Claude Prompting Principles
Stop issuing imperative commands to a frontier model. Assign it a senior role, name the deliverables, and make it reason before it writes. Prompts like "do this", "write code", or "fix this error" treat a senior model like a junior intern and get junior output.
The Core Shift
The instinct with a coding agent is to issue terse commands. That framing caps output at intern level because it hands over a task without a standard, a role, or a definition of done. The stronger pattern assigns a senior persona, enumerates the artifacts expected back, and forces analysis before code. Source: X/@ajitcodes thread, 2026-05-28
Three moves separate a senior prompt from a junior one:
- Assign a role. "Think like a senior full-stack engineer / debugging engineer / systems architect." The role imports the implicit standards of that role.
- Enumerate deliverables. Spell out the artifacts: architecture, file structure, schema, API surface, edge cases, usage examples. The model returns a complete package instead of a snippet.
- Force reasoning first. "First understand the architecture and data flow, then identify problems." Ordering analysis before code surfaces root causes instead of patching symptoms. Source: X/@ajitcodes thread, 2026-05-28
The Eight Archetypes
The thread packages the pattern into eight reusable role-prompts. Source: X/@ajitcodes thread, 2026-05-28
| # | Role | Returns |
|---|---|---|
| 1 | Senior full-stack engineer | Architecture, schema, endpoints, complete scalable MVP |
| 2 | New-joiner refactorer | Architecture summary, problem areas, refactor strategy (behavior unchanged) |
| 3 | Senior debugging engineer | Root cause, why it fails, edge cases, production-ready fix |
| 4 | Systems architect | Component structure, data flow, caching strategy, implementation |
| 5 | Performance engineer | Bottlenecks, inefficient logic, optimized code |
| 6 | Clean-architecture rebuilder | New folder structure, decoupled modules (behavior unchanged) |
| 7 | Four collaborating agents | Architect, Engineer, Reviewer, Optimizer roles in one pass |
| 8 | Senior frontend engineer | Reusable, accessible components with loading and edge states |
Relation to Kevin's Harness
Most of this is encoded structurally in Kevin's setup rather than typed per prompt. Path-scoped rules inject the "senior engineer" standard automatically (see Claude Code Harness); the ethos rule enforces completeness and reasoning-before-code; Harness Engineering treats definition-of-done as a loop invariant rather than a prompt suffix. Prompt 7 (four collaborating agents) is the manual version of subagent dispatch. The thread is a useful baseline for ad-hoc sessions without a harness; in a configured repo the same discipline already lives in rules and skills. Source: compiled from wiki, 2026-06-01
For art-direction and frontend visual prompting, the parallel reference is Prompt Recipes (explicit materials, type, motion budget, banned defaults). The shared principle across both: specificity and explicit standards beat vague commands, and naming the right component or role unlocks better output. Source: compiled, 2026-06-01
Deep Review Verdict
The 2026-06-30 bookmark review did not create a new executable skill from this thread. The thread's useful knowledge is already decomposed into existing operating rules: read architecture before editing, state the role and invariant, return artifacts rather than snippets, verify edge cases, and route large work through subagents or a loop. Creating a separate "senior Claude prompt" skill would duplicate Harness Engineering, Frontend and Design Skills, Debugging Pack, frontend design skills, and the agent identity rules.
The retained lesson is still important for unconfigured sessions: imperative commands create underspecified work. A good prompt names the standard of judgment before the work starts. But in Kevin's brain this should compile into harness structure, not live as eight copy-paste incantations. Source: X/@ajitcodes thread, 2026-05-28
Slash Command Vocabulary
The RoundtableSpace "32 Claude shortcuts" artifact is not evidence of built-in Claude commands. It is a vocabulary card for common prompt moves: explain, summarize, step through, checklist, executive summary, assume a role, adapt audience/tone, compare, use multiple perspectives, preserve context, constrain beginning/end, define role/task/format, emit schema, rewrite, reflect, check bias, slow down, self-evaluate, use parallel lenses, and reason from first principles. Source: X/@RoundtableSpace, 2026-04-06; Source: local image review, 2026-07-03
Treat the card as onboarding shorthand for unconfigured chats. In Kevin's harness, the durable version is structural: role, task, format, constraints, reviewer lens, and definition of done should live in agent files and skills. For reasoning prompts, ask for a short plan, assumptions, and verification evidence; do not require a private chain-of-thought transcript as the output.
Timeline
- 2026-07-03 | Deep-reviewed @RoundtableSpace's "32 Claude shortcuts" screenshot. Promoted it as slash-command vocabulary for unconfigured prompting, not a new Claude feature or executable skill. Source: X/@RoundtableSpace, 2026-04-06; Source: local image review, 2026-07-03
- 2026-05-28 | @ajitcodes posts the "Stop telling Claude do this" thread with 8 senior-role prompt templates. 4,355 likes, 12,814 bookmarks. Source: X/@ajitcodes, 2026-05-28
- 2026-06-01 | Absorbed from enriched X bookmark; principle extracted and cross-linked to harness and prompt-recipes. Source: compiled, 2026-06-01
- 2026-06-30 | Deep-reviewed the enriched record, self-thread, and artifact replies. Decision: keep as a refined concept/harness principle, not a new skill, because Kevin's agent stack already encodes role, deliverables, analysis-first, and reviewer/optimizer behavior structurally. Source: X/@ajitcodes thread, 2026-05-28