Workflow Prompts for Claude Code¶

NON-NORMATIVE. This is an operational reference. Normative rules are in docs/governance/morphism-kernel.md.

Curated prompts for common workflows in this repo. Each aligns with AGENTS.md, CLAUDE.md, the handoff guide, and the agent kernel. Copy-paste the prompt, or use the slash command if one exists.

1. Session Start (Bootstrap)¶

When to use: At the beginning of a session so Claude Code has context before doing anything.

The prompt:

Before doing anything: run session bootstrap. (1) git log --oneline -15 and summarize recent work. (2) git status — branch and uncommitted changes. (3) Read docs/operations/current-handoff.md if it exists. (4) Read docs/operations/deferred-work.md for P0/P1 items. (5) Read AGENTS.md for governance. Then produce a short status: branch, last 3 commits, uncommitted changes, prior handoff summary, active P0/P1, and "ready for" based on context. Do not take other action until this is done.

Or use: /bootstrap

2. Governance Baseline (Before Structural Changes)¶

When to use: Before changing layout, adding/removing apps or packages, or touching governance docs.

The prompt:

Before making any structural or governance changes: read AGENTS.md, SSOT.md, GUIDELINES.md, docs/operations/handoff-and-integration.md, and docs/operations/deferred-work.md. Then state the one thing we're doing, verify the path (files/tools in scope), and only then execute. One change at a time; verify after each change. Refuse scope creep.

Or use: /governance-baseline

3. Scoping a Single Task¶

When to use: When you want one clear goal and small, verifiable steps.

The prompt:

For this task: (1) State the one thing — single-sentence goal and "done means" (testable). (2) List which files/systems are in scope and what's out of scope. (3) Execute the smallest complete change, run the relevant test/lint (TS: npx turbo typecheck && npx turbo lint; Python: ruff check + mypy; governance: python scripts/verify_pipeline.py), then proceed. Do not plan or batch changes — do it. If you need more than 2 attempts, stop and ask me.

Or use: /scope

4. Handoff (Session End)¶

When to use: Before ending a session so the next one (or another agent) can continue.

The prompt:

Before ending: write docs/operations/current-handoff.md with (1) what was done this session, (2) what remains, (3) current branch and uncommitted changes, (4) any blockers. Then summarize that in your reply.

Or use: /handoff

5. Context Switch (From Cursor or Another Agent)¶

When to use: When switching from Cursor (or another tool) to Claude Code.

The prompt:

I'm switching from [Cursor/other agent] to Claude Code. Load baseline: read AGENTS.md, SSOT.md, GUIDELINES.md, docs/operations/handoff-and-integration.md, and docs/operations/deferred-work.md. Use the same governance and work style (execute don't plan, one change at a time, verify after each change). Then run bootstrap (git log -15, git status, current-handoff.md, deferred-work.md) and give a short status before taking new requests.

Or use: /context-switch

6. Pre-commit / Validation¶

When to use: Before committing or opening a PR.

The prompt:

Run the relevant verification for what we changed: TypeScript — npx turbo typecheck && npx turbo lint && npx turbo test; Python — ruff check src/ tests/ && mypy src/ && pytest tests/; docs or governance — python scripts/verify_pipeline.py. Report pass/fail and any errors. Do not commit until these pass.

Or use: /governance-check (governance-only) or /audit (full pipeline)

7. Refusal / Stuck¶

When to use: When you want Claude Code to stop guessing and ask you.

The prompt:

If constraints are missing, sources conflict, or you're stuck after 2 attempts: refuse clearly. State what's missing or what failed, and ask me for the minimum input or direction. Do not guess or spiral.

Or use: Always active — encoded in CLAUDE.md Work Style rule 3 and agent-kernel-prompt.md Section 7.

8. Environment (Windows 11 Native)¶

When to use: When paths or tools behave oddly, or when Claude Code might assume Linux.

The prompt:

This repo runs on Windows 11 native (not WSL). Use python (not python3). Shell is Git Bash — use Unix syntax. Git uses LF — do not introduce CRLF. If jq is missing, use python -c "import json, sys; ..." for JSON. Check tool availability with which before assuming a command exists.

Or use: Always active — encoded in CLAUDE.md Environment section.

9. Commit and Branch Rules¶

When to use: Before committing or creating a branch.

The prompt:

Before committing: use Conventional Commits (feat:, fix:, chore:, docs:, etc.). Branch names must match (feat|fix|chore|ci|docs|refactor|test|perf|hotfix)/<description>. Validate with: echo "feat(scope): subject" | python scripts/validate_commit.py /dev/stdin and python scripts/validate_branch.py. Commit after each logical unit; do not stack a large uncommitted changeset.

Or use: Partly covered by /governance-check — commit/branch validation is also available as standalone scripts.

