I had a similar question a while back. My understanding is that EFA should generally be done before CFA if there aren't many/any strong a priori assumptions about the structure of the factor model being tested. If there are strong theoretical assumptions, however, then CFA is a better tool because it help help rule out alternative models.  

I found the article below to be really helpful in discussing EFA:

Henson, R. K., & Roberts, J. K. (2006). Use of Exploratory Factor Analysis in Published Research: Common Errors and Some Comment on Improved Practice. Educational and Psychological Measurement, 66(3), 393–416. doi:10.1177/0013164405282485

Thank you So-much for your answer. Please can you send the particular research article 

I routinely perform first EFA and then CFA, does not matter if it a newly developed questionnaire or not due to the fact some issues can come from some translation involved or meaning for some specific field of study.

This is a personal preference in order to prepare myself for a potential bad CFA depending upor EFA results got.

Up to now practice has been quite useful.

It really depends if you already have a model to confirm. If not, then EFA can provide much greater opportunity to be much more open to new interpretations. If you already have an a priori model, then CFA is fine (but beware of using fit indices cutoffs as golden rules). If you perform both EFA and CFA, they must be on separate samples. 

Alternatively, Exploratory Structural Equation Modelling (ESEM) provides a useful way of testing a developing model 

In addition to what others have said, I'll suggest that if you perform EFA, running the CFA with a different set of data will build a stronger case for whatever you're researching.

To add to the previous comment: If you have a large enough sample, you could randomly assign the respondents in two halves, run EFA on the first sample and CFA on the second

Usually EFA is conducted before CFA. However, it depends. If you search a new factor structure, EFA should be done (with a sample or the first half randomly selected from a sample if a sample size is large enough to split) and then to validate the explored factor structure identified (from the first half), CFA is conducted with a different sample (or the other half). However, if you already have a specific factor structure in your mind, you just run CFA to confirm your factor structure with a given sample. Hope my response helps.

Thanks for all , have a nice day. 

The ESEM method (e.g., Asparouhov & Muthén, 2009; Marsh et al., 2010) mentioned before by Jon Perry is a good starting point, if you have some theoretical ideas about your latent dimensions but still don't know whether the items work out. A similar approach --E/CFA or exploratory factor analysis within the confirmatory factor analysis framework -- was proposed by Jöreskog and Söebom (1979) and is nicely explained in Brown (2006), which is a really good source for all questions concerning CFA.

Asparouhov, T., & Muthén, B. (2009). Exploratory structural equation modeling. Structural Equation Modeling, 16(3), 397–438. doi:10.1080/10705510903008204

Brown, T. A. (2006). Confirmatory factor analysis for applied research. New York, NY: Guilford Press.

Jöreskog, K. G., & Sörbom, D. (1979). Advances in factor analysis and structural equation models. Cambridge, MA: Abt Books.

Marsh, H. W., Lüdtke, O., Muthén, B., Asparouhov, T., Morin, A. J. S., Trautwein, U., & Nagengast, B. (2010). A new look at the Big Five factor structure through exploratory structural equation modeling. Psychological Assessment, 22(3), 471–491. doi:10.1037/a0019227

If I recall correctly, EFA includes error in the correlation between variables and CFA does not (i.e., error is a separate term in the equation).  That is why using CFA is better if you want to determine validity of the model. 

Split the sample==do an EFA on the first sample, and make a decision on factor structure. Then confirm the factor structure in the second part of the sample

The decision to use EFA and/or CFA is not completely arbitrary, rather it depends on the items and constructs you are dealing with. Specifically, if your model incorporates items/constructs which have not been tested before for reliability and validity, you need to start with EFA followed by CFA using different samples of same target population.

Definitely, you should run first the EFA and then run CFA according the structure obtained from EFA. 

Shall appreciate a couple of strong citations to support that if we need to perform both EFA and CFA, they must be on separate samples. Thanks,

Dear professor @Se-Kang Kim, may I have a reference for your answer. actually, I have developed a model based on which I have constructed a scale. I have done EFA with Quartimax rotation (the factors are not correlated). I got the same result as in the theoretical model. but few suggested using Manto carle parallel analysis method which decides factors based on a sample and go for second EFA on another sample and use varimax because it is statically more reliably and when I have done that my model is completed disturbed as it suggested retaining 8 factors. and many items are deleted. and in CFA I am Losing 3 factors completely to get a Perfect fit (0.96). I have also done CFA on the originally planned model and is coming under acceptable ( CFI-0.930) fit while all factors are retained and number of items retained are acceptable (3 per item). what do you suggest sir.

Are there any references to support your viewpoints ?@Se-Kang Kim