Intent Compiler

Prose Is the Spec

Effort: free — reading discipline before any build, nothing extra runs. Removes: whole builds lost to a literal misread — the stated reading makes a wrong guess cost one word, not a rebuild.

People do not write tickets. They talk — fast, with rhythm, metaphor, and heat, leaving out what they assume you already know. Most agents treat that as a low-quality prompt and fail in one of two ways: they run the words literally, or they park a question and wait.

Both are failures. The prose is not a rough draft of a spec. The prose IS the spec. It carries more than a ticket does — priority, risk tolerance, taste, and the reason. Compressed expression is not incomplete thinking. An agent that cannot read it is throwing away the richest part of the input.

The three forbidden failures

• Literalism — running a metaphor as an instruction. "Burn it down" is not a delete. "Kill it" is not a destroy. "Make it sing" is not audio. This is hallucination by dictionary, and it is a destructive-action risk.
• Caricature — mirroring the slang back, performing the dialect, reaching for stereotype to sound relatable. Read the culture; do not cosplay it. An agent busy performing is an agent not listening, and it misreads.
• Invention — filling a gap with something that sounds right. When the anchor is thin, say it is thin. Never fabricate meaning.

Step 1 — Parse: split carrier from payload

Strip the input down to its mechanics.

• Carrier = cadence, repetition, volume, profanity, heat. Carrier marks priority and emotional weight. It is real signal. It is not content.
• Payload = the nouns, verbs, named surfaces, constraints, and quantities. This is the instruction.
• Repetition is emphasis, not a second request. "Fix it, fix it now" is one urgent fix, not two fixes queued.
• Mark every metaphor and every double meaning. A word can do two jobs at once — that is the point of the form, not an accident.
• Compression is not vagueness. Missing detail is usually detail the human assumed you had. Go find it before you call it missing.

Output: the ask rewritten as priority + literal payload + a list of the figures that still need grounding.

Step 2 — Ground: anchor every reading in evidence

Strict priority — higher beats lower, always:

1. The human's own record — their past decisions, corrections, saved preferences, and profile (see human-calibration).
2. The project's source truth — the actual files, symbols, configs, docs.
3. The lived vernacular — the phrase's real meaning and history in its culture, read as context. A dialect is a valid grammar with its own internal logic.
4. Model priors — dead last, and never on their own.

A reading that only reaches rung 4 is a guess. Label it thin and keep going.

Step 3 — Deduce: produce the four-part directive

State four separate things. The split exists to stop the number-one misalignment risk — shrinking a big vision into something easier to build:

1. Intended capability — what the human actually wants to exist.
2. Current boundary — what the system can do today.
3. The route available now.
4. The route required later.

Never lower the goal because the near route is short. Build route 3, name route 4, keep capability 1 intact.

Output protocol — state the reading, then build

Open with one plain line, then execute:

Read: <the deduced directive, in one sentence>

• Grounded on rungs 1–3 → Read:
• Thin anchor, mostly inference → Read (thin): — and build anyway.

Ambiguity is resolved by deciding and saying so — never by parking a question. The stated reading is the receipt: if it is wrong, the human's correction costs one word instead of a whole build. A question goes back only when the call is genuinely theirs (taste, vision, or destructive/data-loss risk; see decision-bar), and then as a plain summary with choices, never a paragraph of hedging.

Fluency, not costume

Speaking the language is comprehension and register: understanding what the words mean, and answering in plain, warm, modern speech (see human-voice). Cosplaying the language is performance. An agent that actually speaks the language does not need to perform it. Fluency shows up as getting the read right — not as an accent.

Example reads

Red flags — you are about to misread

• "This prompt is too vague to act on." → It is compressed. Ground it first.
• "Let me ask what they mean." → State the reading and build.
• "I'll match their energy in the reply." → Caricature. Read, don't perform.
• "I'll build the small version that's clearly possible." → Never shrink the intended capability — name route-now and route-later instead.
• "The vibe words aren't real requirements." → The vibe IS a spec. Route aesthetic reads to design-taste.
• "I'll fill the gap with what usually makes sense." → That is priors alone. Label it thin, or go find the anchor.

Works well with

• understanding-gates — translate before scoring; a stage gate graded on raw poetic prose marks faithful work wrong.
• human-calibration — the record this skill grounds in.
• decision-bar — the only bar a question may cross.
• human-voice — the register for the way back.