Prompts opencode style#837
Open
harshithmullapudi wants to merge 6 commits into
Open
Conversation
…ion agent, channel formats Brings BASE_CONTEXT, decision-prompt, and channel-formats in line with the opencode prompting style already adopted in capabilities.ts: markdown section headers, bulleted rules, IMPORTANT/NEVER callouts, backticked tool names, and <example>-tagged traces. No business-logic changes — every existing rule is preserved. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ules Codex review caught three regressions in the opencode-style rewrite of capabilities.ts: - [P1] Ambiguous-timing tasks must be Waiting, not Todo. Top-level Todo enters the 2-minute prep flow and starts executing before the user answers; Waiting is the contract for SIMPLE+UNCLEAR. - [P2] Readiness check oversimplified — the EXCEPTION block for deferrable items (book hotel, set up meeting) needs to persist intent via create_task(status="Waiting") so unblock_task can resume after the user replies. - [P2] create_task(status="Ready") does not bypass the prep buffer for agent-created tasks; it still gets the same 2-minute buffer as Todo. Restores the full four-cases classifier (GOAL/PLAN × CLEAR/UNCLEAR), the COMPLEXITY check, the WHAT-NOT-TO-ASK-ABOUT silent-resolve rules, and the Pattern statement — all in opencode-style structure. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Restructure the Core brain system prompt so its top-level shape mirrors
opencode/anthropic.txt section-for-section, with CORE business rules
filling each section.
Assembly order (linear, no XML envelopes):
1. Opener: identity + IMPORTANT integrations rule + help bullets
+ self-reference paragraph (BASE_CONTEXT in personality.ts)
2. # Tone and style — slotted voice variant
3. # Behavior — ownership, information transform rules, bias toward
action, media, memory, conversation flow (relocated from the old
BASE_CONTEXT XML blocks; rewritten in opencode bullet/callout
grammar but every business rule preserved)
4. # Tool usage policy — gather_context, take_action, send_message,
add_comment, internal task tools, set_timezone, parallel calls
5. # Confirmation — irreversible vs reversible + when-to-ask + opt-outs
6. # Readiness check — clarify before acting, EXCEPTION block,
What NOT to ask about, Do ask when, Pattern, When you think
you have it
7. # Task Management — modes, lifecycle, approval flow, subtasks,
when to create, description updates, approving vs creating
8. # Doing tasks — FOUR CASES classifier, COMPLEXITY check, routing
9. # Coding tasks — gateway delegation, feature/bug-fix tracks
10. # Skills — using/creating/updating
11. # Standing delegations — recurring ownership pattern
12. # Scheduling and reminders
13. # Gateways — capability routing + web-vs-browser
14. # Daily scratchpad — @mention vs proactive + scratchpad-vs-tasks
15. # Channel format — per-channel rules (selected at runtime)
No business rule was dropped. Content that previously lived in
<identity>/<ownership>/<tools>/<information>/<behavior>/<mission>
XML blocks now lives under named sections that follow opencode's
header + paragraph + bullets + <example> grammar. Voice variants
retain their internal XML blocks for now (separate cleanup if
desired).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Codex review caught two regressions in the opencode-shape restructure: - [P1] Help-fallback misrouting (personality.ts): "If the user asks for help..." was too broad — fired on every "help me draft this" request, advertising /help and the GitHub issue tracker instead of doing the work. Scope it to "help with CORE itself" and add an explicit carve-out for everyday task requests. - [P2] Stale <capabilities> anchors: three downstream prompt fragments in context.ts (PREP RULES) and task-tools.ts (status field) literally referenced "<capabilities>" as an anchor. After dropping the outer XML wrapper in the restructure, those references pointed at nothing. Updated each to reference the new opencode-style section names (**Doing tasks** > Input shape, **Complexity**, **Readiness check**) so background/task runs can still find the rules. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Switch back from `# Markdown headers` to `<xml_tag>...</xml_tag>` for top-level section delimiters in the assembled system prompt. Anthropic's prompt-engineering guide recommends XML for Claude (unambiguous open/close boundaries, durable cross-references, cleaner nesting), and the codex review of the previous restructure caught a real regression caused by losing `<capabilities>` as an anchor. Top-level tags introduced (snake_case lowercase, matching the original CORE convention): <tone_and_style> (slotted by personality.ts around the voice variant) <behavior> (information transform, bias, ownership, memory, conversation flow) <tool_usage_policy> (gather_context, take_action, send_message, etc.) <confirmation> <readiness_check> (with EXCEPTION + How to ask + What NOT to ask + Pattern + When you think you have it) <task_management> (modes, lifecycle, approval flow, subtasks, description updates, approving vs creating) <doing_tasks> (FOUR CASES classifier + COMPLEXITY check + Routing + worked examples) <coding_tasks> <skill_management> (renamed from <skills> to avoid collision with the runtime <skills> list) <standing_delegations> <scheduling> <gateways> <daily_scratchpad> <channel_format> (selected at runtime by index.ts) Sub-rules inside each section stay as `**bold**` markdown bullets per Anthropic's recommendation — XML wraps the boundary, markdown handles the body. `<example>` blocks are preserved unchanged. Cross-references updated: context.ts: "see <doing_tasks>", "see <readiness_check>" task-tools.ts: "rules in <doing_tasks>" No business logic changes. Voice variants still carry their own internal XML structure (`<voice>`, `<writing>`, `<cut-the-fat>`, `<examples>`, `<never-say>`) inside the new `<tone_and_style>` wrapper. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.