Prompt Optimizer

Optimize prompts with evals. Keep every instruction, example, and external context reference causal.

Load Only What You Need

Step 1: Capture Contract

Record before editing:

• task type: new, refine, port, or debug
• target model family and snapshot, if known
• prompt surface: system, developer, user, tool descriptions, examples, schemas
• layer owners: platform, deployer/persona, retrieved context, user payload
• objective and non-goals
• inputs, tools, and external files available
• required output shape
• success criteria and failure cases
• hard constraints: latency, verbosity, safety, budget, tool use, style

If success criteria or examples are missing, create a small eval set first. If the bottleneck is model choice, retrieval, tool schema, or missing evals, say so before rewriting.

Step 2: Inventory External Context

For repo or agent prompts, list stable context by exact path:

Rules:

• Reference stable files by repo-relative path instead of copying them.
• Paste only excerpts needed for the prompt or eval case.
• Mark whether a file is loaded, referenced, or out of scope.
• Avoid vague context pointers such as “read the docs”.

Step 3: Choose Model Strategy

Read references/model-family-notes.md.

• Known family: optimize for that family.
• Unknown family: write a portable base plus short adapter notes.
• Snapshot changes: rerun evals.
• Cross-family divergence: specialize only the failing layer.

Step 4: Shape Prompt

Read references/core-patterns.md.

• Put stable policy in system or developer.
• Put task-local facts, retrieved context, and variables in user-facing sections.
• Keep one owner per behavior rule.
• Use headings or tags only to separate content types.
• Put tool policy in prompt text; keep schemas in provider-native tools.
• Keep persona light unless it changes behavior.
• Use the shortest wording that preserves the constraint.
• Cut filler, repeated reminders, dead examples, and rationale that does not affect evals.

Step 5: Optimize

Read references/meta-optimization-loop.md for refinements.

1. Baseline the current prompt on the same eval slice.
2. Cluster failures by root cause.
3. Write concrete edit criticisms.
4. Generate two to four candidates:
   • minimal-diff repair
   • structure-first rewrite
   • examples-first or tool-rule variant
   • provider adapter when needed
5. Compare candidates on the same cases.
6. Keep a short optimization log.
7. Validate the winner on holdout cases.
8. Stop on plateau, oscillation, overfit, excessive cost, or non-prompt bottleneck.

Step 6: Return Package

Return:

1. Target
2. Success Criteria
3. External Context
4. Optimized Prompt
5. Adapter Notes
6. Eval Set
7. Optimization Log
8. Residual Risks

For existing prompts, include a concise diff-style note of the main behavioral changes.

Failure Modes

• editing before defining the eval target
• mixing policy, examples, and raw context without boundaries
• duplicating rules across layers
• putting durable policy in user payloads
• asking for chain-of-thought
• keeping contradictory legacy instructions
• overfitting to one or two examples
• retaining examples that no longer improve evals
• fixing tool-use failures only in prompt text when tool descriptions or schemas are weak
• adding markup that does not reduce ambiguity
• using persona as a substitute for behavior rules