engram-codebook

name	engram-codebook
description	Build an engram codebook for Claude prompts — compresses identity (CLAUDE.md) and prompt vocabulary into key=value pairs to reduce per-turn token cost

Engram Codebook Wizard

Build a two-part codebook that reduces Claude context overhead:

• Identity codebook — compresses who Claude is (CLAUDE.md → key=value, sent once per session)
• Prompt codebook — compresses what you ask Claude (base slots + project vocabulary)

Step 1: Choose scope

Ask the user:

"Where should this codebook apply?

1. Session — in-memory only, savings this session, nothing written to disk
2. Project — saved to ./.engram-codebook.yaml in the current project root
3. Global — saved to ~/.engram-codebook.yaml, applies to all sessions"

Store the answer as [scope]. Proceed to Step 2.

---

Step 2: Read CLAUDE.md

Based on [scope]:

• Session or Project → read CLAUDE.md from the current project root
• Global → read ~/.claude/CLAUDE.md

If the file does not exist, inform the user: "No CLAUDE.md found at [path]. You can continue and I'll build the prompt codebook only, or create a CLAUDE.md first."

Proceed to Step 3.

---

Step 3: Derive identity dimensions

Call mcp__plugin_engram_engram__derive_codebook with:

• content: the full CLAUDE.md text read in Step 2

Display the derived dimensions to the user in a table:

Key	Type	Derived value
(from result)

Then supplement with these Claude-behavioral dimensions that derive_codebook typically misses from prose-style CLAUDE.md files. The values below are defaults based on common CLAUDE.md patterns — review each and override if your file uses different conventions:

Key	Type	Suggested value	Common pattern (verify against your CLAUDE.md)
plan_mode	enum: always,non_trivial,never	non_trivial	"Enter plan mode for ANY non-trivial task"
subagent_use	enum: liberal,minimal,never	liberal	"Use subagents liberally"
self_improve	bool	true	"update tasks/lessons.md after corrections"
verify_before_done	bool	true	"Never mark a task complete without proving it works"
elegance_check	enum: balanced,always,skip	balanced	"Demand Elegance (Balanced)"
autonomous_bugfix	bool	true	"When given a bug report: just fix it"
response_style	enum: concise,verbose	concise	"Make every change as simple as possible"
task_tracking	enum: todo_md,inline,none	todo_md	"Write plan to tasks/todo.md"
code_quality	enum: senior,standard	senior	"Senior developer standards"
impact_scope	enum: minimal,broad	minimal	"Impact minimal code"

After review, ask: "Any additional custom identity dimensions to add?" If yes, collect key, type, and value for each.

Proceed to Step 4.

---

Step 4: Build prompt codebook

Tell the user: "Now building the prompt codebook — base slots that compress any task prompt."

Add these base template slots (no user review needed — these are universal):

prompt_task:
  type: enum
  values: [implement, fix, refactor, explain, review, plan]
  description: "The primary action being requested"

prompt_scope:
  type: text
  description: "What file/component/system is affected"

prompt_style:
  type: enum
  values: [concise, thorough, tdd, elegant]
  description: "How the work should be approached"

prompt_constraint:
  type: text
  description: "What NOT to do or hard limits"

prompt_target:
  type: text
  description: "Success criteria for this prompt"

Then scan the project for domain vocabulary. Use Grep to find the most frequent meaningful terms in CLAUDE.md and file names under the project root. Encode the top terms (skip common English words):

vocab_skill:
  type: text
  description: "A Claude Code skill file (SKILL.md)"

vocab_agent:
  type: text
  description: "A Claude Code subagent dispatched via Agent tool"

vocab_subagent:
  type: text
  description: "Subagent handling a focused sub-task"

vocab_plan:
  type: text
  description: "Implementation plan doc in docs/superpowers/plans/"

vocab_codebook:
  type: text
  description: "Engram codebook YAML defining compression dimensions"

vocab_hook:
  type: text
  description: "Claude Code hook script triggered by tool events"

vocab_session:
  type: text
  description: "The current Claude Code conversation session"

Show the user the vocabulary list. Ask: "Any terms to add or remove?"

