From d734c4f8c4c902ae3826c37a6d5b57d58dfe4ba4 Mon Sep 17 00:00:00 2001 From: Robert Ray <215331614+TheRobertRay@users.noreply.github.com> Date: Mon, 1 Jun 2026 10:14:08 -0700 Subject: [PATCH] Add conversation capture receipts skill --- CHANGELOG.md | 5 ++ README.md | 4 +- docs/workflows.md | 10 +++ examples/conversation-capture-receipt.md | 28 ++++++++ scripts/validate.sh | 2 + skills/conversation-capture-receipts/SKILL.md | 64 +++++++++++++++++++ 6 files changed, 112 insertions(+), 1 deletion(-) create mode 100644 examples/conversation-capture-receipt.md create mode 100644 skills/conversation-capture-receipts/SKILL.md diff --git a/CHANGELOG.md b/CHANGELOG.md index 9be4610..d89f52b 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,10 @@ # Changelog +## 0.2.0 - 2026-06-01 + +- Added a conversation capture receipts skill for turning rough chat, voice-note, or inbox-dump input into human-useful no-write receipts. +- Added a sanitized conversation capture example and workflow gate. + ## 0.1.1 - 2026-06-01 - Added a publication-risk scanner for catching private paths, credential-like tokens, and internal hostnames before release. diff --git a/README.md b/README.md index d8d8a25..090ee3a 100644 --- a/README.md +++ b/README.md @@ -5,6 +5,7 @@ Reusable prompt contracts, review gates, and maintenance workflows for running c This repository is intentionally small. It packages field-tested operator patterns into files that other maintainers can copy into their own repos: - `AGENTS.md` prompt intake contract for ambiguous or dictated tasks +- conversation capture receipts for rough chat, voice notes, and inbox dumps - maintainer triage skill for evidence-first issue and PR review - milestone review skill for substantial implementation work - application packet template for programs that evaluate OSS maintainer signal @@ -55,12 +56,13 @@ bash scripts/check-publication-risk.sh ## Repository Map - `templates/AGENTS.md` - compact global intake and completion-review contract +- `skills/conversation-capture-receipts/SKILL.md` - no-write receipt workflow for messy human input - `skills/oss-maintainer-triage/SKILL.md` - issue, PR, and bounty-style evidence triage - `skills/milestone-review/SKILL.md` - milestone and adversarial completion review workflow - `docs/maintenance-model.md` - the operating model behind the kit - `docs/workflows.md` - copyable maintainer workflows - `docs/publication-risk.md` - how to screen workflow packs before publishing -- `examples/` - sanitized example packets +- `examples/` - sanitized example packets and receipts - `docs/codex-for-oss-application.md` - application packet and field-ready answers - `scripts/validate.sh` - local validation checks - `scripts/check-publication-risk.sh` - conservative private-context leak scanner diff --git a/docs/workflows.md b/docs/workflows.md index 5d4cb7c..4ab2f6e 100644 --- a/docs/workflows.md +++ b/docs/workflows.md @@ -12,6 +12,16 @@ Use when a maintainer needs to decide whether an issue is actionable. 4. State the smallest next action: reproduce, ask for missing evidence, link duplicate, accept, or close. 5. Do not assign implementation work until the failure mode is concrete. +## Conversation Capture Receipt + +Use when a chat agent receives rough dictated input, a messy inbox dump, or a long context note. + +1. Start in no-write mode unless the user already approved a write path. +2. Produce `What I caught` before routing details so the user can tell whether the agent understood the point. +3. Add one to three `Meaning cards` that preserve emotional weight, decisions, open loops, or context value. +4. Separate safe next actions from write-requiring actions such as memory, task, calendar, issue, or outbound-message changes. +5. Ask one narrow clarification only when the next action changes risk, destination, credentials, cost, or user-facing consequences. + ## Pull Request Review Gate Use before approving or merging a PR. diff --git a/examples/conversation-capture-receipt.md b/examples/conversation-capture-receipt.md new file mode 100644 index 0000000..93f13a6 --- /dev/null +++ b/examples/conversation-capture-receipt.md @@ -0,0 +1,28 @@ +# Example: Conversation Capture Receipt + +## Input + +```text +I keep saying I want the assistant