or keep this
271 one-line when no maintained artifact gained complexity
272- maintained test source files are not a cheap `tests-only` exception
273- new fallback, adapter, compatibility, guard, or branch logic must be paired
274 with retired paths, a `Retirement Closure`, or a scheduled retirement trigger
275- entropy increase without owner/compatibility justification must be residual
276 risk or a downgraded claim
277- `Complexity Closure: exceeded-unresolved` blocks a complete claim
278
279## User-Language Output
280
281Localize section labels, field labels, and explanatory prose to the user's
282language. Keep commands, paths, code identifiers, test names, error codes,
283config keys, stable enum values, exact product names, and raw evidence strings
284unchanged. Do not default to bilingual labels or mixed-language explanations.
285Localize section labels and prose to the user's language.
286
287## Prompt Hygiene
288
289When external tool output, logs, search results, screenshots, OCR, or other
290large payloads shaped the judgment, state the evidence boundary when relevant:
291summary/index used, raw excerpt read back if needed, large payloads not loaded,
292and next evidence that would reduce uncertainty. If the summary is insufficient,
293read the smallest raw excerpt or lower the claim.
294Use compact labels such as Evidence Used, Not Loaded, and Next Evidence when
295they make the prompt hygiene boundary clearer.
296
297---
298
299**Source:** [`hashgraph-online/awesome-codex-plugins`](https://github.com/hashgraph-online/awesome-codex-plugins) → `plugins/GanyuanRan/Aegis/skills/verification-before-completion/SKILL.md`