Proceed to Step 5.

---

Step 5: Check for existing codebook and confirm overwrite

Determine the target path:

• Session → no file write, skip to Step 6
• Project → ./.engram-codebook.yaml
• Global → ~/.engram-codebook.yaml

For Project or Global scope: check if the file exists (use Read tool and handle the "file not found" case).

If a file exists, display a warning:

"A codebook already exists at [path]. Overwriting will replace all existing dimensions. Continue?"

If the user declines, stop. Offer: "You can cancel and manually merge new dimensions into the existing file, or run the engram:codebook skill's diff subcommand to see what would change."

If the user confirms (or no file exists), proceed to Step 6.

---

Step 6: Assemble and write the codebook YAML

Assemble the final YAML by merging:

1. Auto-derived dimensions from Step 3
2. Reviewed behavioral dimensions from Step 3
3. Prompt base slots from Step 4
4. Project vocabulary from Step 4

Use this format (all dimensions live at the top-level dimensions key — engram's schema):

# Engram codebook — generated by engram-codebook skill
# Identity dimensions compress CLAUDE.md behavioral instructions.
# Prompt dimensions (prefix: prompt_, vocab_) compress task prompts.
name: [project-name or "global"]
version: 1
dimensions:
  plan_mode:
    type: enum
    value: non_trivial
  subagent_use:
    type: enum
    value: liberal
  self_improve:
    type: bool
    value: "true"
  verify_before_done:
    type: bool
    value: "true"
  elegance_check:
    type: enum
    value: balanced
  autonomous_bugfix:
    type: bool
    value: "true"
  response_style:
    type: enum
    value: concise
  task_tracking:
    type: enum
    value: todo_md
  code_quality:
    type: enum
    value: senior
  impact_scope:
    type: enum
    value: minimal
  prompt_task:
    type: enum
    value: null
  prompt_scope:
    type: text
    value: null
  prompt_style:
    type: enum
    value: null
  prompt_constraint:
    type: text
    value: null
  prompt_target:
    type: text
    value: null
  vocab_skill:
    type: text
    value: "skill_file"
  vocab_agent:
    type: text
    value: "subagent"
  vocab_subagent:
    type: text
    value: "subagent"
  vocab_plan:
    type: text
    value: "impl_plan"
  vocab_codebook:
    type: text
    value: "engram_cb"
  vocab_hook:
    type: text
    value: "cc_hook"
  vocab_session:
    type: text
    value: "cc_session"

For Session scope: do not write a file. Instead call mcp__plugin_engram_engram__compress_identity passing the assembled dimensions as a structured object (not a raw YAML string). Pass each dimension key-value pair from the assembled codebook as the identity parameter content. Report what was compressed.

For Project or Global scope: write the assembled YAML to the target path using the Write tool.

---

Step 7: Show sample compression (testing phase)

Note: This step is for the testing phase only. Remove once encoding is validated.

Show the user a before/after example:

Before (plain prompt):

"Create a new skill that uses subagents to parallelize the plan execution and verify it works before marking done"

After (codebook-encoded):

prompt_task=implement prompt_scope=vocab_skill prompt_style=concise prompt_target=parallel_subagent_plan verify_before_done=true

Token count before: ~22 tokens. After: ~12 tokens (illustrative — actual savings vary by prompt). ~40–50% prompt compression typical.

---

Step 8: Confirm and show savings estimate

Report success:

"Codebook saved to [path] (or: held in session memory).

Estimated savings:

• Identity: ~400 tokens → ~20 tokens per session start (~95% reduction)
• Per-turn identity re-send: eliminated after turn 1
• Prompt compression: ~30–50% per encoded prompt

On a 50-turn session this saves approximately 19,000+ tokens (theoretical maximum) — equivalent to ~15 pages of usable context added back to your window.

To check redundancy in future responses, call mcp__plugin_engram_engram__check_redundancy with the response content."