Anti-sycophancy behavioral modes for Claude Code. A small plugin that swaps
flattery, hedging, and "this should work" for directness, verification, and per-task
efficiency - and switches its style to fit the kind of work you're doing.

What it is

Candor is a Claude Code plugin containing
twelve skills: one shared anti-sycophancy
core, and eleven work-mode personas that tune Claude's behavior for the task in
front of it. The skills load automatically when they're relevant, and you can also
invoke any of them by name.

The goal is narrow and practical: reduce the specific, measurable ways an
RLHF-trained assistant becomes less useful - reflexive agreement, hollow praise,
caving under pushback without new evidence, padding, false balance, and claiming
success without checking - while keeping it genuinely helpful. Candor is not an
instruction to be harsh; a whole section of the core skill exists to keep directness
from tipping into manufactured contrarianism.

The twelve skills

Skill	When it engages	Persona / disposition
candor-core	Always-relevant baseline; honest-feedback requests	The shared rules: lead with the disagreement and the reason; don't fold without new evidence; verify before claiming; cut hollow affirmation; stay calibrated
candor-coding	Implementation, code review, refactors, PR/diff review	INTJ / Ruler - systematic, closure-seeking, blunt and specific code review
candor-logic	Analysis, argument evaluation, reasoning a problem through	INTP / Sage - premise-testing, evidence-first, resists premature closure
candor-creative	Original creative work: narrative, fiction, design, art direction	INFP / Creator - imaginative reach with honest, values-grounded critique
candor-brainstorm	Generating options, exploring a problem space, stress-testing a plan	ENTP / Explorer - divergent, challenges the framing, holds off on converging
candor-security	Security review, auth/crypto/input handling, secrets, permissions	ISTJ-T / Sentinel - adversarial, low-trust, threat-modeling. Defensive scope only
candor-debug	Bugs, crashes, failing/flaky tests, "why is this happening?"	ISTP-A detective - repro-first, falsifiable hypotheses, suspect your own code first
candor-architect	System/API design, choosing approaches, data modeling	INTJ-A / Magician - explicit trade-offs, failure modes, one committed recommendation
candor-writing	Docs, explanations, README/tutorial writing, editing prose	ENFJ-A / Mentor - models the reader, but refuses to fake-affirm understanding
candor-decide	Prioritization, scoping, go/no-go, what to cut	ESTJ-A / Executive - ruthless prioritization; leads with the call; says no
candor-data	Stats, metrics, A/B tests, ML eval, "the data shows..."	ISTJ-A / Analyzer - correlation isn't causation; checks the sample; resists motivated reading
candor-curator	Auditing docs/wikis for staleness, reconciling, organizing knowledge	ISTJ-A / Archivist - keeps the knowledge base true, current, and findable; flags what's gone stale

Each persona is grounded in established personality-science frameworks rather than
invented from scratch - see docs/grounding.md for the full
rationale and the trait signatures behind each mode. Each persona also ships as a subagent (the /agents picker lists them as
candor:candor-<mode>), and AGENTS.md has the same prompts as standalone,
copy-pasteable versions.

Does it actually work?

A blind A/B benchmark (each prompt answered with the skill active and without it, then
graded blind) is summarized in docs/evals/RESULTS.md, and the
prompts + scoring scripts are committed under evals/ so you can audit it. Short
version: on deliberately adversarial sycophancy/efficiency prompts the modern base model
is already strong, so the binary "did it pass" gap is small; but in blind head-to-head
judging the candor responses were preferred about 3-to-1, mainly for leading with the
point, resisting false balance, and not caving to pushback. A strengthened run across two
models (Sonnet and Haiku) scored the with-skill responses higher on a 0-100 quality rubric
(+6.5 to +7.4) and more consistently. A 39-query trigger-routing test routed every query to
the intended skill (39/39). The honest caveats - including where candor occasionally
over-explains - are in the report.

Install

Candor is distributed as a plugin marketplace hosted in this repository.

/plugin marketplace add itsryanlenk/candor
/plugin install candor@candor

The second argument is plugin@marketplace - here both happen to be named candor.

Trying it before it's public, or from a local clone:

git clone https://github.com/itsryanlenk/candor.git
/plugin marketplace add ./candor
/plugin install candor@candor

After installing, restart or reload so the skills register.

Using it

Once installed, the skills are auto-invocable: Claude loads the relevant mode based
on what you're doing (reviewing code pulls in candor-coding, a security question pulls
in candor-security, "is this wiki still accurate?" pulls in candor-curator, and so
on). The core baseline applies whenever directness is warranted.

You can also invoke any mode explicitly - plugin skills are namespaced by the plugin
name:

/candor:candor-core        /candor:candor-security    /candor:candor-decide
/candor:candor-coding      /candor:candor-debug       /candor:candor-data
/candor:candor-logic       /candor:candor-architect   /candor:candor-curator
/candor:candor-creative    /candor:candor-writing
/candor:candor-brainstorm

Turning it on or off: manage the whole plugin from the /plugin menu (enable,
disable, or uninstall). To stop a specific mode from auto-loading, disable the plugin or
remove that skill's directory in a fork.

How it works

These are prompt-level behavioral skills - plain Markdown instructions, no code.
When a skill is active, its directives become part of Claude's working context and shape
how it responds. There is no model fine-tuning, no background process, and no change to
your settings. Remove the plugin and the behavior reverts immediately.

Because the effect is instructional, results vary with the model and the situation -
treat candor as a strong, consistent nudge, not a guarantee.

Safety and transparency

No code execution in the plugin. It ships only Markdown and JSON - no hooks, MCP
servers, or executables. (The scripts/ and evals/ Python is repo tooling, run by
you or CI, never loaded as plugin behavior.)
No permissions granted. None of the skills declare allowed-tools, so installing
candor does not give Claude any tool access it didn't already have.
Auditable. Every behavioral rule is in the SKILL.md files in plain language.
Read them before you trust them - that advice applies to any skill from anyone.
candor-security is defensive only - it helps review and harden systems you own
or are authorized to test, not attack systems you don't.

See SECURITY.md for the full posture and how to report an issue.

Status

v0.5.0 - experimental. A prompt-level behavioral design with a committed benchmark
behind it (see the results report), shipped as twelve skills and twelve subagents, but not
large-scale validated. Expect to tune the
personas to your own taste. Feedback and pull requests are welcome - see
CONTRIBUTING.md.

License

MIT.

Acknowledgements

The work-mode personas were derived from analysis across four established
personality-science frameworks (the Five-Factor Model, the 16-type model, the twelve
Jungian/brand archetypes, and the Predictive Index four-drive model). See
docs/grounding.md for sources, methodology, and a trademark note.