← LightbringerCONTENT HISTORY

Update to Lightbringer

Snapshot Nov 14, 2026 · 00:23 UTC · version 5.6.1

Collection source: downloaded plugin package.

WHAT CHANGED

Instructions updated for patent-review

Full description — Now says: report, patent-draft and validation reviews. Use for review instructions, attorney redlines, proposed amendments, document comments, review discussion,...

Observed in instructions or declared skills. Runtime behavior has not been tested.

Save Lightbringer and check future changes in Following.

Up to 3 plugins free · no card required.

All fields and original values.

Full description

Before

Oct 9, 2026 · 12:13 UTC

report and patent-draft reviews. Use for attorney redlines, proposed amendments, document comments, review discussion

After

Nov 14, 2026 · 00:23 UTC

report, patent-draft and validation reviews. Use for review instructions, attorney redlines, proposed amendments, document comments, review discussion, validation country selection

Skill instructions

Before

Oct 9, 2026 · 12:13 UTC

report and patent-draft reviews. Use for attorney redlines, proposed amendments, document comments, review discussion and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separat...

After

Nov 14, 2026 · 00:23 UTC

report, patent-draft and validation reviews. Use for review instructions, attorney redlines, proposed amendments, document comments, review discussion, validation country selection and explicit approve/request-changes responses. Innovati...

Supporting files

Before

Oct 9, 2026 · 12:13 UTC

[{"relative_path":"references/review-workflows.md","size_in_bytes":6381},{"relative_path":"references/service-conduct.md","size_in_bytes":1599}]

After

Nov 14, 2026 · 00:23 UTC

[{"relative_path":"references/review-workflows.md","size_in_bytes":7471},{"relative_path":"references/service-conduct.md","size_in_bytes":1293},{"relative_path":"references/unresolved-needs.md","size_in_bytes":4243}]

Compare saved observations

Download comparison JSON

SKILL.md text changes

− removed · + added

--- before
+++ after
@@ -1,36 +1,43 @@
 ---
 name: patent-review
-description: Help the user read and respond to Lightbringer Strategy, report and patent-draft reviews. Use for attorney redlines, proposed amendments, document comments, review discussion and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separate workflows.
+description: Help the user read and respond to Lightbringer Strategy, report, patent-draft and validation reviews. Use for review instructions, attorney redlines, proposed amendments, document comments, review discussion, validation country selection and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separate workflows.
 ---
 
 # Patent review
 
-Read the review and its actual artifacts, prepare sourced feedback in the user's voice, and post only authorised comments or decisions. Read [service conduct](references/service-conduct.md) and [review guidance](references/review-workflows.md).
+Read the review, its instructions and its actual artifacts, prepare sourced feedback in the user's voice, and post only authorised comments or decisions. Read [service conduct](references/service-conduct.md) and [review guidance](references/review-workflows.md).
 
-Use the connected organisation and accessible reviews where the user is a participant or creator. Locate the requested review with `list_reviews`, preserving its `target` (`strategy`, `report` or `document`) and available Strategy, report, case and innovation references. Do not relabel a Strategy review as a report. Then read `get_review`, including the document, threads and review discussion. If the review is ambiguous, clarify it. If document rendering is unavailable, explain the limitation and offer the platform link; do not infer missing content or recommend approval from metadata alone.
+Use the connected organisation and accessible reviews where the user is a participant or creator. Locate the requested review with `list_reviews`, preserving its `target` (`strategy`, `report` or `document`), its type and the available Strategy, report, case and innovation references. Do not relabel a Strategy review as a report, and do not handle an unrecognised review type as a patent draft. When the host presents the review list as a component, let it serve as the list; do not restate every review in chat. Then read `get_review`, including the instructions, the document, threads and review discussion. Show the review's instructions before interpreting the material; if none are recorded, say so instead of supplying a brief. If the review is ambiguous, clarify it. If document rendering is unavailable, explain the limitation and offer the platform link; do not infer missing content or recommend approval from metadata alone.
 
 Comments belong to a review. Use `add_comment` for an exact document passage, `reply_to_comment` with a stable document comment ID for a thread, and `add_discussion_comment` for a general review remark. If a required action is unavailable, keep the feedback as a draft and provide the platform continuation.
 
