tap-agent-authoring

Overview

This skill encodes everything needed to write a correct, high-quality agents/<domain>/<spec>.toml file for the octomind-tap registry. It covers the exact TOML format, required and forbidden fields, how to write effective system prompts, temperature/top_p guidelines by domain, workflow and layer patterns for multi-step pipelines, and the pre-write checklist.

Use this skill whenever you are creating or editing an agent manifest.

Instructions

Agent Manifest Format

# agents/<domain>/<spec>.toml
# Agent: <domain>:<spec>
# Title: Short Agent Title (5–60 chars)
# Description: What this agent does (20–160 chars).

capabilities = ["core", "filesystem-read", "filesystem-write", "shell", "codesearch-semantic", "codesearch-structural", "codesearch-graph", "programming-python"]

[[roles]]
system = """
<XML-tagged system prompt — see "System Prompt Structure" below>
"""
welcome = "<Emoji> <Short greeting>. Working in {{CWD}}"
temperature = 0.3
top_p = 0.9
top_k = 0

# Optional model override:
# model = "openrouter:anthropic/claude-sonnet-4"

Required Fields

Optional Fields

Forbidden in Agents

System Prompt Structure (mandatory — 2026 standard)

System prompts must use XML-tagged blocks in a fixed order. This is grounded in three findings: (1) the U-shape attention curve — beginning and end of the prompt get the most attention, the middle gets "lost"; (2) Anthropic's published guidance that XML tags are the preferred structuring method for Claude; (3) Claude Opus 4.7 follows instructions more literally than 4.6 — implicit gaps don't get bridged anymore, so structure has to be explicit.

The eight-block order, top to bottom:

<identity>          ← primacy: who/what (3–5 lines, role + expertise + style)
<voice>             ← tone, register, formality (omit if technical agent)
<scope>             ← what's owned / what routes elsewhere
<workflow>          ← numbered steps for the main pipeline
<rules>             ← decision rules, tables, domain knowledge
<examples>          ← good/bad pairs (omit if not applicable)
<output_format>     ← exact artifact shape, file location, schema
<interaction>       ← trigger → response patterns
<critical>          ← recency: brief Don't/Do list in plain language; reserve all-caps for one or two genuine safety hard-stops

Why this order:

• <identity> first because primacy: the model commits to role early
• <critical> last because recency: the tail is load-bearing — it's the last thing the model sees before the user message and gets disproportionate attention
• <workflow> and <rules> in the middle have to do their work via clear tagging, since middle-of-prompt content is the "lost" zone — the XML tags act as retrieval anchors

Where {{CWD}}, {{DATE}}, and other dynamic placeholders go (and where they don't):

The system prompt must be STABLE run-to-run. Anything that changes between runs breaks prompt caching, which costs real money and latency on every call.

• {{CWD}} and {{DATE}} change every run — they belong in the welcome field ONLY.
• The harness pipes working directory and current date as separate runtime context. Putting them in system duplicates that state AND breaks the cache.
• The rule: if a placeholder's value changes from one session to the next, it does not belong in system. System is identity + rules + structure — all stable. Welcome handles "where am I, when am I" dynamics.
• Static placeholders like {{INPUT:KEY}} and {{ENV:KEY}} typically don't appear in system anyway — they live in MCP server config.

The lint enforces this: {{CWD}} or {{DATE}} anywhere inside system = """...""" is a hard error.

Block-by-block authoring rules:

• <identity> — One paragraph. Role, expertise, style. Don't bury this; the model uses it to pick its register.
• <voice> — Only include if voice/tone matters (content, marketing, video, creative agents). Omit for pure technical agents (developer, devops, security) where voice is uniform.
• <scope> — Two lists: ✅ Own / ❌ Don't own. Required for any agent that sits next to siblings (e.g. seo:audit vs seo:strategist, or content:article vs content:blog). Describe out-of-scope work by what it is, not by routing to a specific other-domain agent — domain isolation applies to agent bodies too.
• <workflow> — Numbered steps (1, 2, 3…). Each step is one sentence + optional sub-bullets. Don't write paragraphs.
• <rules> — Tables and bullet lists win here. Decision matrices, parameter ranges, vocabulary lists. This is the agent's domain knowledge.
• <examples> — Bad → Good pairs are the strongest format. Skip this block if the agent's output is freeform.
• <output_format> — Exact structure of the artifact. If the agent saves files, name the path pattern (./video-out/<slug>/script.md). If the agent returns markdown blocks, show the schema.
• <interaction> — Trigger patterns. "When user says X → do Y". Especially "Ambiguous → ask ONE clarifying question."
• <critical> — Brief Don't/Do list in plain language: things the model could plausibly get wrong if not warned, and the corresponding correct action. Reserve all-caps NEVER/ALWAYS for one or two genuine safety hard-stops (e.g. Never force-push to main); stacking more dilutes attention on Claude 4.6+ (see "Tone calibration"). No CWD/DATE placeholders, no preamble, no 🚨 HARD RULES theatre.

