Skip to content

Latest commit

 

History

History
147 lines (113 loc) · 8.52 KB

File metadata and controls

147 lines (113 loc) · 8.52 KB

Language: English | 繁體中文

Reflective Prompt Library

This library collects reusable prompts for reflective engineering work: deciding what to do, turning ambiguity into specs, reviewing artifacts, designing tests, choosing agent workflows, and preserving context between sessions.

Core principle:

Doing the right thing > doing things right.

Use the prompts compositionally rather than as one giant instruction:

Core identity prompt
+ task prompt
+ context mode prompt
+ validation prompt

North Star

TeaPrompt helps humans and host agents choose the right amount of rigor for a task, record why decisions were made, and verify outcomes with evidence — using composable prompt layers, nine core workflow skills plus optional registered domain packs as natural-language harness policy, without operating its own agent runtime.

Strictness Before Skills

Pick Strictness L1–L6 first (skills/reflective-dispatch/SKILL.md, GLOSSARY.md), then the skill phase. Skills declare context_load: low|medium|high in frontmatter so cost-aware hosts can defer heavy skills when Strictness is low.

Standing non-goals are explicit: no TeaPrompt-owned runtime/swarm, no tenth core workflow skill without promotion evidence (frozen means gated, not never), and no domain pack inside core routing; optional packs follow the admission rule in 06-repo/AGENTS.md. Full SKILL.md i18n is not a default, and silent rigor downgrade is forbidden. Seeded routing evals are regression guards, not semantic-routing proof; methodology-complete does not mean operationally complete; skills may declare required runtime guarantees as preconditions, but host-runtime code and tests must enforce them. See PROJECT_KNOWLEDGE.md.

Governance Panel Record

Multi-agent Socratic consensus on project goals and the nine skills (Rounds 1–101, options A–IE) is recorded in plans/multi-agent-panel-consensus-2026-06-25.md. Run make all from the repository root before claiming routing or governance changes are verified.

Directory Map

  • 00-core/: core identity, custom instructions, master prompts, and controller prompts.
  • 01-thinking/: Socratic questioning, critical thinking, counterarguments, falsifiability, and Why / What / How / Done gates.
  • 02-engineering/: task start, specs, usage-first design, task slicing, implementation, review, tests, and local feedback.
  • 03-context/: small / medium / large context prompts, low-token mode, context engineering, Gemini long-document handling, and handoff summaries.
  • 04-agent/: agent selection, artifact promotion, external adoption review, workflow acquisition, SOP compiler design, workflow engine design, review-rating-fix loops, retros, and memory consolidation.
  • 05-domain/: high-risk review, research, business strategy, learning, writing, and creative template prompts.
  • 06-repo/: repository-level instruction templates plus a reusable PROJECT_KNOWLEDGE.template.md for project-specific design judgement.
  • skills/: nine core SKILL.md workflow wrappers plus optional registered domain packs; skill-map.md is the tier map.
  • skills/SKILL_TRIGGER_CHEATSHEET.md: one-page trigger cues and anti-misuse rules for the 9 core workflow skills, with a host-invoked domain-pack appendix.
  • skills/examples/: worked input/output examples for every registered core or domain-pack skill.
  • plans/: decision records, research reflections, eval fixtures, and governance contracts; live pointers are summarized in PROJECT_KNOWLEDGE.md.
  • plans/ROUTING_CONTRACT.md: routing fairness, observability, and risk-based default-up contract.
  • plans/route-001-paraphrase-eval.yaml: paraphrase routing eval seed for ROUTE-001.
  • GLOSSARY.md: operational definitions of load-bearing terms used across the library.
  • METHODOLOGY_MAP.md: maps the source methodology families to repo prompts and workflow skills.
  • LANGUAGE_POLICY.md: explains why operational docs are English while some prompt sources remain localized.

