AI accepts your question as stated. If the framing is wrong, you get the wrong answer delivered fluently. Titan opens with a Socratic stress test, decomposes the work into pillar subagents, adjusts scope between phases as findings emerge, and ships a publication-grade master with verified citations.
/plugin marketplace add brandonmcpeak/titan-research
/plugin install titan-research@brandonmcpeakAI speaks with authority about topics where its actual knowledge is shallow. Users follow confident, fluent answers down wrong paths.
Even with web search wired in, AI fails to identify the right sources, triangulate claims across distinct source types, or validate findings before composing.
Without success criteria, research sprawls. Generic insight replaces decision-ready intelligence. The work feels thorough; the operator still cannot act on it.
Most AI assistants treat the user's framing as fact. Wrong framing yields a fluent, confident answer to the wrong question, and the operator never sees the seam.
Three sub-protocols compose Titan. Each is documented as a standalone reference in the public repo; the orchestrator (SKILL.md) sequences them.
A Titan run is structured as N dependent phases. Each phase decomposes the work into pillars; pillar subagents run in parallel; outputs are synthesized between phases. Phase N+1 pillar scopes are rewritten from Phase N findings, not pre-frozen. Phase 0 (Intake Gatekeeper) and Phase 0.5 (Socratic Stress Test) act as quality gates: the work does not start until the brief survives premise validation.
multi-phase-protocol.md →
After the master draft is composed, Titan runs a two-pass cycle: V1 → CRITIQUE → V2. V1 is the dense, citation-saturated draft optimized for downstream agent consumption. CRITIQUE is a structured editorial pass that flags voice, density, and citation issues. V2 is the rewrite at consultant-briefing voice, with verified citations and a 90-minute-read shape. The pass is reserved for genuinely public-shareable masters.
publication-grade-pass.md →
After every phase, an always-on Verifier-style sub-agent asks: would a stranger reading just this phase output understand what to do with it? If not, the synthesis is sharpened before the next phase runs. The Verifier sub-agent pattern is inspired by companion-inc/feynman; Titan generalizes it across multi-phase outputs as a continuous readability gate rather than a final-step check.
humanizing-patterns.md →
Vocabulary used across the protocol.
Premise validation through targeted questioning before the brief is accepted. "What does success actually look like?" "Is this the root problem or a symptom?"
Adversarial validation before publication. "If this finding is wrong, why?" "What would a domain expert flag?" Catches blind spots and overconfident claims before they ship.
The brief is split into N pillars, each owning a discrete part of the problem. Pillars run in parallel as subagents; their outputs are synthesized between phases.
Mutually Exclusive, Collectively Exhaustive. Pillars are required to cover the problem fully and to not overlap. Forces honest scoping.
Every load-bearing claim is validated across 3 or more distinct source types (academic, industry, practitioner) before composition. Single-source claims are marked as such.
Bottom Line Up Front. Reports lead with the most decision-relevant finding. Section ordering follows attention budget, not chronology.
The titan-research plugin ships 16 files: one orchestrator (SKILL.md), 14 sub-skill references, and one quality-gate script. The orchestrator dispatches the references in a fixed phase sequence. The references are grouped below by where they fire in a typical run.
The orchestrator. Phase 0-5 structure, intake gates, dispatch contract for the 14 references and the gate script.
---
name: titan-research
description: A multi-phase deep research protocol for Claude Code.
user-invocable: true
argument-hint: "[research topic]"
allowed-tools: Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch, Agent
---
# TITAN Research Protocol v2026.05
Sequences Phase 0 (Intake Gatekeeper), Phase 0.5 (Socratic Stress
Test), Phase 1-N (pillar subagents per phase, parallel dispatch),
and the Publication-Grade Pass for shareable masters.
[Preview only. Download for the full file.]Before research begins. The orchestrator runs Phase 0 to build a working memory, gate the brief on definition-of-ready criteria, and decompose the work into pillars. These four files are the canonical references the orchestrator pulls from during intake.
The agent's working memory across a Titan run. First artifact created in Phase 0, updated continuously, never overwritten.