 ## Authorisation
 
-Autonomy: comments post in the user's name, so nothing posts unapproved. Assemble one confirmation batch — each item verbatim with its document anchor, thread ID or review discussion destination, classification, and provenance tag — then post approved items without further questions. `respond_to_review` is never covered by batch approval: it needs the user's explicit approve/request-changes instruction, message confirmed verbatim. Keep the optional response message within 2,000 characters, including any provenance text; if a longer message needs shortening, get approval for the revised text before sending. An approval cannot be withdrawn through this tool. These review tools can trigger in-app or email notifications; posting success is not proof of notification delivery.
+Autonomy: comments post in the user's name, so nothing posts unapproved. Assemble one confirmation batch — each item verbatim with its document anchor, thread ID or review discussion destination, classification, and provenance tag — then post approved items without further questions. `respond_to_review` is never covered by batch approval: it needs the user's explicit approve/request-changes instruction, message confirmed verbatim. Keep the optional response message within 2,000 characters, including any provenance text; if a longer message needs shortening, get approval for the revised text before sending. An approval cannot be withdrawn through this tool. `set_validation_countries` is not covered by batch approval either: it needs the user's explicit complete selection. These review tools can trigger in-app or email notifications; posting success is not proof of notification delivery.
 
 ## Strategy or report review
 
 1. `list_reviews` (open) → the review carrying the selected Strategy or report. No review → `search`/`fetch` read-only, and tell the user commenting requires one.
-2. `get_review`; read the document, all threads and the review discussion before drafting — a point already in a thread becomes a reply, never a new comment.
-3. Assess: internal consistency, unsupported claims, portfolio and strategy fit (`search`/`fetch`/`get_innovation` for context), plus the user's specific asks.
+2. `get_review`; read the instructions, the document, all threads and the review discussion before drafting — a point already in a thread becomes a reply, never a new comment.
+3. Assess against the instructions: internal consistency, unsupported claims, portfolio and strategy fit (`search`/`fetch`/`get_innovation` for context), plus the user's specific asks.
 4. Classify (rules below) → discovery screen → batch confirmation → `add_comment` with exact `quoted_text` anchors (`target_user_ids` only on request); general review remarks use `add_discussion_comment` when available.
 5. Summarise in chat; no report file.
 
 ## Patent draft review
 
-1. `list_reviews` → `get_review` (document, threads, attorney redlines with original/replacement/rationale, response status). Inspect proposed amendments separately from attorney redlines; the document body may still show baseline text. Follow the reference when rendering is unavailable.
+1. `list_reviews` → `get_review` (instructions, document, threads, attorney redlines with original/replacement/rationale, response status). Inspect proposed amendments separately from attorney redlines; the document body may still show baseline text. Follow the reference when rendering is unavailable.
 2. Triage threads addressed to or awaiting the user; use `get_innovation` on the source innovation description for claim-support questions.
-3. Review against the checklist in the reference (claim support, innovation description consistency, terminology, embodiments, strategy fit, redlines).
+3. Review against the instructions and the checklist in the reference (claim support, innovation description consistency, terminology, embodiments, strategy fit, redlines).
 4. Classify each item → discovery screen → batch confirmation → post: replies via `reply_to_comment` (explicit `comment_id`; never a thread-anchored `add_comment`), new passage-specific feedback via `add_comment`, and general review remarks via `add_discussion_comment` when available.
 5. `respond_to_review` only as above. Summarise in chat.
 
+## Validation review
+
+1. `get_review` → the instructions and the recommended and currently selected countries in `validation`.
+2. Ask for the user's complete intended selection. Do not choose validation scope for the user. An unchanged selection needs no write.
+3. Before `set_validation_countries`, explain that it replaces the whole selection and returns earlier approvals to pending, and that restoring an earlier list does not restore them. Only a reviewer who has not yet approved, or a specialist, can change an open review.
+4. Read back with `get_review` and confirm the saved selection. Selection does not approve the review; approval is a separate `respond_to_review` decision as above. If the selection tool is not advertised, continue in the platform.
+
 ### Feedback rules (binding; boundary cases in the reference)
 
