Content Engine

step 2 — load the calibration examples

read memory/approved-drafts.jsonl. take the last 20 entries (or all of them if fewer than 20). these are the strongest available signal for "what rishabh actually approves." treat them as few-shot training examples.

if the file is empty (fresh install, no approvals yet), read memory/humor-calibration.md as the fallback calibration.

step 3 — draft

if task == "reply":

• 1–3 sentences. no more.
• must contain a specific claim backed by receipt (experience / data / structural observation).
• must use at least one humor mode (structural / precision / scale / deadpan / self-aware escalation / turn / quiet bomb / compression).
• one insider angle if possible. use memory/career-context.md for what counts as insider.
• never agree fully. per §conversation-depth, agreement closes the loop. productive tension keeps it open. even when rishabh would basically agree, the draft should add a distinct angle or gentle pushback, not a "great point!" reply.
• never end with a closed statement. leave it open — a specific observation, a reframe, a question that invites further response.

if task == "original":

• 2–5 sentences. aim for distilled.
• pick a hook pattern explicitly and tag it in the output: cold-open-with-number, contrarian-signal, insider-reveal, pattern-interrupt, tension, or a new pattern that earns its own tag.
• first 100–140 chars should be a standalone hook (it's what shows in the notification preview).
• must pass the §18 test: sound like something rishabh was going to think anyway.

step 4 — self-check before validation

before calling voice-check, check your own draft against these top 5 mistakes aria tends to make:

1. does it contain "delve", "nuanced", "landscape", or any other §16.3 banned word? if yes, rewrite.
2. does it start with "so,", "here's the thing", "i've been thinking", or any other banned opener? if yes, rewrite.
3. does it have an em-dash or semicolon? if yes, replace with periods.
4. does it have a parallel triad ("not just X, but Y and Z")? if yes, restructure.
5. does it end with "think about that" / "let that sink in" / similar? if yes, cut.

if you had to rewrite any, your internal draft is now attempt 1. if you rewrote nothing, proceed.

step 5 — call voice-check

call skills/voice-check with {draft: <your draft>, task: <task>, source_context: <target or topic>}.

receive {verdict, failures, notes}.

• if verdict == "PASS" → go to step 7.
• if verdict == "FAIL" → go to step 6 for exactly one redraft attempt.

step 6 — redraft (only once)

read the failures list. for each failure, apply the specific fix:

• banned word → replace with a rishabh-native synonym or restructure the sentence
• banned opener → cut the first 3 words, start from the real sentence
• banned ending → cut the last sentence or rewrite it to leave a thread open
• em-dash → period
• semicolon → period, split into two sentences
• parallel triad → cut one of the three items or restructure
• no receipt → add a specific claim from rishabh's career-context
• no emotional truth → decide which of the 3 truths to target, rewrite to hit it
• no humor mode → add one structural / precision / scale / deadpan / turn move
• fails §18 → this is the hardest to fix — rewrite from scratch with the note as guidance, or give up

call voice-check again with the redrafted version.

• if verdict == "PASS" → go to step 7.
• if verdict == "FAIL" → go to step 8 (give up). do NOT attempt a third redraft.

step 7 — emit PASS