---
type: reference
skill: titan-research
module: thinking-file
date: 2026-05-15
---
# Thinking File
The agent's working memory across a Titan run. First artifact
created in Phase 0; updated continuously; never overwritten.
Source-of-truth scratchpad for all between-phase synthesis.
[Preview only. Download for the full file.]Definition-of-ready gates the work. The brief does not pass to Phase 1 until intake-DOR criteria are met.
---
type: reference
skill: titan-research
module: input-dor
date: 2026-05-15
---
# Input Definition of Ready
The Phase 0 gate. Lists the intake-DOR criteria that the brief
must satisfy before Phase 1 begins: clarified outcome, named
decision, scoped pillars, source-quality minimums.
[Preview only. Download for the full file.]Brief template. Sets pillar count, source-quality minimums, output schema, scope-adjustment policy, and publication-grade flag.
---
type: reference
skill: titan-research
module: brief-scaffolding
date: 2026-05-15
---
# Brief Scaffolding
The template the orchestrator fills during Phase 0 intake.
Encodes pillar count, output schema, scope-adjustment policy,
and whether the run targets the publication-grade pass.
[Preview only. Download for the full file.]Cross-phase glossary. What "pillar," "phase," "master," and "subagent" mean inside Titan, with worked examples.
---
type: reference
skill: titan-research
module: definitions-and-concepts
date: 2026-05-15
---
# Definitions and Concepts
Cross-phase glossary. Defines pillar, phase, master, subagent,
premise validation, MECE tree, triangulation, BLUF, and the
Publication-Grade Pass with worked examples.
[Preview only. Download for the full file.]The orchestration loop. Titan structures a run as N dependent phases; this group of references defines how phase scopes are rewritten between rounds, how pillar subagents are dispatched, and how emergent findings outside the brief are integrated.
Within-run, cross-phase orchestration. Defines how Phase N+1 pillar scopes are rewritten from Phase N findings. Distinct from sequential-chaining (which spans separate runs).
---
type: reference
skill: titan-research
module: multi-phase-protocol
date: 2026-05-15
---
# Multi-Phase Research Protocol (Within-Run)
Within-run, cross-phase orchestration. Defines how Phase N+1
pillar scopes are rewritten from Phase N findings, not
pre-frozen at intake.
[Preview only. Download for the full file.]Pillar object schema and subagent dispatch contract. Includes platform-specific mapping for Claude Code Agent tool.
---
type: reference
skill: titan-research
module: agent-allocation
date: 2026-05-15
---
# Agent Allocation
Pillar object schema and subagent dispatch contract. Maps
Titan's pillar abstraction to the Claude Code Agent tool with
concrete subagent_type and prompt conventions.
[Preview only. Download for the full file.]What to do with findings the brief did not anticipate. Reserves capacity for genuinely emergent intelligence rather than forcing all output through the planned pillars.
---
type: reference
skill: titan-research
module: emergence-protocol
date: 2026-05-15
---
# Emergence Protocol
What to do with findings the brief did not anticipate. Reserves
capacity for emergent intelligence rather than forcing all
output through pre-planned pillars.
[Preview only. Download for the full file.]Output discipline. Once a master draft is composed, these references govern the V1 -> CRITIQUE -> V2 publication cycle, the surgical pass that strips AI-tells from long-form prose, the content-quality gates that run before publish, and the durable register of prior art Titan learned from.
Two-pass V1 -> CRITIQUE -> V2 cycle for public-shareable masters. Produces consultant-briefing voice with verified citations.
---
type: reference
skill: titan-research
module: publication-grade-pass
date: 2026-05-15
---
# Publication-Grade Pass
Two-pass V1 -> CRITIQUE -> V2 cycle. V1 is the dense,
citation-saturated draft; CRITIQUE flags voice and density
issues; V2 is the consultant-briefing-voice rewrite.
[Preview only. Download for the full file.]Surgical patterns to strip AI-tells from long-form prose. Detect-and-fix; not a full rewrite system.