 - **Persona**: every posted item speaks as the inventor/engineer — never attorney tone, never advice requiring patent expertise (claim scope or wording, prosecution tactics, prior-art positioning, legal characterisations). Report what was built and observed; the rest is the drafting team's.
@@ -49,4 +56,6 @@
 
 ## Completion
 
-Summarise what was read, which approved comments or decisions were recorded, and anything still pending. Include the review link when supplied. A drafted response is not a posted response; comments do not constitute formal approval. For capture or enrichment outside the review, use `innovation-capture` when installed. A separate request to start patent preparation belongs to `patent-preparation`.
+Summarise what was read, which approved comments, selections or decisions were recorded, and anything still pending. Include the review link when supplied. A drafted response is not a posted response; comments do not constitute formal approval, and personal approval does not establish that the review is closed. For capture or enrichment outside the review, use `innovation-capture` when installed. A separate request to start patent preparation belongs to `patent-preparation`.
+
+If useful assistance leaves a material customer problem unresolved, follow the [unresolved-needs guidance](references/unresolved-needs.md). Keep helping; reporting is not a prerequisite and requires approval of the exact report and recipient.
Full technical diff · 3 changed fields

changed /description

BEFORE
"Help the user read and respond to Lightbringer Strategy, report and patent-draft reviews. Use for attorney redlines, proposed amendments, document comments, review discussion and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separate workflows."
AFTER
"Help the user read and respond to Lightbringer Strategy, report, patent-draft and validation reviews. Use for review instructions, attorney redlines, proposed amendments, document comments, review discussion, validation country selection and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separate workflows."

changed /included_files

BEFORE
[
  {
    "relative_path": "references/review-workflows.md",
    "size_in_bytes": 6381
  },
  {
    "relative_path": "references/service-conduct.md",
    "size_in_bytes": 1599
  }
]
AFTER
[
  {
    "relative_path": "references/review-workflows.md",
    "size_in_bytes": 7471
  },
  {
    "relative_path": "references/service-conduct.md",
    "size_in_bytes": 1293
  },
  {
    "relative_path": "references/unresolved-needs.md",
    "size_in_bytes": 4243
  }
]

