メニュー
Use when starting work in the FLOSSI0ULLK workspace, re-orienting after context loss, or deciding which canonical docs and corpora to load first before deeper research or code changes.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | flossi0ullk-orient |
| description | Use when starting work in the FLOSSI0ULLK workspace, re-orienting after context loss, or deciding which canonical docs and corpora to load first before deeper research or code changes. |
| version | 0.2.0 |
Regain orientation on minimum tokens, maximum verifiability. Read nothing before you know what is present and fresh.
| Tier | Target | When |
|---|---|---|
| T0 | ≤ 500 | Probe only. Always. |
| T1 | ≤ 3 000 | Read CONTEXT_L0.md + canonical pointers. |
| T2 | ≤ 12 000 | Escalate to CONTEXT_L1.md, INDEX.md, or a single routed corpus root. |
| T3 | open | Only after T1+T2 proved insufficient for the specific task. |
Stop at the lowest tier that answers the task. Do not pre-fetch.
Run the probe to get a deterministic orientation packet. No mutation, no network.
python FLOSS/scripts/orient_probe.py --query "<task in one line>"
The packet reports: presence, mtime, size, and sha256 prefix for each canonical file; .agent-surface/events/ queue depth; whether context_router.py is runnable; and a routed-corpus shortlist (if the router is present).
If the probe script itself is missing, fall back to the Manual probe ladder at the bottom of this file.
In order, read only what the probe marked present and fresh:
.agent-surface/context/CONTEXT_L0.md — cheap re-orientation.INDEX.md — authoritative map of the repo.FLOSS/CLAUDE.md — agent-facing operating notes.FLOSS/FLOSSI0ULLK_Master_Metaprompt_v1_3_1_Kernel.md — only if the task touches governing principles.python FLOSS/scripts/context_router.py "<query>" --format markdown --limit 4
Open only the top-routed corpus roots. If the router is absent or errors, fall back to references/entry-points.md in this skill.
If L0 + routed roots did not resolve the task, then — and only then:
.agent-surface/context/CONTEXT_L1.mdFLOSS/docs/adr/ for governance/architecture tasks.For tasks touching intake, filewatch, consolidation, or cross-agent update flow, load (in order of increasing cost):
FLOSS/docs/architecture/CONTEXT_DAEMON_ARCHITECTURE.mdFLOSS/docs/superpowers/plans/2026-04-19-filewatch-metaharness.mdFLOSS/scripts/watch_intake.pyFLOSS/scripts/process_intake_events.pyls -R, never cat a directory-sized glob. If you find yourself wanting to, re-run the router._reference/ is cold storage. Treat it as a published research library, not live canon. Read only when the router returns it and the task requires it..agent-surface/events/ is runtime queue state. Not canon. Do not read event contents unless the task is the event-processing pipeline itself._reference/.| Missing artifact | Fallback |
|---|---|
orient_probe.py | Use Manual probe ladder below. Then open a P1 issue to restore probe. |
CONTEXT_L0.md | Read INDEX.md + FLOSS/CLAUDE.md only. |
INDEX.md | Use references/entry-points.md in this skill as the map. |
context_router.py | Use references/entry-points.md; flag the regression. |
.agent-surface/ absent | Treat as cold-start. Do not infer runtime state. |
for f in \
.agent-surface/context/CONTEXT_L0.md \
.agent-surface/context/CONTEXT_L1.md \
INDEX.md \
FLOSS/CLAUDE.md \
FLOSS/FLOSSI0ULLK_Master_Metaprompt_v1_3_1_Kernel.md \
FLOSS/scripts/context_router.py \
FLOSS/scripts/orient_probe.py; do
if [ -f "$f" ]; then
printf '%s\t%s\t%s\n' "present" "$(stat -c '%y %s' "$f" 2>/dev/null || stat -f '%Sm %z' "$f")" "$f"
else
printf 'MISSING\t-\t%s\n' "$f"
fi
done
ls -1 .agent-surface/events/ 2>/dev/null | wc -l # runtime queue depth
Before closing the task, answer:
_reference/ or a broad directory? If yes, justify..agent-surface/events/ or source-chain state? If yes, intentionally?references/entry-points.md — canonical file map with tier tags and one-line purpose.references/entry-points.md and scripts/orient_probe.py.Self-learning capture system. Hooks detect correction patterns ("no, use X", "actually...", "use X not Y") and queue them; user runs /reflect to review and apply queued learnings to CLAUDE.md files. Use when the user mentions remembering something for future sessions, after a feature lands, or when context is about to compact with queued items.
Save a tight cross-session handoff note to .remember/remember.md so the next session — same agent or different — picks up cleanly. Use at the end of a working session, before context compaction, or whenever the user says to remember state for next time.
Generate an explorable HTML report of Claude Code session usage (tokens, cache, subagents, skills, expensive prompts) from ~/.claude/projects transcripts. Use at end-of-session, end-of-week, or when investigating token spend / cache-hit anomalies / which subagent or skill is eating budget.
Use when changing or extending the shared FLOSSI0ULLK agent surface: MCP servers, context manifests, skill manifests, materializers, or agent-native projections for Codex, Gemini, Claude-adjacent tooling, or OpenCode.
FLOSSI0ULLK Inline Reasoning Ensemble: use the Router for cheap mode classification (pass_through, single_strong, ensemble), and the Synthesizer for full ensemble deliberation on architectural decisions, multi-file refactors, ADR-class moves, or any substantive decision where being wrong has rollback cost. Also use when the user explicitly asks for debate, a second opinion, or multi-model/lens analysis.
Use when the user asks for recent, current, latest, or open-access research on distributed intelligence, agent-centric architectures, decentralized multi-agent systems, federated learning, distributed vector databases, agent interoperability protocols, or FLOSSI0ULLK-adjacent open knowledge commons. Always use this skill for "recent open distributed intelligence research" prompts, even if the user only names a seed report or asks for a quick update.