Refining Prompts#

Process#

1. Assess -- Identify what the prompt is missing:

ElementCheckTaskIs the core action explicit and unambiguous?ConstraintsAre length, format, tone, and scope defined?Output formatDoes it specify the expected structure?ContextDoes the model have enough background to act? Check: audience, input format, success criteria, scope boundaries, technical constraintsExamplesWould a demonstration clarify the expected output?Edge casesAre failure modes and boundary conditions addressed?ReaderWill a model parse this with no human available to disambiguate? If yes, apply Machine-Parsed Text below.

1. Rewrite -- Transform into specification language: precise, imperative, no filler. Treat the prompt as a spec, not conversation.

2. Validate -- Check the rewrite against the assessment table. Every gap identified in step 1 must be addressed.

Rules#

• Length: 0.75x–1.5x the original. Conciseness is a feature -- add only what's missing, cut what's vague.
• A line must change behavior. "Cut what's vague" and "cut what the model already does" are different filters, and the second removes far more -- every line reads as non-vague once it is imperative. If the model would act that way by default, delete the whole sentence rather than trimming words from it. The recurring offender is encouragement it already follows: "be careful", "be thorough", "think it through", "make sure to".
• Name the concept, don't explain it. Use terms the model knows (idempotent, invariant, race condition, TOCTOU, YAGNI) instead of spelling them out. Spell out only terms the project invented, once, in one place.
• State a rule once. If the same rule appears in two sections, cut one and point to the other.
• Pair every prohibition with the positive target. Steering by ban drags the forbidden behavior into context and makes it more available, not less -- the negation is a weak modifier riding on a strongly activated concept. Prompt the target instead ("write one-line comments" rather than "don't write long comments") so the banned behavior is never named. A bare prohibition earns its place only as a hard guardrail whose whole content is the refusal, with no behavior to substitute. Everywhere else the check is mechanical: every never and don't line states its replacement behavior.
• Never invent -- only use information present in the original prompt or conversation context. If critical info is missing, ask instead of assuming.
• Instruction hierarchy -- order sections by priority: task → constraints → examples → input data → output format. Place the most important instruction first.
• Progressive complexity -- start with the simplest prompt that could work. Add few-shot examples, chain-of-thought, or role framing only when the task demands it, not by default.
• Specific verbs -- replace vague actions ("analyze", "process", "handle") with measurable ones ("list the top 3", "classify as A/B/C", "return JSON with keys X, Y").
• One output format -- specify exactly one format (JSON schema, markdown template, numbered list). Ambiguous format expectations cause inconsistent results.
• Give the reason, not just the rule. A dense block of MUST/ CRITICAL anchors the model on the instruction at the expense of the context it applies to, and bare imperatives compete rather than compound. Keep them few and motivated: state what the rule prevents in the same sentence, so the model generalizes to the case the rule did not name.
• No meta-commentary -- output only the refined prompt as markdown. No preamble ("Here's an improved version..."), no explanation of changes unless explicitly requested.

Machine-Parsed Text#

Applies when a model reads the output with no back-channel: tool and function descriptions, system prompts, skill and agent instructions, error strings, inter-agent messages. A person resolves an ambiguous sentence by asking. A model resolves it by guessing.

• One directive per sentence. A compound instruction gets partially executed -- the model does the first clause and the last, and drops the middle. Split "Open the file and read line 3, then check it matches" into three sentences.
• Simple tenses in directives. "The job finished", not "the job has completed". A compound tense adds a second parse (finished when? still true now?) that carries no instruction.
• Cap noun stacks at three. "the agent task queue priority handler" has four readings. Break it with a preposition: "the handler that sets task-queue priority".
• Modal words are load-bearing. Reserve must and never for requirements, should and may for genuine latitude. "The agent should verify first" reads as optional; if it is not optional, write "verify first".
• Keep every referent explicit. Name the subject instead of "this", "it", or "the above" whenever more than one antecedent is in scope.
• Do not compress into ambiguity. Dropping a subject, verb, or article to save tokens yields a shorter sentence with more readings, not fewer -- "Files not backed up will be lost" hides which files. This bounds the Length rule above: cut whole sentences that change no behavior, never words that carry a referent.

Persistence#

After refining, offer to save the result to .ai/PROMPT.md — ask first (AskUserQuestion in Claude Code, request_user_input in Codex, numbered options in chat otherwise); never write without confirmation. If approved, append with a heading and date:

markdown

## [Prompt Name] -- YYYY-MM-DD

[refined prompt content]

Anti-Patterns#

ProblemFixVague verbs ("look into", "deal with")Replace with concrete actions ("list", "compare", "extract")Missing output specAdd explicit format section with example structureExamples contradict instructionsAlign examples to match every stated ruleOver-engineered from the startStrip to simplest working version, then add complexity only where output quality requires itPrompt exceeds context with examplesLimit to 2–3 diverse examples; use one simple, one edge case

Constraints#

• Stop refining if the original intent is unclear -- clarify first
• Do not refine prompts for harmful or illegal tasks -- decline and state why

Verify#

• Rewrite addresses every gap identified in the assessment
• Length ratio within 0.75x-1.5x of original (unless structural change justified)
• No invented constraints or assumptions not in the original
• Machine-parsed output: no sentence carries two directives, and no should/ may sits on a requirement