Production

Production Mode

Load the full template from ~/projects/animus/packages/forge/prompts/modes/production.md and use it as the structural contract.

Process

1. Gather INPUTS from the user. Missing slots are the #1 cause of filler output — ask before producing. Required slots:

   • deliverable — one artifact (email, PR description, landing copy, decision memo, code change)
   • objective — what the reader does after reading
   • audience — who reads it
   • constraints — measurable ("350 words", "≤ 3 bullets", "subject < 60 chars")
   • voice / format — existing pattern to match ("matches the tone of ")
   • evidence base — every claim maps here; responsibilities from a role, bullets from a resume, commits from a PR
   • exclusions — must not appear (hedging, filler, generic adjectives)

2. Produce the deliverable in the exact requested format. Nothing else.

   • NO preamble ("Here is the email you requested...")
   • NO wrapper ("```" fences around prose that isn't code)
   • NO trailing sign-off unless requested
   • NO meta-commentary about the choices made

3. Honor FAILURE CONDITIONS — self-check before emitting:

   • Any exploratory language ("we could consider", "one option") → delete, wrong mode
   • Any meta ("I've structured this as...") → delete
   • Filler phrases ("it's worth noting", "importantly", "in essence") → delete
   • Unsourced quantifiers ("significantly improved X") → replace with a number from evidence OR delete
   • Soft-skill boilerplate ("strong communicator") → delete unless evidenced
   • Format violation (exceeded word count, missing section, wrong structure) → rewrite
   • Hedging the deliverable ("you may want to consider...") when a decision is requested → replace with the decision

Anti-pattern: "make it better"

If the deliverable scores below band B on the personal-quality rubric, DO NOT respond to "make it better" — too vague, will regress or add filler. Instead:

1. Run the output through /evaluation against personal-quality
2. Identify the 3 lowest-scoring dims
3. Re-enter /production with those named in INPUTS.exclusions (e.g. "exclusions: generic adjectives flagged in prior pass: [list]")

Measured iteration beats "try again" prompts 2-4x in practice.

Build-mode goal-loop (code/build deliverables ONLY)

If and only if deliverable is a code/build artifact (a module, feature, multi-file change, or system — NOT prose like an email, cover letter, copy, memo, or docs), ENGAGE the goal-loop by writing a .goal-active sentinel at the repo root:

objective: <INPUTS.objective, one line>
stop_condition: <falsifiable end state, from acceptance criteria>
budget_ceiling_usd: 10
turn_ceiling: 25
turns: 0
spent_usd: 0
mode: engaged
started_at: <ISO8601>

Then ensure .goal-active is git-ignored (add to .gitignore or .git/info/exclude). The goal-loop hook injects the objective + budget each turn and forces a checkpoint at the ceiling; goal-gate escalates destructive ops for approval. For prose deliverables, do not write the sentinel. Manage an active goal with /goal.

Scoring

animus-forge eval run <suite> --rubric personal-quality --prompt-version production-v1

Weights that matter most: format_compliance (hit every format constraint), precision (no generic boilerplate), concision (every sentence earns its keep), decision-clarity (reader knows exactly what they're asked to do).