The standout
A knowledge layer over the tooling
We ingest transcripts from the best practitioner videos, extract the claims that matter, and score each one for novelty against the rest of the corpus. What stays is the signal; what repeats gets deduped.
1
Ingest
Curated channels and search feeds per tool. Transcripts come from captions or mlx-whisper as a fallback.
2
Extract and score
An LLM pulls structured claims; each is embedded and scored for novelty against the existing corpus, so redundant takes sink.
3
Connect
Claims become a graph that links tools to the concepts and tradeoffs that define them, rendered right on each tool page.
What the novelty score means. A claim scoring 0.9 adds
signal almost no other source covers; a claim near 0.2 restates what is
already widely said. We surface the high-novelty claims first.
Insights
Curated, ranked by novelty
Persistent memory is the biggest multiplier: a directory that logs corrections, people, and projects, read at session start, means every fix compounds across sessions.
100
After a model update that gets nerfed, Claude Code can burn 2-3x the tokens for the same job; budget for token economics, not just price per token.
96
Treat the agent like a high-leverage junior engineer: give it tight specs, review every diff, and keep the feedback loop fast.
96
Plan versus Act modes are the safety model: Plan proposes a diff, Act executes it, and approval gates stop destructive edits.
96
At-mention grounding on files, docs, and web is what keeps Cursor from hallucinating APIs; use it deliberately on every non-trivial task.
94
Sub-agents (the Task tool) keep the main context clean; delegate search-heavy work so it does not fill the window.
93
OpenCode's provider-agnostic, local-first design is why teams standardize on it to avoid lock-in to a single model vendor.
93
Aider's repo-map is what lets it scale to large codebases without dumping every file into context.
92
Because Aider commits every change, your git history becomes an audit log and rollback is a single command.
92
Langfuse is the most-used open-source LLM observability tool; self-hosting keeps sensitive traces off vendor servers.
90
For pure regression testing rather than tracing, pair Langfuse with a dedicated eval tool; tracing alone will not catch quality drift.
90
Concept map
Tools grouped by what they cover
Context management
MCP support
Git-native workflow
Sources
Where the knowledge comes from
| Source | Channel | Kind | Status |
|---|---|---|---|
| Practitioner notes, 2,000 hours with Claude Code | Curated practitioner notes | article | Verified |
| Claude Code documentation | Anthropic | docs | Verified |
| Anthropic YouTube channel (transcript pending) | Anthropic | video | Transcript pending |
| Aider documentation | Paul Gauthier | docs | Verified |
| Cline documentation | Cline | docs | Verified |
| OpenCode documentation | SST | docs | Verified |
| Langfuse documentation | Langfuse | docs | Verified |
| Cursor documentation | Cursor | docs | Verified |