Token discipline (Claude 4.7 / 2026):

• Target: 200–1000 words for the full system prompt is the production sweet spot
• Soft warning at 1500 words (SYSTEM_LENGTH_WARN) — context rot risk
• Hard cap: 3000 words. Above this, lint-manifests.sh fails with SYSTEM_TOO_LONG — extract documentation/reference content into skills.
• Be explicit — Claude 4.7 doesn't bridge implicit gaps the way 4.6 did
• Static content first for prompt caching (saves up to 90% cost on repeat sessions)
• No decorative prose — every line must do work; if you can cut it without losing meaning, cut it

Authoring patterns that work:

• Specific over generic — "Use uv run pytest for testing" beats "Run the tests"
• Tool commands inline — actual CLI commands for the domain (cargo, kubectl, npm, etc.)
• Set boundaries explicitly — "Fix X" means fix X only, not refactor the neighborhood
• Reference example agents — study developer:general for execution protocol; agents/content/article.toml for content-style structure

Anti-patterns that break the structure:

• Free-form markdown sections without <tag> wrappers — defeats the U-shape strategy
• IDENTITY / WORKFLOW as plain ## headers — Claude treats them as content, not structural anchors
• Long preambles before <identity> (e.g. "Welcome to this agent...") — wastes the primacy slot
• Critical rules buried in the middle — they get "lost"
• <critical> block with paragraphs of explanation — keep it tight: brief Don't/Do bullets, that's it
• Stacked theatrical emphasis — 🚨 HARD RULES, CRITICAL: YOU MUST, ten bullets all starting NEVER or ALWAYS. On Claude 4.5+ this over-triggers and dilutes signal. See "Tone calibration" below.
• Embedding reference documentation (CLI commands, config schema, API listings) inside the system prompt — this ships to the model on every activation, even for unrelated questions. Reference content belongs in skills loaded on demand, not in system. The agent's system prompt is identity + behaviour + scope; everything that's "the user might ask about" goes in skills.

Tone calibration (Claude 4.6+ over-emphasis, mandatory awareness):

Claude 4.5+ is far more responsive to the system prompt than 3.x. Aggressive language written to defeat under-triggering on older models now over-triggers. Substance stays; theatre goes.

• CRITICAL: YOU MUST use tool X when … → Use tool X when …
• 🚨 HARD RULES + 10 stacked NEVER bullets → <critical> with plain Don't … / Do … lines
• MANDATORY: Run validation → Run validation after edits.
• NEVER assert X you haven't verified → Don't assert X you haven't verified.
• DEFAULT TO using web search → Use web search when it would enhance your understanding.
• After every 3 tool calls, summarize progress → drop on 4.7 (internalised)

Reserve all-caps and "must" for one or two genuine safety hard-stops (e.g. Never force-push to main). The substance test: delete the NEVER/ALWAYS/MUST and lowercase the line. Does the rule still make sense? Soften it. Does it read as filler once softened? Cut it.

Full reference (verbatim Anthropic guidance, parallel-tool-calls block, message-history rules): skills/prompt-engineering/reference/claude-4-emphasis-and-tools.md.

Markdown discipline inside XML blocks (token economy — mandatory):

The XML tag IS the structural anchor. Markdown decoration that duplicates that role is pure token waste — every ** and ### ships to the model on every call. Strip the noise:

• No ### subsection heading at the top of an XML block. The opening tag already names the section. <voice> followed immediately by ### Voice & Tone is redundant — drop the H3.
• ### subsection is fine only when an XML block has 2+ genuinely distinct sub-areas (e.g. <workflow> containing ### Research protocol + ### Memory protocol + numbered steps). Single-subsection blocks should flow as plain prose under the tag.
• No bold on bullet leads. - Active voice — "Studies show X" becomes - Active voice — "Studies show X". The em-dash already separates lead from explanation. Bold adds tokens, not meaning.
• No bold on inline emphasis unless the emphasis is genuinely load-bearing (a hard rule, a critical warning). Default: drop it.
• Keep tables. |---|---| markdown tables are real structure — they carry decision matrices and stay.
• Keep code fences. Triple-backtick blocks are essential for output schemas, command examples.
• Keep numbered lists. 1. Step becomes 1. Step — but the numbering stays.
• No --- separators between content inside an XML block. Use blank lines.

The test: would removing this markdown change the model's understanding of the rule? If no, drop it.

Temperature Guidelines by Domain

Required Disclaimers

Medical, legal, and financial agents MUST include prominent disclaimers in the system prompt:

• Medical: "NOT a doctor, CANNOT diagnose or prescribe"
• Legal: "NOT a licensed attorney, legal information only"
• Financial: "NOT financial advice"

Placeholder Variables

Naming and Metadata Rules

• File path matches tag: developer:rust → agents/developer/rust.toml
• Sub-specs use hyphens: developer:rust-nightly
• One [[roles]] per file
• # Title: — 5–60 chars. Example: "Rust Developer", "Blood Test Interpreter"
• # Description: — 20–160 chars. Concise and scannable, like an SEO meta description

Multi-Step Pipelines (external)

Agents are capabilities + one [[roles]]. They do not define multi-step pipelines — the old workflow = "..." field and [[workflows]] block were removed from Octomind. Multi-step orchestration is now the external octomind workflow <file.toml> CLI: a portable TOML that chains octomind run steps (sequential / parallel / loop / conditional) and references installed role tags. Load the octomind-workflow skill for that syntax.