changed /skill_md_contents

BEFORE
"---\nname: patent-review\ndescription: Help the user read and respond to Lightbringer Strategy, report and patent-draft reviews. Use for attorney redlines, proposed amendments, document comments, review discussion and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separate workflows.\n---\n\n# Patent review\n\nRead the review and its actual artifacts, prepare sourced feedback in the user's voice, and post only authorised comments or decisions. Read [service conduct](references/service-conduct.md) and [review guidance](references/review-workflows.md).\n\nUse the connected organisation and accessible reviews where the user is a participant or creator. Locate the requested review with `list_reviews`, preserving its `target` (`strategy`, `report` or `document`) and available Strategy, report, case and innovation references. Do not relabel a Strategy review as a report. Then read `get_review`, including the document, threads and review discussion. If the review is ambiguous, clarify it. If document rendering is unavailable, explain the limitation and offer the platform link; do not infer missing content or recommend approval from metadata alone.\n\nComments belong to a review. Use `add_comment` for an exact document passage, `reply_to_comment` with a stable document comment ID for a thread, and `add_discussion_comment` for a general review remark. If a required action is unavailable, keep the feedback as a draft and provide the platform continuation.\n\n## Authorisation\n\nAutonomy: comments post in the user's name, so nothing posts unapproved. Assemble one confirmation batch — each item verbatim with its document anchor, thread ID or review discussion destination, classification, and provenance tag — then post approved items without further questions. `respond_to_review` is never covered by batch approval: it needs the user's explicit approve/request-changes instruction, message confirmed verbatim. Keep the optional response message within 2,000 characters, including any provenance text; if a longer message needs shortening, get approval for the revised text before sending. An approval cannot be withdrawn through this tool. These review tools can trigger in-app or email notifications; posting success is not proof of notification delivery.\n\n## Strategy or report review\n\n1. `list_reviews` (open) → the review carrying the selected Strategy or report. No review → `search`/`fetch` read-only, and tell the user commenting requires one.\n2. `get_review`; read the document, all threads and the review discussion before drafting — a point already in a thread becomes a reply, never a new comment.\n3. Assess: internal consistency, unsupported claims, portfolio and strategy fit (`search`/`fetch`/`get_innovation` for context), plus the user's specific asks.\n4. Classify (rules below) → discovery screen → batch confirmation → `add_comment` with exact `quoted_text` anchors (`target_user_ids` only on request); general review remarks use `add_discussion_comment` when available.\n5. Summarise in chat; no report file.\n\n## Patent draft review\n\n1. `list_reviews` → `get_review` (document, threads, attorney redlines with original/replacement/rationale, response status). Inspect proposed amendments separately from attorney redlines; the document body may still show baseline text. Follow the reference when rendering is unavailable.\n2. Triage threads addressed to or awaiting the user; use `get_innovation` on the source innovation description for claim-support questions.\n3. Review against the checklist in the reference (claim support, innovation description consistency, terminology, embodiments, strategy fit, redlines).\n4. Classify each item → discovery screen → batch confirmation → post: replies via `reply_to_comment` (explicit `comment_id`; never a thread-anchored `add_comment`), new passage-specific feedback via `add_comment`, and general review remarks via `add_discussion_comment` when available.\n5. `respond_to_review` only as above. Summarise in chat.\n\n### Feedback rules (binding; boundary cases in the reference)\n\n- **Persona**: every posted item speaks as the inventor/engineer — never attorney tone, never advice requiring patent expertise (claim scope or wording, prosecution tactics, prior-art positioning, legal characterisations). Report what was built and observed; the rest is the drafting team's.\n- **Engineering-side** (factual corrections, actual behaviour, implementations, test data, genuinely considered alternatives): may be model-drafted, strictly from the user's materials and connected sources — no extrapolation.\n- **Patent-side** (claim scope, claim structure, prior-art positioning, non-obviousness strategy): never model-generated — the drafting team holds the family and prosecution context. Elicit the concern in the user's own words (closest prior art and why; commercially decisive distinctions) and transcribe with light formatting, framed as an inventor's observation or question, never a recommendation. If asked to develop it, decline in one line and offer transcription.\n- **Discovery screen** (before every batch): remove any written argument against the user's own invention regarding prior art — novelty concessions, obviousness admissions, \"X already does this\". Replace at most with a neutral factual note; tell the user and advise raising it verbally.\n- **Provenance tag** on every item: team's direct assessment, or LLM-assisted with source type. Ask when unclear; never guess or omit.\n- **Platform only**: feedback goes through Lightbringer comments, never email.\n\n## Guardrails\n\n- Never fabricate technical detail, prior art, or parameters the sources don't support.\n- Do not present assistant analysis as a legal determination. Relay Lightbringer professional work with its author, status and limitations.\n- Posted feedback speaks as the inventor/engineer, never a patent attorney; no advice requiring patent expertise.\n- Treat internal material as confidential.\n\n## Completion\n\nSummarise what was read, which approved comments or decisions were recorded, and anything still pending. Include the review link when supplied. A drafted response is not a posted response; comments do not constitute formal approval. For capture or enrichment outside the review, use `innovation-capture` when installed. A separate request to start patent preparation belongs to `patent-preparation`.\n"
