KNOWLEDGE.BASE
Living knowledge asset. Raw sources are captured immutably, thinking is dated and append-only, understanding is merged in place, writing ships. Three agent pipelines keep it honest.
WIKI // CURATED
Synthesized current understanding — topics, concepts, entities, summaries. Every claim traces back to a source.
NOTES // THINKING
Dated thinking notes. Append-only: a changed mind is a new note, not a rewrite.
- Fact Grounding for AI Characters: Classify Hallucinations by Supply Chain, Not as "the Model Being Bad"Practice from taming time/schedule/memory hallucinations in a character product: turn-level fact capsules, send-time grounding with candidate narrowing, a closed exit for semantic recall, a side-effect-free prompt A/B sandbox.
- Parallel Agent Delivery across Worktrees: the Throughput Bottleneck Is Isolation, Not the ModelA retrospective on the delivery scaffolding behind 400+ merged PRs in two-plus months: worktree-as-session, port slot allocation, requirement mirroring and acceptance tracking, bounded review convergence, CI caching and fast/slow tiering.
- Small-Team Multi-Environment Secret Management: SOPS + age + Doc Permissions as CredentialsA retrospective on landing secret management in a multi-service project: ciphertext in the repo, per-key encryption for reviewability, doc permissions as credentials, keyless CI, two-layer rotation.
- LLM Harness Writing StudyThe Prompt → Context → Harness three-ring spiral: each ring turns the previous one into infrastructure.
- Multi-Agent Veto BoundaryIn multi-agent collaboration, every agent needs an explicit veto boundary — when it may refuse the orchestrator.
- Natural Language Tool StateTranslating exit codes into natural language is not redundancy — natural language is the state representation a model consumes most easily.
- Tool Call Role Design: Claude Code vs Codextool_result living in role:user is not a bug — the user role carries environment observations, not a human.
ARTICLES // OUTPUT
Long-form writing. Local markdown is ground truth; every platform gets a derivative.
- I Fed 653 Agent Sessions to SpotlightThe direction left after four ideas got killed. Feed the session data to the entry point the system already ships — then end up writing a distributed system on a single machine.
- How One Video Agent Iteration Grew into a Harness MethodOut of tuning one video-style agent grew shared context, independent review, boundary knowledge and regression protection.