agentleFS
m2ai-mcp-serversm2ai-mcp-servers/claude-skills/skills/bitter-lesson-scorecard/SKILL.md

bitter-lesson-scorecard

Skill0 starsChanged 7 months agoMITView on GitHub

Score an agent system design against the Bitter Lesson principle — how much "how" is encoded vs "what", how much bets on model improvement vs locks in current limitations. Flags procedural lock-in, hardcoded orchestration, and domain hacks. Produces a simplification roadmap. Use when designing new agent systems, reviewing agent architecture, or deciding what to simplify. Trigger on "bitter lesson", "score architecture", "agent complexity audit", "simplification roadmap", "how vs what ratio", "are we fighting the model".

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Ask the user to provide ONE of:

    line 12
  2. Distinctive:

    If given a codebase path, look for: orchestration files, state machines, routing logic, prompt templates, tool definitions, multi-agent …

    line 17
  3. Distinctive:

    Every component in the system must be classified (nothing skipped)

    line 98

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Bitter Lesson Scorecard for Agent Architectures

Score an agent system against the Bitter Lesson: computation and learning beat hand-engineering. Systems that encode "how" instead of "what" get worse as models improve.

Phase 1: Intake

Ask the user to provide ONE of:

  • An agent system description or architecture doc
  • A codebase path (will scan for orchestration patterns)
  • A CLAUDE.md or system prompt describing an agent workflow

If given a codebase path, look for: orchestration files, state machines, routing logic, prompt templates, tool definitions, multi-agent coordination code.

Phase 2: Component Inventory

Map every architectural component into one of these categories:

CategoryBitter Lesson AlignmentExamples
Outcome SpecALIGNED — says "what"Goal definitions, success criteria, quality thresholds
Tool InterfaceALIGNED — extends capabilityAPI wrappers, file access, search tools
Hard ConstraintNEUTRAL — business necessityAuth, rate limits, compliance rules, safety gates
Procedural OrchestrationMISALIGNED — encodes "how"State machines, fixed step sequences, hardcoded agent routing
Model CompensationMISALIGNED — bets against improvementChunking strategies, re-ranking, format enforcement, retry heuristics
Domain HackMISALIGNED — freezes current knowledgeHardcoded few-shot examples, domain-specific parsing, manual entity extraction

Phase 3: Scoring

Calculate three scores (0-100):

Alignment Score

(outcome_specs + tool_interfaces + hard_constraints) / total_components * 100

Higher = more aligned with the Bitter Lesson.

Lock-In Score

(procedural_orchestration + model_compensation + domain_hacks) / total_components * 100

Higher = more locked into current model limitations. This is the number to REDUCE.

Improvement Leverage

Estimate: if the underlying model improves 2x in capability, what percentage of the system becomes unnecessary?

High leverage = the system will naturally simplify with better models. Low leverage = the system fights improvement.

Phase 4: Scorecard Output

## Bitter Lesson Scorecard: [System Name]

| Metric | Score | Rating |
|--------|-------|--------|
| Alignment | XX/100 | [STRONG/MODERATE/WEAK] |
| Lock-In | XX/100 | [LOW/MODERATE/HIGH] |
| Improvement Leverage | XX% | [HIGH/MODERATE/LOW] |

## Component Breakdown
| Category | Count | % of System |
|----------|-------|-------------|
| Outcome Specs | N | X% |
| Tool Interfaces | N | X% |
| Hard Constraints | N | X% |
| Procedural Orchestration | N | X% |
| Model Compensation | N | X% |
| Domain Hacks | N | X% |

## Top Bitter Lesson Violations
1. [Component] — [Why it bets against improvement]
2. ...
3. ...

## Simplification Roadmap
### Quick wins (delete now, test)
- ...

### Medium-term (replace orchestration with outcome specs)
- ...

### Strategic (requires model capability validation)
- ...

Phase 5: Recommendations

For each MISALIGNED component, suggest:

  1. What it would look like as an outcome spec instead of a procedure
  2. What model capability would need to exist to delete it
  3. Whether that capability likely already exists in current models

Verification

  • Every component in the system must be classified (nothing skipped)
  • Scores must be mathematically consistent with component counts
  • Simplification roadmap must have at least one item in each tier
  • Recommendations must be specific to the system, not generic advice

Source

Nate's Newsletter (2026-04-01): The Bitter Lesson applied as a practical audit tool for agent architectures — scoring systems on "how" vs "what" encoding and producing simplification roadmaps.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.