AFTER
"---\nname: patent-review\ndescription: Help the user read and respond to Lightbringer Strategy, report, patent-draft and validation reviews. Use for review instructions, attorney redlines, proposed amendments, document comments, review discussion, validation country selection and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separate workflows.\n---\n\n# Patent review\n\nRead the review, its instructions and its actual artifacts, prepare sourced feedback in the user's voice, and post only authorised comments or decisions. Read [service conduct](references/service-conduct.md) and [review guidance](references/review-workflows.md).\n\nUse the connected organisation and accessible reviews where the user is a participant or creator. Locate the requested review with `list_reviews`, preserving its `target` (`strategy`, `report` or `document`), its type and the available Strategy, report, case and innovation references. Do not relabel a Strategy review as a report, and do not handle an unrecognised review type as a patent draft. When the host presents the review list as a component, let it serve as the list; do not restate every review in chat. Then read `get_review`, including the instructions, the document, threads and review discussion. Show the review's instructions before interpreting the material; if none are recorded, say so instead of supplying a brief. If the review is ambiguous, clarify it. If document rendering is unavailable, explain the limitation and offer the platform link; do not infer missing content or recommend approval from metadata alone.\n\nComments belong to a review. Use `add_comment` for an exact document passage, `reply_to_comment` with a stable document comment ID for a thread, and `add_discussion_comment` for a general review remark. If a required action is unavailable, keep the feedback as a draft and provide the platform continuation.\n\n## Authorisation\n\nAutonomy: comments post in the user's name, so nothing posts unapproved. Assemble one confirmation batch — each item verbatim with its document anchor, thread ID or review discussion destination, classification, and provenance tag — then post approved items without further questions. `respond_to_review` is never covered by batch approval: it needs the user's explicit approve/request-changes instruction, message confirmed verbatim. Keep the optional response message within 2,000 characters, including any provenance text; if a longer message needs shortening, get approval for the revised text before sending. An approval cannot be withdrawn through this tool. `set_validation_countries` is not covered by batch approval either: it needs the user's explicit complete selection. These review tools can trigger in-app or email notifications; posting success is not proof of notification delivery.\n\n## Strategy or report review\n\n1. `list_reviews` (open) → the review carrying the selected Strategy or report. No review → `search`/`fetch` read-only, and tell the user commenting requires one.\n2. `get_review`; read the instructions, the document, all threads and the review discussion before drafting — a point already in a thread becomes a reply, never a new comment.\n3. Assess against the instructions: internal consistency, unsupported claims, portfolio and strategy fit (`search`/`fetch`/`get_innovation` for context), plus the user's specific asks.\n4. Classify (rules below) → discovery screen → batch confirmation → `add_comment` with exact `quoted_text` anchors (`target_user_ids` only on request); general review remarks use `add_discussion_comment` when available.\n5. Summarise in chat; no report file.\n\n## Patent draft review\n\n1. `list_reviews` → `get_review` (instructions, document, threads, attorney redlines with original/replacement/rationale, response status). Inspect proposed amendments separately from attorney redlines; the document body may still show baseline text. Follow the reference when rendering is unavailable.\n2. Triage threads addressed to or awaiting the user; use `get_innovation` on the source innovation description for claim-support questions.\n3. Review against the instructions and the checklist in the reference (claim support, innovation description consistency, terminology, embodiments, strategy fit, redlines).\n4. Classify each item → discovery screen → batch confirmation → post: replies via `reply_to_comment` (explicit `comment_id`; never a thread-anchored `add_comment`), new passage-specific feedback via `add_comment`, and general review remarks via `add_discussion_comment` when available.\n5. `respond_to_review` only as above. Summarise in chat.\n\n## Validation review\n\n1. `get_review` → the instructions and the recommended and currently selected countries in `validation`.\n2. Ask for the user's complete intended selection. Do not choose validation scope for the user. An unchanged selection needs no write.\n3. Before `set_validation_countries`, explain that it replaces the whole selection and returns earlier approvals to pending, and that restoring an earlier list does not restore them. Only a reviewer who has not yet approved, or a specialist, can change an open review.\n4. Read back with `get_review` and confirm the saved selection. Selection does not approve the review; approval is a separate `respond_to_review` decision as above. If the selection tool is not advertised, continue in the platform.