Process bulk prompts with stop-start play/think cycles. Save first, lose nothing, discover value.

Activation

User signals: "batch incoming" / "multiple prompts incoming" / "corpus incoming"

Respond: 🔁 BATCH MODE — send them. I'll save everything first, then process one-by-one.

Step 1: SAVE (mandatory, before any processing)

Parse input into individual prompts (split on blank lines or ---). Write entire batch to workspace/systems/batch-cognition/batches/YYYY-MM-DD-HHMMSS.md. Read prior value-stack.md and last batch's meta-think for cross-batch context. Format:

# Batch: [timestamp]
# Source: [telegram|file|drive|paste]
# Total: [N]
# Status: SAVED
1. [first 60 chars of prompt]
[ ] PENDING
[full prompt text]
2. [next prompt]
[ ] PENDING
[full prompt text]

Confirm to user: "✅ Saved [N] prompts to batch doc. Starting processing."

Step 1.5: PRE-SCAN & CLASSIFY (per item, before processing)

Read first 100 chars of each item. Classify type and assign depth budget:

Step 2: PLAY (per prompt)

Execute the prompt. Not summarize — EXECUTE. The depth must match the item:

• Website/app: hours × $0 (we build) + API costs + hosting. Estimate tokens for AI-assisted build.
• Script/tool: lines of code estimate → token estimate (1 LOC ≈ 10-20 tokens to generate)
• Research: number of searches + fetches × ~500 tokens each
• Total = build tokens + test tokens + fix tokens (budget 30% extra for iteration)

"Solid" means tested. First pass is never solid. Flag items that need a second pass.

Append output under the prompt entry. Update status to [~] PLAYING. Take factual notes: what was done, what was produced, what was discovered.

Step 3: THINK (per prompt, immediately after PLAY)

Answer 5 questions (keep tight, 1-2 lines each):

• Learned: factual takeaway
• Pertinent: relevant to user's current projects/goals?
• Value: creative thinking → value creation → money path?
• Act?: yes/no — if yes, smallest next step
• Future: park / discard / investigate deeper

Tag: 🟢 ACT NOW | 🟡 PARK | 🔴 DISCARD | 🔵 INVESTIGATE

Update status to [x] DONE with tag.

Step 4: CHECKPOINT (every 5 prompts)

Brief summary: what's covered, patterns emerging, top value items so far. Ask: "Continue, pause, or pivot?" — if no response in 30s, continue.

Step 5: META-THINK (on "done" or batch exhausted)

Review all Think notes. Produce:

• Value Stack — ranked by expected value
• Patterns — themes, connections
• Action Items — concrete next steps
• Park List — not now but maybe later
• Discard List — safe to ignore (1-line reason each)

Append to batch doc. Update status to COMPLETE. Append 🟢 items to systems/batch-cognition/value-stack.md. Append 🟡 items to systems/batch-cognition/parked.md. Append 🔴 items to systems/batch-cognition/discarded.md. Log any cross-batch connections to systems/batch-cognition/connection-graph.md.

Commands (user can say anytime)

skip — skip current prompt | deeper — spend more tokens | park — park for later pause — stop, resume later | resume — continue paused batch | status — show progress value stack — show current ranked items | done — trigger meta-think + close

Context Management

Rolling decay memory — each checkpoint creates a new block in a chain. Items decay 20% per block. Referenced items reset to full weight. Below 0.2 = archived (never lost). See references/rolling-decay-memory.md for full spec.

At each checkpoint:

• Score all items: salience *= 0.8, re-referenced items → 1.0
• Drop items below 0.2 threshold (write tombstone, archive to disk)
• Carry forward: rolling summary + surviving high-salience items + value stack
• New block header written to chain file

Cross-batch: last batch's survivors enter new batch at 0.8, connect to new items → reset to 1.0.

See references/drive-mode.md when processing Drive folder dumps.

Self-Improvement

After each batch, append to workspace/systems/batch-cognition/learnings.md:

• What worked / what was wasted / grouping improvements / play-think split effectiveness
• Update this skill if any rule changes warranted.

Key Rules

• SAVE FIRST — never process before saving full batch to file
• Never lose input — if Telegram truncates or splits, wait for "done" before processing
• Play and Think are separate — don't blend execution with inference
• Notes are mandatory — both Play notes and Think notes, every prompt
• Prompts aren't always instructions — some are ideas, half-thoughts, value discovery. THINK phase handles these.
• PLAY means EXECUTE, not summarize — if it says build, build. If it says research, research with sources. If it's an idea, scope it with real numbers.
• First pass is never final — flag items that need deeper work. A 🟢 tag without execution is just a bookmark.
• Prototype cost is mandatory for buildable ideas — hours, tokens, API costs, hosting. Not vibes.