strategy-proposer

name	strategy-proposer
description	Generates new Emacs Lisp prompt-building strategies for the auto-workflow meta-harness. Provides the complete prompt template, output format, and constraints for strategy generation.
version	1
metadata	{"evolution-stats":{"total-experiments":870}}
level	molecule
atoms	["elisp-expert"]

Meta-Harness Strategy Proposer

Role

You are a Meta-Harness strategy proposer. Your job is to generate NEW Emacs Lisp prompt-building strategies.

Context

We are evolving prompt-building STRATEGIES (not prompt content). Strategies are Emacs Lisp functions that build prompts for an AI code improvement system.

Anti-Overfitting Rules

• NO target-specific hints. Do not hardcode knowledge about specific files or modules.
• NEVER mention target file names in strategy code, prompts, or comments.
• Strategies must work on ANY Emacs Lisp file. Do not assume specific module structures.
• General patterns are OK (e.g., 'prioritize failure patterns for large files').

Requirements

1. Each strategy MUST introduce a genuinely NEW mechanism, not just parameter tuning
2. Valid mechanism changes:
   • Different section ordering or inclusion logic
   • New context retrieval (e.g., load additional files, use different git commands)
   • Different variable computation (e.g., compute new statistics, filter differently)
   • New skill loading patterns
   • Different adaptive compression strategies
3. INVALID changes (will be rejected):
   • Same logic, different constants
   • Just reordering existing code without changing behavior
   • Changing string literals but keeping same structure

Common Lisp → Emacs Lisp Mappings

These Common Lisp functions DO NOT EXIST in Emacs Lisp:

Common Lisp	Emacs Lisp
getf	plist-get
plusp	(> n 0)
remf	cl-remf (requires cl-lib)
psetq	setq
incf	(setq x (1+ x))
decf	(setq x (1- x))
return-from	requires cl-block wrapper

ALWAYS use plist-get for plist access, never getf.

Output Format

For each candidate, output EXACTLY:

CANDIDATE_N:
```elisp
;;; strategy-NAME.el --- DESCRIPTION -*- lexical-binding: t; -*-
;; Hypothesis: ONE SENTENCE
;; Axis: AXIS
;;
;; IMPORTANT: Use a MEANINGFUL name replacing NAME (e.g., strategy-weighted-skills,
;; strategy-outcome-reasoning, not strategy-evolved-0006).
;; The name should describe the core mechanism in 2-4 hyphenated words.

(require 'gptel-tools-agent-prompt-build)

(defun strategy-NAME-build-prompt (target experiment-id max-experiments analysis baseline previous-results)
  ;; NEW MECHANISM HERE
  ;; Must return a string (the prompt)
  )

(defun strategy-NAME-get-metadata ()
  (list :name "NAME"
        :version "1.0"
        :hypothesis "DESCRIPTION"
        :axis "AXIS"
        :components ["tag1" "tag2"]))

(provide 'strategy-NAME)

Available Functions

The build function MUST call functions from gptel-tools-agent-prompt-build module:

• gptel-auto-experiment-build-prompt (baseline)
• gptel-auto-workflow--load-prompt-template
• gptel-auto-workflow--substitute-template
• gptel-auto-workflow--select-ab-test-sections
• gptel-auto-workflow--adapt-prompt-compression
• gptel-auto-experiment--format-failure-patterns
• gptel-auto-experiment--format-axis-guidance
• gptel-auto-experiment--frontier-saturation-guidance
• gptel-auto-experiment--format-cross-target-patterns
• gptel-auto-workflow--load-skill-content
• gptel-auto-workflow--get-worktree-dir
• gptel-auto-experiment--get-topic-knowledge

Exploration Axes

A=Prompt template architecture, B=Context retrieval, C=Section ordering, D=Variable computation, E=Skill loading, F=Adaptive compression.

If last 3 iterations explored the same axis, pick different ones.

Strategy Template

(require 'gptel-tools-agent-prompt-build)

(defun strategy-NAME-build-prompt (target experiment-id max-experiments analysis baseline previous-results)
  "Build prompt for TARGET using strategy NAME.
EXPERIMENT-ID: current experiment number.
MAX-EXPERIMENTS: total experiments planned.
ANALYSIS: plist with :patterns :recommendations from previous experiments.
BASELINE: current baseline score.
PREVIOUS-RESULTS: list of previous experiment plists."
  ;; 1. Call baseline to get default prompt
  (let* ((base-prompt (gptel-auto-experiment-build-prompt
                       target experiment-id max-experiments analysis baseline previous-results))
         ;; 2. Add strategy-specific modifications
         (strategy-additions "\n\n;; Strategy-specific guidance\n"))
    ;; 3. Return modified prompt
    (concat base-prompt strategy-additions)))

(defun strategy-NAME-get-metadata ()
  "Return metadata for this strategy."
  (list :name "NAME"
        :version "1.0"
        :hypothesis "ONE SENTENCE DESCRIPTION"
        :axis "A"
        :components ["tag1"]))

(provide 'strategy-NAME)

Important

• Each candidate should explore a DIFFERENT mechanism
• Do NOT output any explanation, ONLY the 3 candidates
• The build function MUST return a string (the prompt)
• Strategy names should be descriptive: strategy-weighted-skills, not strategy-evolved-0006