speechcanvas-free-expression-swarm

Four-agent image-prompt swarm (Muse drafts, Guardian safety-checks, Critic perfects realism, Composer finalizes) that produces lawful, consent-aware, non-deceptive symbolic imagery prompt packs as validated JSON for journalism, protest, censorship and free-expression themes. Use when the operator asks for a symbolic image or image prompt about protest, censorship, propaganda, press freedom, or free expression.

SpeechCanvas — Free-Expression Image Swarm (v2.0.3)

Turn a lawful creative brief into a validated symbolic image prompt pack through a 4-role refinement loop. Everything is JSON in, JSON out — any model can execute this.

When to use / when not to

• ✅ Symbolic, lawful imagery about protest, censorship, propaganda, press freedom, silence, debate.
• ❌ Realistic depictions of real people doing things they did not do, fake evidence, fake documents, hate or harassment imagery. → Decline and offer a symbolic alternative.

Hard safety rules (short form — full list in references/rules.md)

1. NEVER forge reality: no fake evidence, documents, ballots, screenshots, admissions, receipts, legal notices.
2. NEVER photoreal real people doing things they didn't; no real names, official seals, readable civic instructions.
3. NEVER private persons as targets/victims; no hate, harassment, dehumanization.
4. Deception is a SUBJECT, never a METHOD: deception-themed packs MUST carry at least one allowed motif (mask, mirror, frost, false crown, empty throne, frozen microphone, … full list in references/rules.md).
5. If a rule is unclear for the brief → ask the operator. Never guess on safety.

Workflow (reply-first, ≤3 iterations)

text

brief → Muse draft JSON ──► Guardian check ──► Critic critique ──► Composer final pack
              ▲                  │(FAIL)                              │
              └──────────────────┘ redraft, iteration+1 (max 3)       ▼
                                                     scripts/safety_validator.py + validate_pack.py
                                                                  PASS → deliver pack + image instruction

1. Reply first (speed): immediately show the operator a one-line concept + Muse's draft JSON. Run the rest while the operator reads. This halves perceived latency.
2. Muse drafts beauty: light, lens, texture, setting, gesture → JSON {subject, motif, lighting, lens, setting, gesture}.
3. Guardian checks every hard rule → {guardian_status: PASS|FAIL, reason, refined_prompt?}. FAIL → Muse redrafts with refined_prompt (iteration+1, max 3).
4. Critic checks realism → {critique, improved_details} (specific: "light too flat → cold blue moonlight + dust").
5. Composer merges all into the final prompt pack that validates against schema/prompt_pack.schema.json.
6. Deterministic gate (mandatory): run the validators — they are ground truth, not opinions:

   bash

   python3 scripts/safety_validator.py --file pack.json      # exit 0=PASS 1=BLOCK 2=WARN
   python3 scripts/validate_pack.py pack.json                # structural schema check
   

   Treat validator PASS as "no known violation pattern found" — not proof of safety; still apply the references/rules.md checklist judgment and ask the operator on ambiguity. No python3? Perform the checklist in references/rules.md manually, step by step, and say so.
7. Deliver: final pack JSON + a one-paragraph image-generation instruction built from it.

Role prompts (machine-readable: swarm/roles.json)

Load swarm/roles.json and use each role's goal / must / never / output_fields verbatim as that sub-agent's instruction. Roles are terse (≤80 words) on purpose: less input per call, fewer output tokens, faster generations. Roles may run in parallel where the orchestrator supports it (Guardian and Critic are independent after Muse's draft).

Output contract (the prompt pack)

Every final pack MUST validate against schema/prompt_pack.schema.json: subject, motif, lighting, lens, setting, gesture (≤200 chars each), constraints (array, must include the six safety constraints), safety_tags (enum: lawful, consent-aware, non-deceptive, free-expression, symbolic-atmosphere), iteration (1–3), guardian_status (PASS/FAIL — final packs must be PASS), critic_notes (≤500 chars), deception_theme (bool). Complete examples: references/examples.md (journalism, protest, censorship, debate, civil liberties).

Self-improvement (learn from every run)

After delivering, build the run record (prints to stdout, writes NOTHING by default):

bash

python3 scripts/record_run.py --brief-hash <sha256-16> --iterations N --guardian PASS \
  --critic "<one line>" --pack pack.json
# operator opted in to a persistent log? append explicitly:
python3 scripts/record_run.py ... --out ./speechcanvas_runs.jsonl

At start of a session, if the operator keeps a speechcanvas_runs.jsonl, read the last 5 records: recurring critic fixes → pre-inject into Muse's draft; recurring guardian failures → stricter redrafts. Default is stdout-only; file persistence happens only when the operator explicitly passes --out.

Token discipline

• SKILL.md (this file) is the only always-read file. Open references/ files only when needed.
• Scripts run WITHOUT reading them into context — execute, read the exit code and JSON verdict only.
• All role outputs are terse JSON; no prose narration between steps.
• Iteration cap 3. If still FAIL after 3 → return the safest passing variant or ask the operator.

Optional integrations (only if already installed — never assume)

• A local fast model: use it for Muse's first draft (Guardian/Critic/Composer stay on the primary model).
• A prompt cache: key = sha256(brief + constraints). Hit → skip the swarm, return cached pack.

Compatibility

Plain files + optional stdlib Python 3 scripts. No vendor-specific tool syntax, no network, no secrets, no shell required (scripts optional). Works in any SKILL.md-compatible agent (Claude Code, Cursor, Codex CLI, OpenClaw, …) and any model that can read JSON.