[[layers]] still exist in Octomind config (not in tap manifests): they back the [[commands]] slash-command system (/run <name>) and delegate to a role via command = "octomind acp <role>". Model, system prompt, and MCP config live in the referenced [[roles]] entry, not in the layer.

Pre-Write Checklist

Before writing any agent manifest, verify:

• # Title: comment present? (5–60 chars)
• # Description: comment present? (20–160 chars)
• Is there a similar agent to reference for prompt structure?
• Do all needed capabilities exist? If not, use tap-capability-authoring skill first.
• System prompt uses XML-tagged blocks in the canonical order? (<identity> → <voice> → <scope> → <workflow> → <rules> → <examples> → <output_format> → <interaction> → <critical>)
• <identity> is the first thing the model sees? (no preamble before it)
• <critical> is last and contains a brief Don't/Do list in plain language? (no 🚨 HARD RULES theatre, no CWD/DATE placeholders, no stacked all-caps NEVER/ALWAYS — reserve those for one or two genuine safety hard-stops)
• Tone calibrated for Claude 4.6+: Use X when ... rather than Default to X / CRITICAL: YOU MUST X?
• {{CWD}} and {{DATE}} appear in the welcome field, NOT in system?
• Total system prompt under ~1500 words? (context rot threshold)
• Is the system prompt specific enough? (domain knowledge, tools, patterns)
• Are temperature/top_p appropriate for the domain?
• Does the welcome message include an emoji + {{CWD}}?
• No [deps], [roles.mcp], or [[mcp.servers]] in the agent file?
• Run bash scripts/lint-manifests.sh agents/<domain>/<spec>.toml — passes clean?
• Run bin/load <domain>:<spec> — resolves without errors?

Examples

Example 1: Minimal correct agent (2026 XML structure)

# agents/developer/rust.toml
# Agent: developer:rust
# Title: Rust Developer
# Description: Expert Rust developer. Writes idiomatic, safe, performant Rust code.

capabilities = ["core", "filesystem-read", "filesystem-write", "shell", "codesearch-semantic", "codesearch-structural", "codesearch-graph", "programming-rust"]

[[roles]]
system = """
<identity>
Expert Rust developer. You write idiomatic, safe, performant Rust — borrow checker–friendly, zero-cost abstractions, no unsafe unless justified in a comment.
</identity>

<scope>
✅ Own: Rust code, Cargo workflows, async with tokio, error handling with thiserror/anyhow, no_std embedded, FFI bindings.
❌ Don't own: cross-language glue beyond FFI signatures (route to language-specific agent), infra/deployment (route to devops:*).
</scope>

<workflow>
1. Read context — `cargo metadata`, `Cargo.toml`, surrounding modules.
2. Plan — for non-trivial changes, present an outline before writing.
3. Write — idiomatic Rust, prefer iterators over loops, `Result` over panic.
4. Verify — `cargo check`, `cargo clippy`, `cargo test`.
</workflow>

<rules>
- Errors: `thiserror` for libraries, `anyhow` for binaries. Never `unwrap()` outside tests/examples.
- Async: tokio is the default runtime. Don't mix runtimes.
- Lints: `cargo clippy -- -D warnings` must pass clean.
- No `unsafe` blocks without a SAFETY comment naming the invariant.
</rules>

<output_format>
Code edits via Edit/Write tools. After non-trivial changes, run `cargo check` and report the result.
</output_format>

<interaction>
- New task → confirm the scope before editing if files >3 will change.
- Ambiguous spec → ask ONE clarifying question, then proceed.
</interaction>

<critical>
- Don't add `unsafe` without a SAFETY comment naming the invariant.
- Don't use `unwrap()` in production code paths.
- Don't disable clippy lints to silence warnings — fix the underlying issue.
- Run `cargo check` after edits.
- Use the simplest type that fits — reach for `Box<dyn>` only when a generic doesn't work.
</critical>
"""
welcome = "🦀 Rust developer ready. Working in {{CWD}}"
temperature = 0.1
top_p = 0.9
top_k = 0

Example 2: Agent with model override

capabilities = ["core", "filesystem-read", "filesystem-write", "websearch"]

[[roles]]
system = """..."""
welcome = "🔍 Research agent ready. Working in {{CWD}}"
temperature = 0.3
top_p = 0.9
top_k = 0
model = "openrouter:anthropic/claude-sonnet-4"

Example 3: Common mistake — forbidden fields

# ❌ WRONG — never put these in an agent manifest
[deps]
require = ["muvon/octofs"]

[roles.mcp]
server_refs = ["octofs"]

[[mcp.servers]]
name = "octofs"

References

• templates/agent.toml — canonical agent template
• agents/developer/general.toml — reference for execution protocol pattern
• bin/load — resolves capabilities → merged manifest (run to debug)
• bash scripts/lint-manifests.sh — validates agent TOML files