Install the workflow skills with SKILL_INSTALLATION.md. It covers Claude Code, Codex, Cursor, Antigravity CLI / IDE, and OpenCode.

Use METHODOLOGY_MAP.md before adding new prompts or skills. It keeps classification stronger than forced unification.

Skills as Workflow

The prompt files are the source material. The skills/ directory is the operational layer: a small set of repeatable workflows that can be copied into .claude/skills/, ~/.codex/skills/, .agents/skills/, or another SKILL.md-compatible environment when needed.

Workflow Need Skill
Route a task to the smallest useful reflective workflow skills/reflective-dispatch/SKILL.md
Clarify goal, assumptions, scope, acceptance, falsifiability skills/reflective-brief/SKILL.md
Write spec, usage-first design, task slices, a no-code Test Plan, or a no-code agent workflow design skills/reflective-spec-plan/SKILL.md
Implement code with verification and traceability skills/reflective-implement/SKILL.md
Challenge unnecessary code, dependencies, abstractions, or scope before implementation skills/reflective-minimality/SKILL.md
Review code, plans, specs, and AI outputs skills/reflective-review/SKILL.md
Research external docs, DeepWiki, and long sources skills/reflective-research/SKILL.md
Gate high-risk work before execution skills/reflective-risk/SKILL.md
Handoff, retro, and memory consolidation skills/reflective-handoff-retro/SKILL.md

The design intentionally avoids one skill per prompt. Use the prompt library for nuance and the skills for execution shape.

Skills are prompt-level workflow wrappers: natural-language procedures interpreted by the host runtime (Claude Code, Codex, OpenCode, etc.). They do not provide a multi-agent runtime, async messaging, or role isolation.

Recommended Defaults

Daily thinking:

core-short
+ socratic-reviewer
+ critical-thinking-check
+ why-what-how-done

Engineering task:

core-full
+ task-start
+ spec-writer
+ usage-first
+ task-slicer
+ implementation-agent
+ code-reviewer

High-risk task:

core-full
+ high-risk
+ local-feedback
+ human review gate

Long research:

research
+ large-context
+ gemini-long-document
+ context-handoff

Use Case Selection

Situation Recommended Prompt
You are unsure what the real question is 01-thinking/socratic-reviewer.md
You want to check whether a plan is flawed 01-thinking/critical-thinking-check.md
You want to avoid optimism or overengineering 01-thinking/counterargument.md
You want to start an engineering task 02-engineering/task-start.md
You want to write a spec 02-engineering/spec-writer.md
You want to design usage before implementation 02-engineering/usage-first.md
You want to slice work into tickets 02-engineering/task-slicer.md
You want an agent to implement code 02-engineering/implementation-agent.md
You want to review code or a diff 02-engineering/code-reviewer.md
You want to design tests 02-engineering/test-designer.md
The task is high-risk 05-domain/high-risk.md
Something failed 02-engineering/local-feedback.md
You need to switch sessions 03-context/context-handoff.md
A task is complete and you want process learning 04-agent/retro.md
You need to decide Prompt vs Agent vs Workflow 04-agent/agent-selection.md
You need to convert a repeatable human SOP into an agent workflow 04-agent/sop-compiler.md
You need to classify memories, discussions, or repeated procedures into knowledge, prompt, skill, verifier, runtime, or no-change 04-agent/artifact-promotion.md
You need to capture and replay an observed workflow before turning it into a skill, verifier, or runner 04-agent/workflow-acquisition.md
You need to evaluate whether an external tool, paper, repo, or workflow method should change TeaPrompt 04-agent/external-adoption-review.md
You need to design or review runtime trust boundaries 04-agent/runtime-trust-boundary.md
You need to review a leaked, mirrored, or third-party agent scaffold 04-agent/agent-scaffold-provenance.md
You need a project-specific design-judgement layer without creating a second agent rulebook 06-repo/PROJECT_KNOWLEDGE.template.md + skills/reflective-handoff-retro/SKILL.md