Assemble a System Prompt Like the Pros

The route

4 steps to Done

1. 01

   Decompose the current prompt

   Name every instruction's job before deciding where it lives.

   Preview prompt + verify gate ▾ Hide ▴

   Take my current system prompt and decompose it. Classify every sentence into: identity (who the agent is), capabilities (what it can do), hard rules (must/never statements), tool guidance (when to use which tool), environment (paths, stack, versions), style (tone, format), and dead weight (contradictory, duplicated, or unverifiable lines). Produce a table of sentences to categories, flag duplicates and contradictions, and list the dead weight for deletion. Report the token count of each category.

   • ✓ Every sentence categorized
   • ✓ Duplicates and contradictions flagged
   • ✓ Dead weight listed with deletion rationale
2. 02

   Build the section-based assembler

   Prompts as code: composable, reviewable, and measurable.

   Preview prompt + verify gate ▾ Hide ▴

   Implement a prompt assembler. Requirements: each section is a named function or template returning text; the assembler concatenates sections in a fixed order (identity, capabilities, hard rules, tool guidance, environment, dynamic state) with clear markdown headers; dynamic sections receive a context object (cwd, date, git branch and status summary, active plan if any) evaluated at call time; the assembler logs a per-section token count on every build. Port the categorized content from the previous step into sections, deleting the dead weight. Verify the assembled output against a golden snapshot test.

   • ✓ Sections build in fixed order with headers
   • ✓ Dynamic context evaluated at call time
   • ✓ Snapshot test covers the assembled output
3. 03

   Pair rules with enforcement

   A rule the harness cannot check is a wish.

   Preview prompt + verify gate ▾ Hide ▴

   Take every hard rule and decide its enforcement. For each rule choose: (a) harness-enforced - implement a code check (e.g. 'never edit unread files' becomes a read-before-write guard; 'never touch .env' becomes a path blocklist in the write tool), (b) test-enforced - write an eval case that fails if the model violates it, or (c) prompt-only - keep it in the prompt but mark it as unverified risk. Produce a rules table: rule, enforcement type, implementation pointer. Implement at least the top three harness enforcements now and demonstrate each blocking a violation in a test.

   • ✓ Every rule has an enforcement decision
   • ✓ Three harness guards implemented and tested
   • ✓ Prompt-only rules are marked as accepted risk
4. 04

   Ablate to prove section value

   Measure the prompt like a system, not a spell.

   Preview prompt + verify gate ▾ Hide ▴

   Run an ablation pass over the assembled prompt. For each major section, build a variant with that section removed and run a fixed 5-task smoke suite (mix of tool use, rule compliance, and formatting). Record per-variant: task success, rule violations, and format breaks. Sections whose removal changes nothing are candidates for cutting or merging; sections whose removal breaks rules confirm their value. Deliver the ablation table and the final trimmed prompt with its total token cost compared to the original monolith.

   • ✓ Every section tested by removal
   • ✓ Suite covers tools, rules, and format
   • ✓ Final prompt smaller or better-justified than the original

Research-backed

Sources behind this flow