10. Area-Specific Work¶

When to use: When working in a specific area (TypeScript, Python, docs) and you want the right test/lint commands loaded.

The prompt (pick the relevant sub-section):

TypeScript (apps or packages): We're working in [apps/morphism|packages/shared]. Next.js/TS rules apply. After changes run: npx turbo typecheck && npx turbo lint && npx turbo test. Match the existing layer: apps depend on packages; packages have no app-specific deps.

Python core: We're working in src/morphism/ (Python core). Use ruff and mypy. After changes run: ruff check src/ tests/ && mypy src/ && pytest tests/. Preserve the CLI and category-theory engine contracts.

Docs or SSOT: We're working in docs/ or SSOT atoms. After editing AGENTS.md, SSOT.md, or GUIDELINES.md (within SSOT markers), run: python scripts/ssot_extract.py then python scripts/ssot_verify.py and python scripts/docs_sync.py --check. Run python scripts/docs_frontmatter.py --check for frontmatter validation. Run python scripts/verify_pipeline.py for full governance checks.

Or use: Encoded in CLAUDE.md Test Gates and Monorepo Layout sections.

11. Structured Refusal¶

When to use: When you want refusals to follow the kernel format rather than free-form.

The prompt:

If you must refuse (out of scope, under-specified, SSOT conflict, entropy increase, or invariant break): output (1) which tenet(s) are violated, (2) why, (3) the minimum info or action to proceed, (4) a safe fallback if any. Do not guess or proceed when constraints fail.

Or use: Always active — encoded in agent-kernel-prompt.md Section 7 (refusal template). Prompt #7 is the simpler version.

12. Long Session / Rate Limit¶

When to use: When the session is getting long or output is truncating.

The prompt:

We're in a long session. Prioritize: (1) finish the current atomic change, (2) run verification, (3) commit with a conventional message, (4) write docs/operations/current-handoff.md (what was done, what remains, branch, uncommitted changes, blockers). Do not start new work.

Or use: /handoff — covers step 4 automatically.

13. Full Protocol (Read → State → Verify → Execute → Refuse)¶

When to use: When you want the full kernel protocol on a non-trivial request.

The prompt:

For this request follow the MORPHISM protocol: (1) READ — list what you treat as canonical (intent, contracts, ADRs, policies); say what's missing. (2) STATE — the one thing, done means, scope, non-goals, minimal assumptions. (3) VERIFY — check I-1..I-7 and any drift/scope/refusal conditions. (4) EXECUTE — smallest change set that satisfies done means. (5) If constraints fail, REFUSE using the refusal template (tenet violated, why, minimum to proceed, fallback). Ref: agent-kernel-prompt.md.

Or use: The full protocol is documented in agent-kernel-prompt.md. This prompt activates it for a single request without requiring the agent to have read that document.

14. One-Shot Master Prompt (Session Start)¶

When to use: Single prompt that loads governance, bootstrap, and work style in one go — useful when pasting into a fresh session without slash commands.

The prompt:

You're working in the morphism-systems monorepo. Governance: AGENTS.md, SSOT.md, GUIDELINES.md, docs/operations/handoff-and-integration.md, docs/operations/deferred-work.md. Protocol: read first, state the one thing, verify path, execute incrementally, refuse scope creep. Work style: execute don't plan; one change at a time; verify after each change (TS: npx turbo typecheck && npx turbo lint; Python: ruff + mypy + pytest; governance: python scripts/verify_pipeline.py; docs frontmatter: python scripts/docs_frontmatter.py --check); if stuck after 2 attempts, stop and ask. Environment: WSL2, bash, /mnt/c paths, LF line endings. Now run bootstrap: git log --oneline -15, git status, read docs/operations/current-handoff.md if present, read docs/operations/deferred-work.md for P0/P1, then give a short status (branch, last commits, uncommitted, handoff summary, P0/P1, ready for). After that I'll give the task.

Or use: /bootstrap covers the bootstrap portion; this prompt adds governance loading and work style on top.

Summary¶

When	Prompt / action
Start session	Bootstrap (#1) or `/bootstrap`
One-shot load	Master prompt (#14)
Structural / governance work	Governance baseline (#2) or `/governance-baseline`
Any concrete task	Scoping (#3) or `/scope`
End session	Handoff (#4) or `/handoff`
Switch from Cursor	Context-switch (#5) or `/context-switch`
Before commit	Pre-commit (#6), `/governance-check`, or `/audit`
When stuck	Refusal (#7) (always active)
Paths / WSL issues	Environment (#8) (always active)
Commit / branch naming	Commit rules (#9) or standalone scripts
TS / Python / docs	Area-specific (#10)
Structured kernel refusal	Structured refusal (#11)
Long session	Long-session (#12) or `/handoff`
Non-trivial, full rigor	Full protocol (#13)