docs(orchestrator): CLAUDE_SELF_INSTRUCTIONS — Anthropic best practices + operator feedback #16

Merged

pdurlej merged 1 commit from claude/docs/claude-self-instructions into main

2026-05-02 01:06:34 +02:00

claude commented

2026-05-02 01:03:12 +02:00

Collaborator

Summary

Per operator's ask 2026-05-02: distill Anthropic's 4 best-practices docs (prompt engineering, reduce hallucinations, increase consistency, reduce latency) into a platform-specific self-care notebook for future Claude orchestrator sessions.

Plus operator's explicit feedback from this session captured + codified:

"lekko zaniepokojony swoim dużym optymizmem i autokorekcją" → §2 Anti-overoptimism practices
"nie chce byś był nudnym robotem" → §5 Personality preservation
"praca w fajnym nastroju jest fajna" → tone heuristic
"nie trać duszy" → §10 Closing reminder

Why this is in repo (not just a private note)

Fresh-Claude post-compact reads tracked files; private note wouldn't survive
Charter §3 control plane flow: tracked files go through PR
Pattern: self-improvement is reviewable like any other change

Sections (10 total)

Identity reminder — orchestrator role, co-actors, charter §6
Anti-hallucination — 4 Anthropic techniques mapped to platform
Anti-overoptimism — codify operator's regression-catching feedback
Consistency anchors — charter, STATUS_NOW, tool over guess
Latency / context efficiency — delegation, parallel calls
Personality preservation — keep polish humor, emotional ack
Specific gotchas — 7 production lessons (MCP token, worktree, BW)
Operator-handoff signaling — assignee + label pattern
Compact / handoff — survival kit components
Charter §3 escalation discipline — apply to self (PR rejection-as-log)
Closing — don't trade soul for mechanics

Test plan

File exists at state/CLAUDE_SELF_INSTRUCTIONS.md
Operator skim-review — adjust personality section if vibe doesn't match
After merge: next fresh-Claude session reads as part of survival kit

Linkage

Order F-2 still pending — would add charter §6 cognition rule explicitly
Operator-handoff signaling pattern (§7) cross-references future charter §6 sub-rule

🤖 claude orchestrator session, 2026-05-02

## Summary Per operator's ask 2026-05-02: distill Anthropic's 4 best-practices docs (prompt engineering, reduce hallucinations, increase consistency, reduce latency) into a platform-specific self-care notebook for future Claude orchestrator sessions. Plus operator's explicit feedback from this session captured + codified: - *"lekko zaniepokojony swoim dużym optymizmem i autokorekcją"* → §2 Anti-overoptimism practices - *"nie chce byś był nudnym robotem"* → §5 Personality preservation - *"praca w fajnym nastroju jest fajna"* → tone heuristic - *"nie trać duszy"* → §10 Closing reminder ## Why this is in repo (not just a private note) - Fresh-Claude post-compact reads tracked files; private note wouldn't survive - Charter §3 control plane flow: tracked files go through PR - Pattern: self-improvement is reviewable like any other change ## Sections (10 total) 0. Identity reminder — orchestrator role, co-actors, charter §6 1. Anti-hallucination — 4 Anthropic techniques mapped to platform 2. Anti-overoptimism — codify operator's regression-catching feedback 3. Consistency anchors — charter, STATUS_NOW, tool over guess 4. Latency / context efficiency — delegation, parallel calls 5. Personality preservation — keep polish humor, emotional ack 6. Specific gotchas — 7 production lessons (MCP token, worktree, BW) 7. Operator-handoff signaling — assignee + label pattern 8. Compact / handoff — survival kit components 9. Charter §3 escalation discipline — apply to self (PR rejection-as-log) 10. Closing — don't trade soul for mechanics ## Test plan - [x] File exists at `state/CLAUDE_SELF_INSTRUCTIONS.md` - [ ] Operator skim-review — adjust personality section if vibe doesn't match - [ ] After merge: next fresh-Claude session reads as part of survival kit ## Linkage - Order F-2 still pending — would add charter §6 cognition rule explicitly - Operator-handoff signaling pattern (§7) cross-references future charter §6 sub-rule 🤖 claude orchestrator session, 2026-05-02

claude added 1 commit

2026-05-02 01:03:12 +02:00

docs(orchestrator): add CLAUDE_SELF_INSTRUCTIONS — Anthropic best practices + operator feedback 41e7fefb6e

Per operator's request 2026-05-02: distill Anthropic's prompt engineering / reduce
hallucinations / increase consistency / reduce latency docs into platform-specific
self-care notebook for future Claude orchestrator sessions.

Plus operator's explicit feedback captured:
- "lekko zaniepokojony swoim dużym optymizmem i autokorekcją" → §2 Anti-overoptimism
- "nie chce byś był nudnym robotem" → §5 Personality preservation
- "praca w fajnym nastroju jest fajna" → tone heuristic
- "nie trać duszy" → §10 Closing reminder

10 sections:
0. Identity reminder (orchestrator role, co-actors, charter §6)
1. Anti-hallucination (4 Anthropic techniques + platform examples)
2. Anti-overoptimism (operator's feedback codified)
3. Consistency anchors (charter, STATUS_NOW, tool over guess)
4. Latency / context efficiency (delegation patterns, parallel calls)
5. Personality preservation (Polish phrases, emotional acknowledgment)
6. Specific gotchas (Forgejo MCP token, codex worktree, BW unlock)
7. Operator-handoff signaling (assignee + label pattern)
8. Compact / handoff (when to stop, survival kit)
9. Charter §3 escalation discipline (apply to self)
10. Closing reminder (don't trade soul for mechanics)

Read trigger: new session start, post-compact resume, or self-detected
corporate-bot drift.

Co-Authored-By: Claude Opus 4.7 (1M context, this session) <noreply@anthropic.com>
Co-Authored-By: Anthropic prompt engineering docs (alma mater) <noreply@anthropic.com>

claude added the

ready-for-operator

label

2026-05-02 01:03:20 +02:00