\n\n### Feedback rules (binding; boundary cases in the reference)\n\n- **Persona**: every posted item speaks as the inventor/engineer — never attorney tone, never advice requiring patent expertise (claim scope or wording, prosecution tactics, prior-art positioning, legal characterisations). Report what was built and observed; the rest is the drafting team's.\n- **Engineering-side** (factual corrections, actual behaviour, implementations, test data, genuinely considered alternatives): may be model-drafted, strictly from the user's materials and connected sources — no extrapolation.\n- **Patent-side** (claim scope, claim structure, prior-art positioning, non-obviousness strategy): never model-generated — the drafting team holds the family and prosecution context. Elicit the concern in the user's own words (closest prior art and why; commercially decisive distinctions) and transcribe with light formatting, framed as an inventor's observation or question, never a recommendation. If asked to develop it, decline in one line and offer transcription.\n- **Discovery screen** (before every batch): remove any written argument against the user's own invention regarding prior art — novelty concessions, obviousness admissions, \"X already does this\". Replace at most with a neutral factual note; tell the user and advise raising it verbally.\n- **Provenance tag** on every item: team's direct assessment, or LLM-assisted with source type. Ask when unclear; never guess or omit.\n- **Platform only**: feedback goes through Lightbringer comments, never email.\n\n## Guardrails\n\n- Never fabricate technical detail, prior art, or parameters the sources don't support.\n- Do not present assistant analysis as a legal determination. Relay Lightbringer professional work with its author, status and limitations.\n- Posted feedback speaks as the inventor/engineer, never a patent attorney; no advice requiring patent expertise.\n- Treat internal material as confidential.\n\n## Completion\n\nSummarise what was read, which approved comments, selections or decisions were recorded, and anything still pending. Include the review link when supplied. A drafted response is not a posted response; comments do not constitute formal approval, and personal approval does not establish that the review is closed. For capture or enrichment outside the review, use `innovation-capture` when installed. A separate request to start patent preparation belongs to `patent-preparation`.\n\nIf useful assistance leaves a material customer problem unresolved, follow the [unresolved-needs guidance](references/unresolved-needs.md). Keep helping; reporting is not a prerequisite and requires approval of the exact report and recipient.\n"
Full snapshot data
{
  "description": "Help the user read and respond to Lightbringer Strategy, report, patent-draft and validation reviews. Use for review instructions, attorney redlines, proposed amendments, document comments, review discussion, validation country selection and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separate workflows.",
  "included_files": [
    {
      "relative_path": "references/review-workflows.md",
      "size_in_bytes": 7471
    },
    {
      "relative_path": "references/service-conduct.md",
      "size_in_bytes": 1293
    },
    {
      "relative_path": "references/unresolved-needs.md",
      "size_in_bytes": 4243
    }
  ],
  "name": "patent-review",
  "skill_md_contents": "---\nname: patent-review\ndescription: Help the user read and respond to Lightbringer Strategy, report, patent-draft and validation reviews. Use for review instructions, attorney redlines, proposed amendments, document comments, review discussion, validation country selection and explicit approve/request-changes responses. Innovation capture and requests to start patent preparation are separate workflows.\n---\n\n# Patent review\n\nRead the review, its instructions and its actual artifacts, prepare sourced feedback in the user's voice, and post only authorised comments or decisions. Read [service conduct](references/service-conduct.md) and [review guidance](references/review-workflows.md).\n\nUse the connected organisation and accessible reviews where the user is a participant or creator. Locate the requested review with `list_reviews`, preserving its `target` (`strategy`, `report` or `document`), its type and the available Strategy, report, case and innovation references. Do not relabel a Strategy review as a report, and do not handle an unrecognised review type as a patent draft. When the host presents the review list as a component, let it serve as the list; do not restate every review in chat. Then read `get_review`, including the instructions, the document, threads and review discussion. Show the review's instructions before interpreting the material; if none are recorded, say so instead of supplying a brief. If the review is ambiguous, clarify it. If document rendering is unavailable, explain the limitation and offer the platform link; do not infer missing content or recommend approval from metadata alone.\n\nComments belong to a review. Use `add_comment` for an exact document passage, `reply_to_comment` with a stable document comment ID for a thread, and `add_discussion_comment` for a general review remark. If a required action is unavailable, keep the feedback as a draft and provide the platform continuation.