---
type: reference
skill: titan-research
module: humanizing-patterns
date: 2026-05-15
---
# Humanizing Patterns
Surgical patterns to strip AI-tells from long-form prose:
filler phrases, hedge stacks, em dashes, banned vocabulary,
and three-clause sentence rhythms. Detect-and-fix.
[Preview only. Download for the full file.]Five content gates that run before publish: em-dash check, section-symbol check, banned-vocab check, banned-filler check, bracketed-multi citation form check.
#!/usr/bin/env bash
# acceptance-gates.sh
# Five content gates that run before publish.
set -euo pipefail
FILE="${1:-}"
[ -z "$FILE" ] && { echo "usage: acceptance-gates.sh " >&2; exit 2; }
fail=0
grep -q $'\u2014' "$FILE" && { echo "FAIL: em-dash found"; fail=1; }
# ... 4 more gates (section-symbol, banned-vocab, banned-filler, citation form)
exit $fail
[Preview only. Download for the full file.] Durable register of external systems and prior art that inform Titan. Includes the Feynman attribution and other influences.
---
type: reference
skill: titan-research
module: upstreams
date: 2026-05-15
---
# Upstreams
Durable register of external systems and prior art that inform
Titan. Includes companion-inc/feynman (Verifier sub-agent
pattern) and other influences.
[Preview only. Download for the full file.]After the master ships. These references support multi-pass research where a later run builds on an earlier one, synthesis across multiple completed reports, and the discussion-document format used for stakeholder review.
Master report schema, decision matrix template, citation rules. The shape every Titan master report meets.
---
type: reference
skill: titan-research
module: report-standardization
date: 2026-05-15
---
# Report Standardization
Master report schema: BLUF, decision matrix, evidence layer,
citation rules. The shape every Titan master report meets.
[Preview only. Download for the full file.]Multi-pass discipline across separate runs. How a later research run inherits scope and findings from an earlier one without re-doing the work.
---
type: reference
skill: titan-research
module: sequential-chaining
date: 2026-05-15
---
# Sequential Chaining
Multi-pass discipline across separate runs. How a later
research run inherits scope and findings from an earlier one
without re-doing the work.
[Preview only. Download for the full file.]Synthesis across multiple completed reports. Used when a research question spans 3+ prior runs.
---
type: reference
skill: titan-research
module: meta-analysis
date: 2026-05-15
---
# Meta-Analysis
Synthesis across multiple completed reports. Used when a
research question spans 3 or more prior runs and the meta-
level findings need their own master.
[Preview only. Download for the full file.]Post-run interactive artifact format. Frames a master report for stakeholder discussion, with explicit decision prompts.
---
type: reference
skill: titan-research
module: discussion-documents
date: 2026-05-15
---
# Discussion Documents
Post-run interactive artifact format. Frames a master report
for stakeholder discussion with explicit decision prompts and
branching question paths.
[Preview only. Download for the full file.]Honest about state. 0.x permanent per STABILITY.md; no LTS, no backports, no API stability promise. Pin to a tag for stability.
Two install paths. Choose by where you work.
Requires Claude Code (free CLI). Works with any paid Claude plan that includes Claude Code access. Run from a terminal in any directory:
/plugin marketplace add brandonmcpeak/titan-research
/plugin install titan-research@brandonmcpeakThen invoke from any Claude Code session:
/titan-research <your research topic>Titan will Socratically stress-test the brief, build a thinking file, allocate pillars, and run the multi-phase loop. First run takes 30 to 90 minutes depending on phase count and source density.
If you do not have Claude Code, Titan also runs as a Claude.ai project. You will upload 16 files and set SKILL.md as the project instructions. The multi-phase orchestration falls back to single-phase (you run the parallel work yourself), but the methodology, report shape, acceptance gates, and humanizing patterns all apply.
Built deliberately small. Permanent 0.x: no LTS, no backports, no API stability. The repo stays solo-maintained with a tight scope; that scope is the survival plan, not an apology. Pin to a release tag if you need stability in production.
Built by Brandon McPeak. Independent. Released under MIT.