Skip to content

features.md

Latest commit

 

History

History
63 lines (45 loc) · 4.89 KB

features.md

File metadata and controls

63 lines (45 loc) · 4.89 KB
Raw

Effective Claude Code: Presentation Features

10 candidates extracted from 22K-line transcript of building Angelic Intelligence over 5 days with Claude Code. Each tested with FOR/AGAINST evidence, self-review, and external references.

✅ Definitely In

# Feature Confidence One-liner Key evidence Reference
-1 Agent Modes -- Audience needs to understand Claude Code vs Subagents vs Agent Teams before patterns make sense Most patterns in this deck use subagents; audience needs to know what's available today vs experimental agent_modes.md
0 METR: AI Can Slow You Down -- METR RCT: experienced devs 19% slower with AI, while believing 20% faster 39-point perception gap; followup study (57 devs) still no speedup; sets the stage -- we're not cheerleading metr.md
1 It's All Meta-Programming 9/10 Your job shifts from writing code to designing agent workflows 9-agent workflow designed in natural language before any code; evolved to 12-agent 5-phase pipeline; "dude have an agent do the work don't do it yourself" evidence_meta1.md

🤔 Maybe -- Need to Decide

# Feature Confidence One-liner Key evidence Biggest caveat Reference
2 Frozen Specs Prevent Drift 7/10 Git-tagged immutable specs as single source of truth for parallel agents 16 stories frozen, 4 agents audited ~30 files, only 3 minor fixes needed Prevents spec-to-spec drift but NOT spec-to-code drift (5/6 modules violated SKILL.md) evidence_arch1.md
3 Build -> Test -> Audit -> Select 7/10 Parallel builders, sequential test gate, parallel auditors, then select the best Used 7+ times; cross-auditors caught 19 real issues across 3 build tracks Human still intervened constantly; rate limits killed all 3 builders once evidence_arch3.md
4 Selenium QA as Parallel Gate 7/10 E2E tests as binary pass/fail gate on every agent's output Caught avatar mismatch, stale refs, React state bug after multi-agent refactoring Most critical bugs were caught by pytest/manual, not Selenium; Chromedriver crash never resolved evidence_qa1.md
5 Dedup Poisons Tests 7/10 Caching failed API responses creates a hidden test-poisoning trap 9 debugging steps before finding cached failure; hit the same trap twice; blocked flagship feature More of a gotcha than a pattern; could fold into "failure modes" slide evidence_qa2.md
6 Three-Agent Fact-Checking 7/10 Researcher -> Auditor -> Summarizer catches errors before sharing Auditor found 4/27 claims had errors (15% error rate caught) Only used twice; 3x latency cost; errors were citation-level not substantive evidence_agent1.md

❌ Out -- Not Strong Enough

# Feature Confidence One-liner Why it's out
7 Quorum Rule for Frozen Docs 5/10 Require 3+ agent agreement before modifying frozen docs Never triggered. No agent ever proposed a change. Can't tell if it deterred anything or just wasn't needed
8 Shared Mutable State 4/10 Agents sharing Docker/DB/browser state causes clobbering Overstated -- only affected Chrome and DB mutations; vast majority of parallel work was read-only and fine
9 Dual Hard-Gate Layers 3.5/10 Hard gates in both scoring and optimizer was redundant Actually the architecture was defensible (defense in depth). Real problem was implementation bugs, not the design

Cross-Cutting Patterns

These recur across multiple findings and may deserve their own slide:

Pattern Where it appeared Confidence
Delegate, don't do meta-1, arch-3, agent-1 High
Freeze inputs, version them arch-1, agent-3 High
Cross-audit > self-review arch-3, agent-1, qa-3 High
Structured output for delegation qa-1 High
Isolate agent contexts arch-1, agent-2 Medium
Serial for mutations, parallel for reads agent-2, arch-3 Medium
Same-model auditors may be lenient agent-1, agent-3, qa-3 Noted

What Failed or Was Overrated

Worth including as an honesty slide:

  • Full autonomy doesn't work: Human intervention needed constantly for quality ("THERE IS NO WAY THEY DID A GREAT JOB"), direction, and resource management (rate limits)
  • Specs alone don't enforce themselves: 5/6 modules violated the frozen spec -- need active audit agents or CI checks
  • Quorum rule: Untested safety net that never activated
  • Shared-state problems: Real but narrow (only mutations, not the common case)