\n\n## Authorisation\n\nAutonomy: comments post in the user's name, so nothing posts unapproved. Assemble one confirmation batch — each item verbatim with its document anchor, thread ID or review discussion destination, classification, and provenance tag — then post approved items without further questions. `respond_to_review` is never covered by batch approval: it needs the user's explicit approve/request-changes instruction, message confirmed verbatim. Keep the optional response message within 2,000 characters, including any provenance text; if a longer message needs shortening, get approval for the revised text before sending. An approval cannot be withdrawn through this tool. `set_validation_countries` is not covered by batch approval either: it needs the user's explicit complete selection. These review tools can trigger in-app or email notifications; posting success is not proof of notification delivery.\n\n## Strategy or report review\n\n1. `list_reviews` (open) → the review carrying the selected Strategy or report. No review → `search`/`fetch` read-only, and tell the user commenting requires one.\n2. `get_review`; read the instructions, the document, all threads and the review discussion before drafting — a point already in a thread becomes a reply, never a new comment.\n3. Assess against the instructions: internal consistency, unsupported claims, portfolio and strategy fit (`search`/`fetch`/`get_innovation` for context), plus the user's specific asks.\n4. Classify (rules below) → discovery screen → batch confirmation → `add_comment` with exact `quoted_text` anchors (`target_user_ids` only on request); general review remarks use `add_discussion_comment` when available.\n5. Summarise in chat; no report file.\n\n## Patent draft review\n\n1. `list_reviews` → `get_review` (instructions, document, threads, attorney redlines with original/replacement/rationale, response status). Inspect proposed amendments separately from attorney redlines; the document body may still show baseline text. Follow the reference when rendering is unavailable.\n2. Triage threads addressed to or awaiting the user; use `get_innovation` on the source innovation description for claim-support questions.\n3. Review against the instructions and the checklist in the reference (claim support, innovation description consistency, terminology, embodiments, strategy fit, redlines).\n4. Classify each item → discovery screen → batch confirmation → post: replies via `reply_to_comment` (explicit `comment_id`; never a thread-anchored `add_comment`), new passage-specific feedback via `add_comment`, and general review remarks via `add_discussion_comment` when available.\n5. `respond_to_review` only as above. Summarise in chat.\n\n## Validation review\n\n1. `get_review` → the instructions and the recommended and currently selected countries in `validation`.\n2. Ask for the user's complete intended selection. Do not choose validation scope for the user. An unchanged selection needs no write.\n3. Before `set_validation_countries`, explain that it replaces the whole selection and returns earlier approvals to pending, and that restoring an earlier list does not restore them. Only a reviewer who has not yet approved, or a specialist, can change an open review.\n4. Read back with `get_review` and confirm the saved selection. Selection does not approve the review; approval is a separate `respond_to_review` decision as above. If the selection tool is not advertised, continue in the platform.\n\n### Feedback rules (binding; boundary cases in the reference)\n\n- **Persona**: every posted item speaks as the inventor/engineer — never attorney tone, never advice requiring patent expertise (claim scope or wording, prosecution tactics, prior-art positioning, legal characterisations). Report what was built and observed; the rest is the drafting team's.\n- **Engineering-side** (factual corrections, actual behaviour, implementations, test data, genuinely considered alternatives): may be model-drafted, strictly from the user's materials and connected sources — no extrapolation.\n- **Patent-side** (claim scope, claim structure, prior-art positioning, non-obviousness strategy): never model-generated — the drafting team holds the family and prosecution context. Elicit the concern in the user's own words (closest prior art and why; commercially decisive distinctions) and transcribe with light formatting, framed as an inventor's observation or question, never a recommendation. If asked to develop it, decline in one line and offer transcription.\n- **Discovery screen** (before every batch): remove any written argument against the user's own invention regarding prior art — novelty concessions, obviousness admissions, \"X already does this\". Replace at most with a neutral factual note; tell the user and advise raising it verbally.\n- **Provenance tag** on every item: team's direct assessment, or LLM-assisted with source type. Ask when unclear; never guess or omit.\n- **Platform only**: feedback goes through Lightbringer comments, never email.\n\n## Guardrails\n\n- Never fabricate technical detail, prior art, or parameters the sources don't support.\n- Do not present assistant analysis as a legal determination. Relay Lightbringer professional work with its author, status and limitations.\n- Posted feedback speaks as the inventor/engineer, never a patent attorney; no advice requiring patent expertise.\n- Treat internal material as confidential.\n\n## Completion\n\nSummarise what was read, which approved comments, selections or decisions were recorded, and anything still pending. Include the review link when supplied. A drafted response is not a posted response; comments do not constitute formal approval, and personal approval does not establish that the review is closed. For capture or enrichment outside the review, use `innovation-capture` when installed. A separate request to start patent preparation belongs to `patent-preparation`.\n\nIf useful assistance leaves a material customer problem unresolved, follow the [unresolved-needs guidance](references/unresolved-needs.md). Keep helping; reporting is not a prerequisite and requires approval of the exact report and recipient.\n"
}

SHA-256 of public snapshot: 61b8d799787f59796fa5946fb2c1daaceac60d7edd8f4f1294dfb765ca0b3c9e