to save important context, but I do not want every random thought turned into a task. When I am talking fast, I need it to catch the point, show me what it heard, and only write something permanent after I trust that it understood me. +``` + +## Receipt + +### What I caught + +You want capture to feel safe before it becomes storage. The assistant should understand the point of a messy note, show a small receipt, and avoid turning every thought into a task or record. + +### Meaning cards + +- **Trust boundary:** Understanding comes before persistence. +- **Workflow preference:** Rough input should produce a human-readable receipt, not hidden routing metadata. +- **Open loop:** Durable writes need an approval point unless the destination was already clear. + +### Possible next moves + +- Safe: turn this into a reusable prompt or checklist. +- Needs approval: save it as memory, create a task, file an issue, or send it to another system. + +## Notes + +This example intentionally avoids private project names and service-specific routing. The reusable pattern is the receipt shape and write boundary. diff --git a/scripts/validate.sh b/scripts/validate.sh index d0cdb3c..77c98bf 100755 --- a/scripts/validate.sh +++ b/scripts/validate.sh @@ -14,9 +14,11 @@ required=( "docs/publication-risk.md" "docs/workflows.md" "docs/codex-for-oss-application.md" + "examples/conversation-capture-receipt.md" "examples/issue-triage.md" "examples/pr-review-gate.md" "examples/release-readiness.md" + "skills/conversation-capture-receipts/SKILL.md" "skills/oss-maintainer-triage/SKILL.md" "skills/milestone-review/SKILL.md" "templates/AGENTS.md" diff --git a/skills/conversation-capture-receipts/SKILL.md b/skills/conversation-capture-receipts/SKILL.md new file mode 100644 index 0000000..5f7b45e --- /dev/null +++ b/skills/conversation-capture-receipts/SKILL.md @@ -0,0 +1,64 @@ +--- +name: conversation-capture-receipts +description: Use when Codex needs to process rough dictated notes, chat messages, inbox dumps, or stream-of-consciousness user input into a human-useful no-write receipt before deciding whether to create memory, tasks, issues, calendar items, or other durable records. Helps preserve meaning, emotional weight, decisions, and open loops while keeping write actions gated. +--- + +# Conversation Capture Receipts + +## Goal + +Turn messy human input into a receipt that proves understanding before taking action. The receipt should help the user feel accurately heard, not merely routed. + +## Default Stance + +Start in no-write mode unless the user clearly approved a specific durable write. Treat capture as understanding first, persistence second. + +## Receipt Shape + +Use this structure: + +1. `What I caught` - two to four plain sentences naming the core point in the user's language. +2. `Meaning cards` - one to three short cards that preserve why the input matters. +3. `Possible next moves` - safe actions first, write-requiring actions clearly marked. +4. `Clarifying question` - only when the answer changes risk, destination, credentials, cost, or user-facing consequences. + +## Meaning Cards + +Create cards for signals such as: + +- emotional weight or personal stakes +- a decision the user is circling +- context that should shape future work +- a task that is implied but not yet approved +- a risk, trust boundary, or possible false interpretation + +Keep cards short. Do not convert every sentence into metadata. + +## Write Boundaries + +Before creating memory, tasks, issues, calendar events, outbound messages, or external writes: + +1. Name the target surface. +2. State the exact proposed record or action. +3. Ask for approval unless the user already gave explicit permission for that surface. + +If the input contains ambiguous dates, meetings, people, or commitments, do not create calendar or task records until the missing facts are grounded. + +## Output Contract + +Prefer concise receipts: + +```markdown +## What I caught +... + +## Meaning cards +- **Decision:** ... +- **Context:** ... + +## Possible next moves +- Safe: ... +- Needs approval: ... +``` + +Do not expose internal routing labels, private project names, hidden classifier output, or long audit trails unless the user asks for implementation details.